# CLAUDE CODE MARKETING FULL COURSE (6 HOURS)

Channel: Nick Saraev
Video: https://www.youtube.com/watch?v=yulWjh3rq28
Duration: 6 hr 3 min
Language: English
Words: 81678
Transcript page: https://viewrankai.com/tools/youtube-transcript/yulWjh3rq28

---

[0:00] Hello and welcome to the 0ero to1 course all on cloud code for marketing. By the end of this video, you guys are going to know how to use modern AI tools like Cloud Code to automate virtually the entirety of the marketing function. Uh whether you guys run your own company or working inside of somebody else's, that's all good. My name is Nick. I run a marketing business that just did over $500,000 in revenue last month using an overwhelming amount of cloud code. I've also been working with AI for over 7 years now, way back when the smartest model was called GPT2. Uh and I've trained some big teams on AI now, like

[0:29] the Mr. beast team and a couple of others. So, I know all about how to take somebody from never having used or applied an AI tool all the way up to, you know, generating large returns in a real functioning business. So, all that to say, you don't need any programming experience. You don't need any prior AI experience to understand all this. I don't have a degree in coding or anything like that. I learned everything I'm about to teach you from free YouTube videos just like this one. We're going to start by looking at a really brief outline and then I'm going to walk you through it. And you don't need to jump

[0:55] around. Literally just follow the course as I laid it out and by the end of it you'll know everything you need in order to automate your marketing. Let's get into it. So the course is going to be focused on actionable real implementations. In order to get there, first I'll show you how to download, install, and set up Claude Code. I've uninstalled it from my computer and I'll actually walk you guys through every step including signing up, creating an account, and then what every button in the interface means. So it'll be very simple. Then I'll cover the five marketing functions that we'll be automating today using Claude code. I'll

[1:26] also show you guys a framework I'm going to use for that. Then I'll run you guys through the difference between prompts, skills, loops, and cloud routines. And the way we're going to approach the rest of the course is I'm going to be showing you guys how to automate things using Claude uh at each of these four levels. And you can think of prompting as the simplest level, Claude routines as the most advanced level. Prompting requires you working on your computer whereas cloud routines effect effectively is a loop that is scheduled that runs in the cloud without you. Then we'll actually go and build stuff. So I'm first going

[1:58] to show you guys top of funnel how to automate the generation of creative. You guys will be able to automate ads, creative for organic. We'll be doing this via images, video, even some audio. Then I'll show you guys the next function which is the personalization of copy. And so we'll be doing newsletters, we'll be doing pulled emails, we'll basically be taking a template, and then I'll show you guys how to use Claude to automatically um automate the process of filling really fuzzy variables that it goes out and finds on the internet and so on. In the middle of funnel sense, I'll then show you guys how to make

[2:29] speed tole systems for appointment booking. These are systems that allow you to squeeze way more juice out of ad spend and then organic growth. Then I'll show you guys how to set up data collection, tracking, dashboards, and so on and so forth. Actually building your own proprietary analytics platform, uh, interpreting the data through cloud, and then ultimately scanning that up somewhere on a dashboard so that you or your clients can see what that looks like. And then finally, at the end, I'll show you guys bottom of funnel, which is

[2:56] how to automate highquality follow-ups. Now, this is basically nurturing. It's starting to verge onto sales, and I don't want this to be a sales course. Um, but I'll show you guys how to automate the process of sending emails, SMS, notifications, and so on and so forth, and doing so in a way that seems very human despite being entirely automated. Finally, I'll show you guys how to maintain and then upgrade these systems over time before giving you guys some advanced tips that I have hard learned, I sort of learned the hard way after working with a lot of large businesses, multi-billion dollar enterprises, and so on. Okay? So, you

[3:25] don't need to worry too much about any of this stuff right now. We're going to go step by step. We're going to start slow and by the end of the course as mentioned you'll be able to do everything that I just talked about. The first step we need to do is we need to create an account and the way to do that

[3:37] is to head over to claude.com. So I'm going to do that first with you guys. claude.com right here. Now in order to create an account all you need to do is click continue with Google or continue with email. The interface may look a little bit different by the time that you guys are looking at it. But what I'll do is I'll just go continue with email and I'm just going to pump in

[3:57] my own little email address over here. So this is the name of my startup. I'm just going to go continue to click the link sent to this. So I'm now I'm going to go to my email. Okay. And I have it right over here. I'm now going to click sign in to claude.ai. It's going to take me back to this page right here. Now this will quickly

[4:16] basically ask you a couple of questions. I agree. I consent. And I think I'm not going to subscribe to the promotional emails and notifications. Sorry, Anthropic. Now, we're going to have to do some onboarding questions. So, just for the sake of not doxing myself, I'm just going to put in a bunch of random

[4:30] dates here. We'll go 010190. Makes me squarely millennial. Uh, I'm going to be using this for personal use, and I'd recommend that you do as well. They're trying a team/enterprise feature, but if it's just you for now, just click personal use. Now, you will have to pay in order to use claude code. Hopefully that's clear. This is a financial transaction. The idea being you invest, in my case, $24 Canadian per month into this. Then you make far more than $24 Canadian a month back in ROI. I can say as somebody that operates a business that does half a million per month. Um, this is by far my number one

[5:06] investment. I probably receive an ROI uh something to the tune of an absurd 10,000% or something like that just off my Cloud Code subscription. Um, you can use Cloud for free. It's just you can't use the code aspect. And as you'll see, the code aspect is very foundational to us being able to use Cloud to do really, really high level automation super quickly. So, I'd recommend that you guys do this for sure. But if you don't have the funds, know that you can also just

[5:31] watch me build things in the course. Maybe wait until you have the 28 Canadians a month, let's say, to sign up and then circle back after you watched it end to end. So, don't worry too much about that. Uh, but if you want to do things with this, you will most definitely have to eventually pay. Okay. Okay. So, monthly right up here. I'm going to go down to payment method to actually put in some info. Okay. And I just wrapped that up. Cool. So, now once we're done, it'll automatically take us to the next step. As mentioned, if your interface looks a little bit different, that's okay. What we do after this is

[6:02] we're going to download it according to our operating system. Now, in my case, my operating system is MacOSS. If you guys couldn't already tell, just based off the way the Windows look and stuff like that. It'll be a little bit different uh if you guys are using Windows, but it'll automatically pop up right over here based off of like a little browser cookie thing. So, I'm

[6:20] going to give this button a quick click. And as you can see, a lot of the time on Chrome, it will automatically block the download. So, what you can do is you can just go to download unverified file in the top right hand corner. Okay? And then you'll actually be able to see it. So, claude 6.d. Okay. And then once you're done with that, all you have to do is just drag this little Claude app over to the applications folder. So, I'm doing that. And I just heard a little chime in the back end. That's because it has gone through and actually copied Claude to my applications. That's just

[6:48] how you do things on a Mac. On a Windows, you'll have to go through that whole prompt uh that you guys are probably pretty familiar with. And then once that's done, all you need to do is actually launch Claude. And then we will have opened, you know, the Claude desktop app. So on Mac, all I do that uh in order to do that, all I do is I take this little window. This is called Raycast. Then I just type Claude. You

[7:07] can see I'm actually right over here. Don't pay attention to the rest of the stuff. Once I give that a quick click, it will then open Claude. And if I go down to the bottom here, it'll say it's an app downloaded from the internet. Are you sure you want to open it? After you're done, it'll then look something like this. And all we have to do is click get started. I'm now going to relog in. So, continuing with my email address here. I'm then going to have to go to my email address and then enter the verification code, which I will get right over here. When you click sign in,

[7:34] it'll take you back to the app and we should be good to go. Cool. So, I've already created the account to be clear. Um, what it's doing here is it's just double-checking a lot of the stuff. So, I'm just going to click continue. We'll start with my name, Nick. What kind of role we do, I'm just going to say product management. I'm then going to say I have my own topic. And now we are in. So, first of all, there's a lot going on on this interface and you'll see so many little widgets and buttons and tools and things you can push and pull. Uh, I'm not going to cover every

[8:06] single one because I don't think that would be a very good use of time. And what you'll quickly find is Anthropic, the parent company of Cloud, is testing a lot of different user interface features and they're trying to collect a lot of data on what people actually like. So they'll change these up pretty

[8:18] often and they're pretty short-lived. But what I will do is show you a couple of them. Um top lefthand corner. Okay, we have obviously the three little close the window buttons. If you're on a Windows, it'll probably be in the top right hand corner. Then you have a tab open or close button, which is how you open or close the sidebar. Then you have a search chats and projects button which allows you to basically um search through all conversations that you've

[8:42] had to find more or less any info. That's actually pretty useful. Next up you have the home versus code tab. And so basically there's almost like two products here. There's like a clawed product and then there's like a clawed code product. And so basically what we want to do is we don't want to use the home product, okay, which is simpler to be clear. We want to use the code product. When you jump over to the code product, a lot will change. you know, if it's the first time you've run it, it'll look something like this. When you don't have any conversations in the queue or

[9:07] sessions going on, what it'll do is it'll give you a quick overview of all of the conversations that you have started and uh basically had over the course of I think it's like the last 6 months or something like that, 3 months anyway, a lifetime basically from when you created your account. Now, in my case, you know, I've created the account quite a while ago and I've had a lot of conversations with Claude. So, as you can see here, it's saying that you sent 12,28 messages, 1,123, you know, 3,289. I've I've had a lot of back and forth with them. As a result, I've used more 241 times more tokens

[9:38] than Pride and Prejudice supposedly, which is pretty cool. You can also break that down on a 7-day basis, 30-day basis, or all. And, um, you can also see which models you use the most. In my case, I use the smartest model as of the time of this recording, Fable, pretty often. Okay. And you can see that my usage started right around July. So

[9:55] that's enough of everything over here. Let's go back to the left hand side. Underneath new is artifacts. In case you guys didn't know, an artifact is basically a visual that Claude can create for you. And so if you wanted to create a brief visual, what you can do is build a code artifact. Um well, in this case, I just clicked new and it automatically populated this. I want to build a code artifact, a self-contained web page published with the artifact

[10:18] tool. Um make this page about Nick Sar. Okay. Okay. And then I'm going to press enter. And what it will do now is it'll actually go through presumably ask me a couple of questions. And then it's going to design a visual interface for me, an artifact inside of Claude that I can then access anytime I want. So now it's confirming the angle, gathering a couple key facts before designing. And now it's going to go through and get me some questions. So what's this page's job? A personal bio portfolio site. I'm just going to say it's a personal bio portfolio page emphasizing the YouTube channel and automation agency. And what

[10:53] you'll notice just happened there is Claude actually suggested an autocomplete down at the bottom. And so all I had to do when that autocomplete popped up is I just pressed tab and then it automatically went all the way to the end. This is basically now letting Claude drive your conversation. And it's not always the best thing to do to be clear. Um but it is pretty neat and I would say it knows what I want to do next maybe 30 40% of the time which helps. And you can see it's now coming up with a bunch of uh information based off of publicly available work like

[11:21] leftclick maker school my subscribers where I live and so on and so forth. Not all this is up to date because it's been a while since you know the web pages that hosted all this information um have been created but it's pretty dang solid. It did that in just a few seconds. Okay. Now here we have what's called a request. This is asking to see if I should allow Claude to run this command because this command requires my approval. And so you can see there are

[11:47] three things that have popped up here. There's a deny button, a one, an always allow button, which is two, or an allow once button, which is three. Generally, it thinks that allow once is the right answer, which is why it's popping up as that little highlighted thing. In my case, I'm just going to go always allow because this is just building a simple web page. It's probably not going to try and delete my computer by accident. Now, once we've done this, you can see the interface change more. There's this little tab that says creating. If you click on this tab, it'll actually show you guys the commands that it's running

[12:15] under the hood. If you guys feel like supercomputer programming, you want to see what's going on. And you can actually see as well the amount of total time that it's spent working as well as the total number of tokens that it is spent as well. Now, tokens are important to understand because tokens are a big part of how you are build at the end of the month. They're also a big part of how much uh usage you basically have with Claude. The more usage you have with Claude, the more tokens you can, you know, run through. And typically, the bigger and sexier the project, the more tokens you will require. So, it's

[12:45] important to understand and sort of have visibility into your token budgets, which is why Anthropic does it right here. Now, I'm running this in real time because I want you to actually be able to see everything that is running. Like, I want you to know how long it'll take to run one of these things. Um, I find a lot of the time on YouTube people will just cut to the end. And while I will be doing some cutting later on in the course when we do larger and bigger builds, let's say ones that are, I don't know, 15 or 20 minutes, I'm not going to force you guys to sit down and and watch

[13:09] all of that, um, it's important to at least see what a real flow looks like start to finish. Okay, once all of this is done, if I just move my big fat head out of the way, you can see that we now have um, you know, we actually have a web page that we built. Okay, and this web page is the artifact. Now, this isn't on the internet to be clear, and it's not going to be perfect all the time, but I'd say it's pretty good. You can see it basically built me this little portfolio site saying Calgary, Canada, building in public automation systems that run without him in the

[13:36] room. Oo, super fancy. So, it gives a bunch of information right over here about my YouTube subscribers, views on the channel, the number of maker school graduates. You can then see that uh, you know, I went to behavioral neuroscience in school. Uh, this was actually wrong because I think somebody made a big post about how I left college or something that ended up getting really popular, but I didn't actually leave college. I graduated with my degree. Um, you can then see this sort of like timeline here

[14:02] with maker school and YouTube and stuff. All all very cool. Anyway, the point that I'm making is not, you know, you can make a website. Hopefully you guys know that by now. If I move my head back here in the top right hand corner, you can um click this little button here to expand it or open it in a default browser. So now this is actually open in my Chrome and the really cool part about opening it in your Chrome is once you have the artifact online you can actually share it with people. So what you can do is you can actually share it with people that are in your uh company

[14:29] using a teams or enterprise plan or you can share it with people um that are just on the internet like basic any basically anybody. And so I'm now gone to share anyone with the link. If I go copy here, then I open this up in let's say an incognito tab and then I paste this, you'll see that I'm now accessing that artifact despite the fact that I'm not even signed in. And so in this way, you can basically host web pages now with anthropic and and cloud. And it's really useful particularly for like dashboards and for managing projects for clients. You know, I do a lot of client

[14:58] work now and a big chunk of my client work these days is actually building like web interfaces and internal tools for people. Well, Artifacts solve that for me completely. I no longer need to worry about actually hosting a website. I can just whip up a quick artifact and then give somebody the link. So that's

[15:12] that artifacts tab is right here. Okay. Next up, you have the ability to customize. I know the second you click this, it'll look really intimidating and frankly quite terrifying. Uh but this customize button is is quite valuable. It'll basically take you all the way down to the skills tab under customize, which is where you can add uh what is probably going to be one of the most important parts of this course. skills which are almost like programs that you can run on knowledge work. So, Enthropic will have created a couple for you assuming you're watching this at least sometime in the next month or two uh

[15:42] like morning and uh they'll also have done I don't know setup writing style and skill creator. These are just meta skills to show you how they work but I'm just going to run you through this one here morning to start. Now the idea is this is a skill. This is something you actually have access to. Render the user's morning brief as a styled HTML artifact or set it up as a recurring weekday task. Use only when the user explicitly asks to run, see, or set up their morning brief or if they invoke /morning by name. You may be wondering exactly what that means. I'll show you

[16:13] guys in a second. A question about their day, schedule, or calendar is not by itself a request for the brief. So, answer it directly instead. And so what this is going to do basically is this is going to automate the process of coming up with a prompt and telling a model what to do. Um you know before I said hey make me a website. The website is a one page we're about nixar. Well now what we have is we have a skill which all we need to do in order to actually generate a page is invoke it by going /wmarning. When you do it'll pump this

[16:40] prompt in to claude which is quite long as you guys can see here. You can actually add a lot of detailed information. Then all the way down at the bottom it'll basically generate a page for you. Okay. So why is this under the customize button? This is under the customize button because this is how you customize your cloud instance. This is

[16:57] how you optim uh customize cloud code. You add skills which are basically domainspecific programs that you can run to do things for you. And so for instance when we do ad creative uh later on in the course when I actually automate that for you using cloud code we will be using like an ad creative skill. I will create one alongside you and that ad creative skill will automate the process of like producing a bunch of ad variants that we can test for marketing. You know, if you create a uh uh I'm not sure a dashboard skill or something. What you can do is I don't know once a day you can quickly set that

[17:30] on a loop to automatically pull all of the different data sources. Let's say Google Analytics API, a big Google sheet that contains data, a bunch of I don't know s you know agency analytics software and stuff like that and then just pull all that in and use it to populate a dashboard. You can do all that stuff just using skills. So, it's very important to understand um sort of what they are and how they work and we're going to go a lot into detail about that later. But that's not the only way you can customize your cloud code instance in the app. You can also do so with connectors. Now, connectors,

[17:57] those you guys that don't know, are basically just uh ways to hook up apps to cla. And so what we have here is we have the ability to connect our Gmail. We have the ability to connect Google Drive. We have the ability to connect Slack. And so as I create these automations with you guys, we'll be using connectors to integrate a variety of different data sources into cloud code like emails, like you know, Slack channels and so on and so forth. And you can automate virtually whatever the heck you want. I mean, as you guys see, we're

[18:26] even doing this with Spotify, right? Spotify is connected right now and has access to get the current track, get the player state, get the volume, get the position, get the repeat, get the shuffle. These are all functions inside of this connector. And so, you know, you could actually ask uh Claude to do something for you, like, hey, you know, I want you to turn up the volume of Spotify and it and it can do so via connections. So, you can connect things, disconnect things, and so on and so forth. A lot of really cool stuff to play around with there. Finally, you have plugins. Now, plugins sort of

[18:54] depend on exactly what you want to do with Claude. And as you guys can see, I have a couple set up right here. One's called Claude Mem, another Superbase, another's clanged LSP. You don't need to know what any of these three mean, but I just wanted to show you guys what this looks like in an interface. So, when you click on it, okay, you'll see that there is a source version, and an author. And this source is pretty interesting because if you go on, if you click that button, you'll see there's actually a variety of plugins that have been created by Anthropic and Anthropic's partners to allow Claude to do things

[19:23] easier. So for instance there is a finance plugin which automates the process of creating journal entries, financial statement reconciliation, streamlining your accounting workflows and so on. Legal which speeds up contract review, NDA triage, compliance workflows for in-house legal teams and so on. There is bio research which connects to pre-clinical research tools and databases. There's code modernization which is a plugin used to modernize really old code. There's a variety of different plugins here and they do everything from changing how Claude looks. So actually changing I don't know in this case the output style or I don't know doing structured actions for you that are similar to skills but

[20:03] not exactly skills like code review. Okay. So we're going to show you guys how to do all of that in this course which I think is going to be quite valuable. Then finally at the very bottom you have memory. Memory is just sort of what Claude knows about you. And so I don't know, remember I'm, you know, we talked about height, you know, a couple days ago with Claude and I remember asking it to do some BMI stuff for me and it was like, I don't know how tall you are, so how tall are you? So what I said now was, hey, you know, I'm 6'2. So what it's done is it's created a

[20:35] memory profile with basic identity facts about the user details is 6'2 tall. You know, you could put in weight information. My dog's name is beans. you know, you could put in uh information about your career, about how much money you make, about, you know, your marketing channel, about uh the priorities of your business and so on and so forth. And so that is what the customiz tab is for. Next, you also have this more tab, which exposes routines, which is going to be quite an important part of our work um as we get on a little bit later. What this allows you to do is basically create templated

[21:08] skills. They call them routines, but they're templated skills that are kicked off on schedule by API or web hook. Now, you don't need to know what API or web hooks mean, but obviously a schedule is just when you want it to. So, what you can do, for instance, is you can build a skill, which is a program like I talked about earlier. And then you can just run that at, I don't know, 5:30 a.m. every morning. And so, this is the last step of automation. This is like the max automation where we're not actually talking to Claude anymore, nor are we even running skills, nor are we looping

[21:39] those skills at a regular interval. We're actually scheduling these to run on the cloud outside of our computer. And this is where you really unlock like fully automated business stuff, which is typically what people work with me for. Um, the full sort of endto-end pipeline automation of a aspect of their business. So, it's actually quite easy to do and I'll show you guys how to do that soon. Finally, you have the ability to customize the sidebar here. So, I'm actually going to open that routines tab so we can also see that underneath you have the various conversations and folders that you're working in. So, I'm actually working in a folder called for

[22:10] YouTube right now, which is why it says for YouTube here. If I just zoom in on this just a tad, you also have the ability to create new conversations by clicking this plus button. Then you can change a bunch of information about these the status of your conversation, the environment that you're working on. um you know all of the stuff is a little bit more technical and not going to be super relevant to us immediately. So I'm

[22:31] just going to leave that for now. Okay. And then that takes us to the actual chat window itself. Now as you guys can see here when you don't have any sessions in play you have this big um sort of like dashboard with all the information about how many tokens you consumed and stuff like that. So the number of sessions, the number of messages, the number of total tokens and stuff like that. You also have the ability to check model specific usage and break that down. Well, I think they just included this information so that you wouldn't have to look at a blank screen and because this looks way sexier

[22:57] than a blank screen. The parts that are important for us are first of all this button which is how you switch between running stuff locally on your own computer running stuff in the cloud K which is Anthropic computer using remote control which is running it on another computer that you own and SSH which stands for secure shell which is uh just a protocol used back in the good old days to communicate with computers in a secure way using passwords and stuff like that. So we're probably not going to touch on the SSH stuff. I'm going to show you guys remote control just sort of so we all know how to run marketing

[23:30] workflows um on another computer. But really the two we're going to be using the most are local and cloud. And local, as I'm sure you guys can guess, is actually quite easy. Um basically what it is is it's my local computer. And when you click on that little settings gear icon, you can actually update that local environment with a bunch of usernames and passwords. Now the usernames and passwords are written in kind of intimidating ways. Nodeenv equals production. But what these are,

[23:52] these are just usernames and passwords. You know, you can imagine it be like user nick atclairvo.io, I don't know, password, hey bros, I love claude code for marketing, whatever. Well, now we've actually stored like a credential called user, which equals nicer.io, and a credential called password, which equals that. So, we can save that. Now, my my local environment actually has that information. You can see it's even sort of hidden some of that from me. Um, so I'm just going to save the changes back here so that we don't have anything that confuses Claude, but you can use that to log into apps and various APIs like the meta ads

[24:25] API and whatnot will typically require you to have like env keys and stuff like that. Okay, then we also have the cloud button here. I can click add cloud environment that'll take me to a panel here which basically allows me to do the same thing. Cloud is speaking loosely here. What this is is this is on Anthropic's own computer. So, Anthropic is running this for you and you just pay Anthropic instead of paying, you know, a computer company or a server company to host. Um, I won't get super deep into the pricing here because we're probably not going to use cloud environments for most stuff. We may use them a little bit

[24:56] later, but this is valuable in larger teams. And the reason why is because if you think about it, right now I'm running stuff on my computer specifically, but my computer works a little bit differently than other people's computers, right? And so because of that, sometimes you'll create something that works on my computer uh but it won't work on somebody else's computer. And so the benefit to using cloud environments a lot of the time is you're standardizing it. Everybody uses the same cloud environment. It's the same server with the same environment variables, same usernames and passwords, same setup scripts, same everything. And then as a result, you just know that

[25:26] like the app that you're creating for your cloud code for marketing course is going to work the same whether it's on my computer or yours. Down over here, you can actually select the folder that you want to play in. And so I'm just going to do um for YouTube. I'll keep that. You can also just add a new folder here anytime that you want. So I have a couple more. Clarbo backend, Fable 25, and so on and so forth. U over here on the right, you actually have this little claude widget. And if you just kind of move your mouse around, it'll do some cute fun stuff. Uh kind of funny cuz

[25:51] this thing is actually galaxy sized brain, but the company made it look super cute. It's breaking out of your shell. Uh and then on the bottom lefthand corner, you'll see the mode, which is very important. My recommendation is to run most things in auto mode. Basically what modes are is you guys remember when um I just launched Nixrive webpage before and asked me hey do you want to allow the ability to run XYZ command? Well a manual will always ask you before it does anything that is even remotely considered like you know autonomy in your computer. And that's good if you're working on like something really secure,

[26:26] but obviously eventually it becomes a pain in the butt, right? Because I don't want to have to like wait, you know, around to see what it needs me for. So, what I typically do is I'll typically use um accept edits, plan, auto, or bypass permissions. Now, what accept edits is is it won't be able to create new files, but it will be able to edit pre-existing old files. And so, that's pretty useful, right? You can run it a mock in like a codebase or I don't know, some marketing folder or something like that and have it upgrade everything. um but it won't be able to create or delete

[26:54] any files, which is sort of a a certain level of control that you could imagine would be useful. You can also use plan mode, which is where instead of making any changes, you create a really in-depth plan about how you're going to do it instead. You can then do auto mode, which is where Cloud will handle most permission decisions for you. It'll only ask for your help if there's something that it considers to be kind of scary. You know, if you're playing around in your file systems, and I don't know, the file systems on your Windows

[27:19] will delete a certain system 32 folder. um you know, Claude will cla will tread very lightly and it'll ask you uh as you work on things that are close to that folder, but if you're just working on like random ad creative or something, for the most part, you'll never actually have to worry about a little um dialogue box popping up. Then finally, you have bypass permissions mode. Uh I love bypass permissions mode. It's basically just like full you only live once, let's go, Claude, 1v one me. Uh you know, it's

[27:45] not going to ask you a single thing. Claude's just going to do whatever the heck it wants all the time. Um, I am typically in the loop in some way, shape, or form. And I typically do not work on things that are super sensitive. You know, cyber security projects, super sensitive like passwords, credentials. I don't do any of that stuff. So, you know, I'm just working on little marketing tasks. I'm usually going to be auto or bypass permissions. So, as you can see, right now, I'm actually on bypass permissions. When you click on that button, you can then scroll all the way down to the page and it'll actually

[28:11] ask you, hey, do you want to allow bypass permissions mode? And you have to do that check because obviously there is the possibility of some sort of long-term impact. So what I'll do is I'll just go back here. And now once I've accepted or allowed this, I can now just go five and bypass permissions. When you do so, it'll kind of make your screen seem kind of scary. So, you know, don't do it if you're not like actually okay with it. Um, auto mode, plan mode,

[28:36] accept edits mode, everything is fine. Uh, I just wouldn't really recommend using manual because you'll have to check in like every 5 seconds as it does anything. Okay. Okay. Then a little bit to the right, we are now exposing the capability to add files or photos. So I can actually add any file or photo to my claude code conversation. So I have some videos here that you could see I did with, you know, Jack Roberts. I have a big list of community members. I can actually export all of these and actually upload them directly into the chat. So this is a picture of uh me, Alexi, and Sam Ovens at uh you know the

[29:06] school games where they were interviewing me about stuff. For instance, maybe I want to ask it questions about how to make this better or have it generate a new image using another model or something like that. I could do so really easily. You can then add a whole folder as well, which is basically the entire folder's worth of context. It'll just go through everything in that folder, consume it, and then, you know, sort of understand what is going on. And then you have the ability to expose slash commands, which if you guys remember earlier were those slash warnings, the skills and so on and so forth. So slash commands are kind of

[29:36] cool. Um, there are a lot, so I'm not going to go through every single one, but I am just going to show you guys what the interface looks like. If you type slash, you'll see there is a massive list of all of the different possible things that you could do here. There's slash schedule/btwreview

[29:51] slrefork/mcp/feedback/model/reame. You know, this is sort of like um you reading through the help docs of a a very complicated piece of software. There are a lot of these and we're adding more and more and more all the time. But what is really cool is this will also include um skills. So just like you know we had um you know all of those different functions here we also have anthropic-skills colon morning. And so for instance that would actually run the morning flow the same thing that um you know we were reading through a moment ago. So as you can see I'm saying anthropic-skills morning. This will take

[30:28] a few minutes. Let me check what's connected before gathering. And the reason why it's doing that is because it's looking to see, hey, do you have email? Do you have calendar? Do you have chat, Slack, Gmail, or Outlook? Because those were the five platforms or six platforms that were used in the skill in order to, I don't know, give you your morning brief. The value to the morning brief is composed of data you would find on places like this. So, the fact that I don't have any of those platforms connected means that it's probably not going to be all that valuable of a skill run, right? But as you guys see, these

[30:54] things are quite organic and uh Claude will let you know as you proceed through building skills and stuff like that, whether you have the appropriate permissions to make that skill valuable in the first place. Okay, there are a couple of these um slash commands that I would recommend understanding. Um one is called BTW. This allows you to ask a quick side conversation before you actually stop the prompt. Um so for instance, I could say, "Hey, what's my name?" When I do this, what it'll do is actually open up this little side chat

[31:22] right next to your main conversation. And you can ask it a question without actually interrupting the conversation that you guys are having down here. And so that's what's going on up here. Okay, I'm actually just having a conversation with it. And it's still technically in the thread. It has all the context of the thread, but um we're talking without actually interfering with anything down here. So then I'm just going to close out the /btw. And you can see here it's now asking me for, hey, can claude open this folder? I'm now going to click allow once. And you can see that it's actually gone through and created the

[31:51] page for me. But the page is pretty boring. It says nothing's connected yet. So there is no day to show you. Nick, you should get your stuff together, right? And so that's what that works. It's going to ask me multiple times. I'm just going to keep on closing this out. I don't really need to worry about that. Okay. Now, I don't like the fact that this has been talking for a while. So what I want to do is I actually want to pause it. And how do you do so? You can click this button right here. You can also press escape. The reason why that's valuable is because sometimes cloud gets

[32:16] caught in loops where it just reopens the thing over and over and over again because you keep exiting it or it keeps on editing a file for you or you change tac and you don't want to use /b2w or something like that. Now moving down slash commands. Another one that I really like is slash usage. This will show your usage plan as well as the um rate limits and how far you are through them. So you know I've done a couple of things with you here on my $20 plan. And so uh you can see my 5-hour limit resets in 4 hours and 21 minutes is at 9%. So

[32:44] I've used 9% of my 5h hour limit. To make a long story short, Anthropic currently gives you five whole hours um as a window in which you can use Claude and then they give you a certain amount of cloud usage within those 5 hours. After you hit 100% you can no longer proceed in that 5 hours. Then when it resets, you go back down to zero and then you can run it up all the way up to 100% again. In addition to the five hour limit, they will also give you a weekly limit. And so this allows you to access all models. The reason why they do this is because sometimes they

[33:17] give you limits to specific models, usually new ones, which are very in demand. Uh, and so this one here resets Friday, 9:00 a.m. And you can see I'm 1% of the way through my weekly limit, 9% of my way through the 5h hour limit. [gasps] It'll also give me some information about the session cost. So this cost an additional 88 cents. I think cuz I was using uh I don't know some smarter model or some API or something. Um looks like I was using an API for 1 minute and 2 seconds and I was using mostly sonnet for this task which is one of the many claude models. Then

[33:47] we also have the token breakdown which is quite valuable as well as what is actually using my limits. 67% of all of my limits were used by that one skill. Okay. And you can see that consumed like I don't know I guess like 6% or so of my total 5 hour limit usage which means I could probably run warning maybe every I don't know I could run it like 20 times

[34:06] in a row. So that's pretty cool. Okay. And then finally uh the last skill that I wanted to show you guys was uh if we scroll down here a little bit there's loop right here. Now loop is valuable because it allows you to define an interval. So I could say, I don't know, let's just say 5 seconds. And then

[34:27] I'll just say just say hey Nick. And so what this will do is every 5 seconds it will start a loop that every 5 seconds responds to me as hey Nick. [gasps] Now there's a rounding rule currently in place. I think that's because there's a one minute floor currently put on by um Claude. But anyway, just to show you guys, let's do the one minute loop. Um, what this will do now is this will set a loop where basically once a minute, every minute, this will just say the words, "Hey, Nick," to me. And so, as you can see, this is a pretty straightforward thing to do. Uh, but as I'm sure you guys can

[35:07] imagine, we can go much more complicated than this. And in fact, we can actually run skills like this /morning skill, and we can do that on a daily or that on a one minute loop or that every, you know, 45 minutes or whatever. We could also very trivially set skills that'll do things like check pages for us. And so, you know, for a lot of SEO or newsletter purposes, that's pretty valuable because what you do is you'll like add connectors and then you will check, let's say you're running a big campaign, a flash sale. Well, you actually set a loop where cloud will check to see every

[35:36] 5 minutes how many new customers have signed up. Let's say you can take all those customers that have signed up and then provision them access to some asset or whatever. Um, and it'll just continue doing this in the background over and over and over and over again. You can see we've now done that three times. Okay. So, if you want to stop the loop,

[35:51] just say something like stop the loop. Cloud will kind of hear what you have to say, find the tool, and actually like cancel the tool call that it just generated. And then, uh, yeah, you're good to go. All right. So, hopefully that was a reasonable walkthrough of all of the slash commands. Last thing I want to do here, um, is, you know, obviously we have those connectors, which I've talked about before. We have the ability to add plugins. You'll see there's a lot of um redundancy built into this because they have the same buttons three or four different places just because they're testing to see which ones people like

[36:20] the most. Next, we have the microphone feature, which is basically how you um basically talk to Claude via your microphone. So, if I hold command D or just press this button, you'll hear those little two and then you can talk to Claude however long you want. And so, that's what I'm doing right now. I'm actually having a conversation with Claude via voice transcription tools. This is built into Cloud. Now, I also can use my own voice transcription tools to pump uh words into this little field if I wanted to. But you'll find that this tends to work pretty well. And you can actually see real time the text that you're

[36:52] generating being transcribed. So, that's kind of neat, right? In practice, I will actually usually end up prompting just through my voice these days. The reason why I tend to prompt through my voice is not because I think I am a better speaker than I am a better writer. Most people tend to be far better writers, but because when you uh use dictation, you can typically just generate a lot more in a short period of time than if you write, the average human being speaks at around three to four times faster than they write with their fingers. So, um you know, this is just a quick and easy way to like brain dump

[37:20] whatever you need into cloud. And I think you'll find that there are a lot of situations in which you'll probably want to brain dump stuff into cloud. I'll show you guys some as we build. Anyway, you can change your microphones and stuff like that right over here if you want. Then we have the actual model. So, there's Fable 5, which requires usage credits right now. There's Opus 5,

[37:37] which is not Sonnet 5, Haiku 4.5. Variety of different ones here, too, that are kind of old last generation ones. Uh, I'm just going to use Sonnet 5 right now because it is pretty smart. Does most of what I want it to do, and I'm not super worried about, I don't know, solving the mathematical formulas of the galaxy. Finally, we have the effort tab, which is basically a slider that allows you to um define how hard you want Claude to work on one of your

[38:02] tasks. This allows you to basically say like, "Hey, like is this work that I'm doing today super important? If not, I don't want you to think too hard. I just kind of want you to go with your best guess. Whereas, you know, if you go all the way to the right, it's like, hey, I don't want you to try guessing at anything. I actually want you to triple, quadruple, sex tuple check everything that you tell me." And in doing so, um, cloud will consume a lot more tokens, but the probability of a high quality answer goes up way more. It's also note also worth noting that the speed changes

[38:27] quite a bit here. Um, when something is at effort low, it usually works much much faster than when something is at effort smarter. So, keep that in mind. If you're going ultra code, it'll take, I don't know, like way way longer than if something is effort low. Effort low might take like 2 seconds. Uh, effort

[38:41] ultra code might take like 5 minutes. So, worth usually operating somewhere in the middle. I like defaulting to medium or high for the most part. Um, sonnet high tends to work pretty quick while also delivering pretty good results. Now, you can also see your context window here in the bottom right hand corner. What that is is the total number of things that are consuming your tokens. So, I just move my again fat head out of the way so you guys could see this a little better. Um, as you can see, we have a certain number of messages up here at the top that are consuming a certain section of our

[39:10] context window. Certain number of system tools, system prompts, MCP tools, skills, memory files. You don't need to know what all these things mean. What you do need to know though is that Claude can only hold a certain amount of information in its head. And that information is currently bounded by tokens. And so a token is very similar to a word. And pretty soon, I'm sure they will just be called words. But for now, they're just a little bit different. Uh right now, you know, sonnet 5, which is the model that I'm

[39:33] using, can host 967,000 of those tokens. We've consumed 112,000, which puts us as 12% of the way to the entirety of Sonnet 5's context window. When you hit the context window, usually what'll happen is Claude will compact your context, meaning it will just like try rewriting it so that it's way shorter and then stick all that information up at the very top. The reason why that's valuable is now you can continue the conversation, but a lot of the BS words are sort of taken out of your context and you'll usually get build less for it, too. And that does take me to the sort of billing and usage aspect. The

[40:06] longer your context window is while you talk with Claude, the faster you will burn through your usage limit. So, my recommendation is try not to keep a thread running all the 967,000 tokens uh that are available. I'd like working somewhere between 0 to maybe 500k or so. I find that's also my performance sweet spot. If you add more information than that, cloud starts to perform a little poorer. That's just because it gets confused. It doesn't fully know what it is it's doing. You've given it a lot of instructions at that point that it needs to juggle around in its memory and obviously that can be pretty tough. Okay, so that context

[40:36] window can be consumed of messages. As mentioned, system tools are just different things that it internally has access to. System prompt is uh basically a little snippet of text that you can define that will be prepunded at the beginning of every conversation. MCP tools is another sort of thing like a skill. It's just skills that other people host for you. So what you can do is you can load basically external skills via MCP tools that other people have created very easily. And then memory files are the things that I showed you guys earlier that basically define I don't know like in my my case

[41:05] my height earlier 6'2 and stuff, right? The rest of the stuff not super valuable. Obviously that's free space. We're just going to skip out of that and you have your plan usage limits. Okay, so I told you this would be comprehensive, didn't I? Yeah, we've covered uh probably 90% of the entire interface. There's just one more thing. At the top right hand corner, you have the ability to spawn a terminal. You're probably not going to do this, but for those of you guys that don't know, a terminal is just um kind of the way that we used to interact with computers back in the day. We had a variety of

[41:34] different commands that you could, you know, access cdls, print working directory and stuff like that. You could see all the files that are in this directory and so on. Um, but anyway, sometimes cloud actually needs to use this terminal in order to do things. We also have the ability to check your diff, which is just the changes to the code that you have made. If you're developing an application, that can be quite valuable. You also then have a browser. So we now have a browser built into cloud that can do things like access web pages for you. That's pretty valuable as I'm sure you can imagine

[42:02] because you can use that to automate browser activities. And then finally at the top right hand corner you also have a couple of things that are pretty cool. One is an iOS simulator. So you can actually set up like apps in cloud code. Now um these are iPhone apps and they're fairly cool. Um you can like download

[42:19] you know Xcode which is how you do it. Open up the Mac app store and so on and so forth. I won't do that just for uh my own sanity. I don't like looking at the app store any more than I have to. Uh but anyway, you can you can create like iPhone apps now, which is pretty badass. Uh and then you can also open in whatever you want. You can open in a new window, VS Code Finder. You can rename your transcript. You see this Nyx Drive web page right over here. Well, if I go to the top rightand corner and I click

[42:43] rename, I'll say Nyx demo. You can see we've actually fundamentally and foundationally changed that. Uh you can also see the transcripts. you can actually like see all of the thinking that Claude has done if you just expand this transcript view. So, in my case, I'm just going to do summary. And what it'll do is it'll just summarize everything that we have done, our whole conversation back and forth. I'm now going to get a summary, which is pretty cool. Uh, as well as a full transcript of the conversation at any time. So, transcripts over here, summaries over here. This is going to take a fair amount of time to spawn, unfortunately,

[43:14] so probably won't happen entirely real time. Okay. And yeah, now you can see it's giving me a summary of this conversation. First, I wanted to build a self-contained publish artifact page profiling Nixx as a personal bio portfolio page. And then uh I don't know there's a bunch of things like the outcome, you know, what had what ended up happening in the session. Well, the next drive artifact is now done and sharable, right? Then you can fork a conversation where you just start

[43:37] another conversation at the same place. So that's me forking the conversation right now. And then you can also archive a conversation if you don't want to look at it anymore. So that's what I just did to that Nyx demo fork. And then uh you can also delete a conversation if you want to. That's what I just did to

[43:50] Nick's demo. Okay, we are now done. Virtually every button in this godforsaken interface. Just kidding. Uh there's a lot and that's because the Anthropic team has added a tremendous amount of value to their app in just uh I don't know maybe 3 or 4 months. So the app was not always like this. The app is very high quality now. The app will probably look a little bit different from what you guys see just because they are consistently and constantly updating it. But now that you guys know how to use the app, you have no excuses. we know every little literal button on this entire interface. Um

[44:20] it'll be very very easy for you to navigate and do virtually everything that you need in order to make something cool happen. Which takes us to learning the interface right here which is now check. We have an optional task here called set up the CLI. CLI in case you didn't know stands for command line interface which is just a different way of communicating with cloud code. As mentioned, you know, they've given us many different ways to talk to Claude and and work with Claude. Um, so what I'm going to do here is I'm going to I'm going to set up another one just as a demonstration. And I may use the CLI. I

[44:53] may end up using the cloud desktop app. Um, it's really just about what is simpler and easier for you. Before I show you guys that, it's worth a brief uh kind of piece of education here. Uh, just like we have different websites on the internet, okay, that are accessible through different browsers. Like in this case, this website Google can be accessed through Chrome, Safari, Firefox or something else. We also have the same concept of clawed code which is accessible through I don't know your desktop app, your terminal VS code or something else. And so there are variety of different ways basically to get to the same resource. And so what I just

[45:24] showed you here was I actually just showed you the desktop app, right? And so that's what we just looked at, but you can also use a terminal to do the same thing. You can also use VS Code. And so what I'm going to do now is I'm just going to show you guys what it looks like in terminal version. And it's very very straightforward and easy. All I'm going to do is I'm going to go to Ghost TTY. When I open up ghost tty, you'll see I have a terminal. Same thing that we had before. It's inside of a folder called business. That's just sort of the layout. And in order to access

[45:48] Claude, all I need to do is type cla-verbose. I like verbose because it allows me to see what it's thinking. And if I scroll and zoom way in, you'll see that, you know, our interface looks a little bit different. You know, our cloud is up here. And I don't know what what is it doing? It's looking up. It's going down a couple times, but it says Claude Code Fable 5 with medium effort in business. So, what's going on? So maybe I'll say, "Hey, um, what you'll notice is this is the exact same product. It's just I'm accessing the product now in a different way. Instead of me using that app that

[46:18] we saw earlier, I'm using my own app and I'm talking to it through a channel. If I type slash, you'll see that there are a variety of different skills that I could use as well. And so in my case, this is actually my preferred way of communicating with cloud code. And the reason why is because I just, you know, uh, uh, I I've done a lot more development stuff over the last 6 or 12 months. I'm just more used to communicating with it in like a terminal style interface. That said, it's not the only way to do so. And hopefully you guys see now that uh the cloud desktop

[46:45] app is a very very strong and viable candidate. And just for simplicity sake, I'll prefer that wherever possible. But there are some situations in which I I do like just diving into the terminal. Again, it's almost like a different car, right? Different browser, different car. Everybody has slightly different preferences about how they drive it. you could have the exact same horsepower, but obviously if you're accessing cloud code or if you're accessing the the same horsepower of the engine through a different model, um it'll react differently, handle differently and stuff like that. And that now takes us to the end of the first step, which was to download, install, and set up cloud

[47:16] code. Next, I want to cover the five marketing functions that we'll be automating today and the framework I'm going to use to show you guys how. So I want to talk a little bit about business because obviously you're doing this not for the sake of you know building up just a cool personal portfolio but ultimately we want to we want to build our marketing skills here with cloud code to do something some sort of economically valuable outcome and so the way that I see most businesses and it's a service business up here but uh really the more I think about it I think it's all businesses is as the following

[47:49] pipeline. All businesses will start with some form of marketing. Okay, that is just to get strangers to notice you and that is basically to get eyeballs on the business itself or the brand. Then you will have some sort of sales transformation event which is what turns these eyeballs and this interest into some form of money. Afterwards you have fulfillment which is where you deliver what you promised and then finally at the end you have some sort of administrative wheel that is what keeps the machine running. Now the whole idea here is basically you know during this process as you go from top to bottom you convert somebody who is a total stranger

[48:26] into some form of profit and you can imagine the same thing applies across a variety of different business stacks. You can do this with ecom. You know in ecom you have strangers that notice a product. You sell that product. they receive delivery of the product and you have some sort of admin to keep uh I don't know the purchasing and the inventory in line with like uh expectations and forecasts. You have hiring, you have vendor relationships, logistics managed with your your manufacturers and stuff like that. Uh and then you know ultimately these four working in perfect concert is what makes you profit. So you know this isn't like

[49:01] an MBA or anything like that. Certainly I don't think I could hope to do uh you know an MBA in this this this time period that I've allotted for this course. But you know I want to talk mostly about this marketing and then this sales aspect. Um and so this is really going to be like our focus. This is what we're going to spend most of our time working through in this course. Um but I think it's also important to understand that marketing and sales they work in tandem, right? It's not just one or the other. Realistically there's there's there's back and forth here. So a lot of the stuff that we are going to

[49:30] build for marketing today with cloud code um we will also sort of build with uh or build for our sales department at least to like understand and and know what's going on. And so this is kind of secondary and I'm going to have future courses that are focused specifically on sales. So you don't need to worry too much about that. Now the framework we're going to be using to really crush these

[49:49] cloud code automations is called race. race stands for reach, acquire, close and expand. And the differentiator or sort of the delineator between sales and marketing in my view occurs right here. So everything to the left of this is technically marketing and everything to the right of this is technically sales. Which means our course today is going to be focused on these two which is how do I get seen which you know if you think about the key performance indicators of getting seen how do I get impressions how do I get views so how do I get my brand in front of more eyeballs and then how do I turn these impressions and

[50:34] views into some form of intent so how do I acquire you know so how do I I don't know book meetings or get form fills get people to sign up to my newsletter get people to opt built in to some offer and so on. And so all of the automations we're going to be developing are really going to focus on uh these levers. Okay, it's all going to be about reach and acquire. And if you think about business in this framework and if you you know race to implement automations that help you with reach, acquisition, close, and expansion, typically uh you'll allocate your time a lot more effectively than

[51:10] otherwise. So, I'm going to come back to race pretty often. Uh, when, you know, somebody wants to work with Leftclick, my AI automation agency. Um, you know, typically they're like, "Oh, like I want to do this crazy admin automation. I want to do this backend thing." And I'm like, "Well, why don't we start with race? Why don't we start with reach, acquire, close, and expand, and then we can worry about that later." I find results are are through the roof. Uh, people focus on revenue and growth uh

[51:34] way less than they realistically should. Race is how I tie them back to, you know, revenue and growth levers constantly. And on the right hand side here we have close and expand. So you and if you think about it, this is like a sales transformation threshold. Um, you know, once we've booked the meeting or gotten some sort of intent or something like that for our product, it's not really a marketer's job to close them and expand them. Obviously, we can tee them up for, you know, the greatest spike of all time. Um, but that's sort of out of our hands at this point. So, we're going to work on

[52:01] optimizing reach and acquisition. And then later on I'll have courses on sales that'll also allow you guys to implement systems that improve your ability to close and your ability to grow the lifetime value of your deals. Okay. So if we think about this now a little more simply what does that mean for us? Well under reach there really three major places that I'm going to uh help you

[52:23] guys build cloud automations with today. The first is creative. Now this is the generation using artificial intelligence and orchestrated by claude of images, videos, audio and so on and so forth typically for the purpose of advertising although also for organic uh kind of brand reach and brand growth. So the cool thing is with claude if you think about this as like one of these analog dials on like a control panel we can pump this puppy all the way up to the top super easy. you can significantly reduce any bottlenecks at the creative stage just by throwing, you know, tons of uh a couple of different creative

[53:01] models to Claude and then giving Claude the ability to to press buttons and arrange prompts um you know, in its own way. I'll show you guys how to do a define system for this. This is the same system I've used at uh over a company that does over a billion dollars a year in revenue to massively increase their top of funnel and then ultimately the total number of pieces of content and ads that they can get out. After that, you also have copy which is sort of in

[53:25] tandem with creative for advertising. Um, but we can also use, you know, cloud code to customize a lot of the copy in our marketing mix. So, we could do the customization of newsletter copy for instance, which I think is a big lever not a lot of people are pulling realistically. You know, your newsletters don't have to say, "Hey, my offer does X, Y, and Z." You know what you can do is you can actually customize it to say, "Hey, Nick, just wanted to send you this because I know you talked about X, Y, and Z last month in your LinkedIn. You know, my offer does X, Y,

[53:54] and Z and I think it can help you." That sort of stuff does really, really well. And it also starts blurring the line between direct outreach, too. A lot of the time people think marketing and then they go, "Okay, this is all inbound. So, I make a brand, I make some sort of ad

[54:07] campaign and then people come to me." Um, well, you know, marketing can also be done outbound, which is where, you know, I have a product or service and I'm I'm actually actively looking for customers for it. And so, this is where a lot of my own experience comes in because I used to do door-to-d dooror marketing. So, I would literally acquire my leads by walking up to a business, opening the door, shaking the business owner's hand, and just giving them a big pitch. So, I'm sure you can imagine it didn't work well super often, but uh when it did, it did. And so, I have a

[54:31] lot of experience around applying AI, cloud code, specifically to the outreach stack. And I think it would be remiss of me if I did not show you guys how to do that as well. It's a very high opportunity lever for outbound growth. Okay, so all of these improve your reach. Okay, so that's sort of the R in race. Now on the right hand side here, um we also have a couple of, you know, ways to improve the total fraction of people that opt in or book an offer. And so the first one is speed to lead. Now what is this? This is basically just capitalizing on interest. you know,

[55:00] presumably between the reach and the acquisition stage, if you think about it here, there's some sort of like um opt-in, which is where somebody will fill out a form or something like that to to demonstrate interest. You would be surprised at how many hundreds of millions of dollars are lost between these two stages. Um if you just, you know, respond to leads really quickly, which is what speed to lead means. If you build systems that allow Claude, let's say, to orchestrate a conversation within 30 seconds of somebody filling out a form, you can make companies 300 to 400%. And um that's what the hat that I'm wearing uh right now represents. I

[55:32] run a company that actually just does speed to lead all day. And we've produced crazy outcomes like taking a home services company that was previously doing around $3 million a month in sales. Um literally just by implementing the greatest speed to lead system ever. We took them to $9 million a month. That's 300% increase in total revenue simply by optimizing one of

[55:52] these, you know, six levers here. Okay. Another is booking percentage. So we can actually improve booking percentage through speed to lead, but also through economizing um you know a lot of the form fills and a lot of the systems that people use right now to to gain interest. So we can have Claude, for instance, do a lot of research on the prospect to eliminate the total number

[56:11] of fields we need in an opt-in form. Typically, if you have, let's say, 10 questions in an opt-in form, you break that down to three or four questions, you'll see a significant improvement in booking percentage. You do so at the cost of qualification, but Claude Code can now do research on the prospect for you sort of autonomously. um which means you don't actually need all that much information. Then also following up. So this is like nurturing. Um what we can do is we can use cloud code to orchestrate autonomous follow-ups and then do so in like really customized ways that people seem and feel like they're are being treated like as people

[56:40] not as dollar signs which is kind of funny if you think about it because like you know we're using AI for this stuff but the end result is our prospects that people are marketing to are a lot more likely to feel like we're talking to them directly. If you think about it, one of the reasons why many companies have been forced to treat people like dollar signs, you know, not personalize their outreach, not even personalize their their billing or their communications or the templates of their invoices and stuff like that is simply because we haven't had the manpower to

[57:07] do that and provide a good experience. But what I really like about cloud code and AI is that it allows us the ability to produce the manpower to give every single person that works with our business a customized experience. They'll have customized creative, customized copy, customized outreach. Their speed to lead will look really customized. We'll we'll customize, you know, information by doing research automatically on people to enrich, you know, fields in our newsletter databases will u customize follow-ups. Everything will be super super unique and and also very personalized. Before we actually build things, you need to understand bottleneck thinking, just sort of constraintbased thinking. And this is

[57:44] quite important in marketing because what you'll find is a lot of the companies that you work with or maybe your own company um they have a lot of one of the things that we talked about. Okay? Maybe they have a lot of creative, maybe they have a lot of capacity around the follow-up end, but you'll find that there are certain parts of their funnel that just really suck. And the way that bottlenecks and pipelines work are the entire operation is always constrained by the narrowest step of that chain uh or like the weakest link in the chain is another way of thinking about it as

[58:19] well. The classic example is a factory. And for those of you guys that have played those like I don't know factorial games or whatever um you'll intuitively understand this but let's say I have a factory and the way it works is every day I get 100 units of some raw material. I don't know let's say it's a rock or something like that. And now this enters my big factory. Okay. And

[58:45] this is what my factory looks like. [sighs and gasps] And basically what happens is you know what we have to do every day is we have to put this raw good through two steps. We have step A and step A has a capacity of 100 units per day and step B has a capacity of 10 units per day. How many units in one day can we get through this pipeline? How many units can we get through this factory? Well, if we start with 100 units on the left side, okay, and then we pass it through A, A has a capacity of 100 units per day. So, what that means is the

[59:32] intermediate between, you know, A and B, like we'll we'll generate 100 units. That's fine. You know, we've made it past the first step. You'll see B only has 10 units per day, which means, you know, despite the fact that we're feeding in 100 units, we can't get all of them. Um what we can do is we can transform 10 units per day. What that means is what is the sum total daily output of the factory? It's 10 units per

[59:53] day. You know basically the entire factory can only work as quickly as the weakest step in their process. Now if I give you an mission to improve the total capacity of this product you know in total number of units generated per day um you know and then I say hey like which step would you improve in order to increase the total output of the factory would you improve step A or would you improve step B you know let's say I allow you to double any of these well if you doubled A you'd be at 200 units per day what would the total capacity of the factory now be well you

[1:00:30] wouldn't actually have changed anything because the bottle neck the thing that all other steps are sub subservient to is still 10 units per day. However, if you changed B from 10 to 20 units per day, then the total output of the factory would go up. Okay, so anyway, this is just a very simple and straightforward um analogy hopefully that allows you to think about how to optimize a marketing process. The reality is, you know, we have a lot of processes in our business that are very similar to what I just talked about. And in this hypothetical example, what we have is we have um you know a lot of

[1:01:02] capacity at the creative step. We have a lot of capacity at the lead step, but we don't have a lot of capacity the booking step. So this is sort of our you know uh example that is analogous to 10 units per day. It doesn't really matter if we improve our creative in this example. Doesn't really matter if we improve our follow-ups or the total number of leads we have. you know, none of that's going to matter because the pipeline, the narrowest part of our whole business is this booking stuff. So, what that means, too, is it's not bad. It's actually quite freeing. Um, we can spend zero

[1:01:34] work on creative leads or following up at this point. If we just improve our booking by a tiny little bit, our entire business will improve by that same multiple. For instance, if we can move it from 10 units a day of capacity to 20 units a day of capacity, what we've done is we've effectively doubled our business. Okay? Then we do it 30. you know, maybe we'll triple our business and so on. And so I I just bring you guys this analogy because I find a lot of the time people want to rush to automate literally everything that they can in a business. And despite the fact

[1:02:05] that that sounds nice and it's sexy, you'll make way more money for yourself or for the companies that you work with, if you spend a little bit of time trying to figure out what the bottleneck is and then building systems that alleviate that bottleneck versus just trying to do everything everywhere all at once. So all the systems that I'm going to show you from now on, I would like you to

[1:02:22] think of as solutions to bottlenecks. But don't just throw these systems at the companies that you work with because that isn't necessarily going to do anything unless that is actually the current constraining step. Like if creative for instance was a very narrow bottleneck in this instance, you know, if it looked maybe something more like this, right? Um and it was a lot more narrow, then widening it with the creative system that I'll show you guys in a minute would make sense. You know, if booking was really the bottleneck as mentioned here, then widening it, widening it with a booking system and doing more speed to lead, that would

[1:02:53] make sense. But if you guys are already really really good across the stack and everything that you guys are attempting to process automate is already sort of not the bottleneck in a process, you're not actually going to see uh much monetary return. Okay, so that takes us back to the five marketing functions we'll be automating today and the framework to do it. Why don't we now talk about how to prompt versus build skills versus schedule loops versus create routines? And once we're done with this, we can actually start the building of our cloud code systems. So, a simple and easy way for you to think about how we are going to progress

[1:03:27] through this course is for all of the systems that I'm going to be building with you guys, we're going to start them very simply as prompts. Now, prompts are sort of the first step in the automation pipeline. What they are is it's kind of like uh an old school Flintstones car where you know you would pick the car up and then in Flintstones they would sort of run with their feet because they

[1:03:48] supposedly hadn't invented wheels yet. It is manual every single time that you do the process you need to be in the driver's seat actually steering it and asking the AI to do the next step and getting the AI to you know do things in sort of a self-directed manner. And as a result, it's also quite slow. But obviously, I mean, if it's the first time you've done a process, how else are you going to do it? And so when we build all of our automations in our course, we're going to do them primarily through prompting to start. But after a while, obviously, business processes tend to be

[1:04:23] quite similar. We tend to just do the same thing over and over and over again, right? That's kind of how business works. We find a hack in the market that pays us more money than the work and money that went into it. And then we just repeat it over and over and over again. So we'll transform our prompts to the next stage which are skills. Now you can basically think about skills as just a saved and much more efficient instruction set that you can reuse. And so we showed you guys how to do that

[1:04:49] earlier through slash commands, right? There was that morning skill. Well, you can get really really granular with skills. And once you have a big set of skills, you can also just run those skills as a meta skill. uh and in that way chain together you know a process that might have previously taken many many man-h hours into one. So for example I have a bunch of skills that manage my YouTube channel and a lot of people wonder how I'm capable of self-managing such a large output. Well the reason why I can is because I have skills that manage everything for me now from the generation of title candidates

[1:05:18] to the generation of thumbnails to the generation of uh you know descriptions to the actual publishing of the content the selection of the keywords. It automatically adds things like uh end screen cards and so on and so forth. You know, this is stuff that previously every time I wanted to publish or record a video, I would have had to, you know, sit down and probably spend an extra hour or two uh doing. Well, now I I don't need to do any of that at all. And once I built all of those granular indivi individual skills, I then created master skills that just ran all of them

[1:05:47] in sequence, sort of one skill after the other, skill after the other skill, and in that way chained together small tasks, turned them into big ones, and eventually replaced entire roles. Okay. Now after the skills which are themselves quite powerful, we are still technically constrained by you running the skill. And so that takes us to the next step which is the loop. Now what a loop is is it's basically a skill except it's a skill that runs on its own without your involvement. And so in the example of, you know, that big like meta um skill that I talked about that like let's say manages my YouTube channel,

[1:06:20] you can imagine how me having to run the skill at the end of every big publishing session or whatever is kind of a pain in the butt, right? What if instead I just set a loop that automatically ran the skill on my computer once every day? And maybe, you know, my deadline was 12:00,

[1:06:34] so I just had it run every day at 12:00. Well, in that way, I'm no longer really in the driver's seat. What I've done is I've prompted built up a saved instruction set and now I have a scheduleuler that's just running that instruction set maybe on a set of resources. You could see that with like the morning brief idea, right? You could just set the morning brief to run completely autonomously without you maybe at like 5:00 a.m. every morning or maybe you know you're doing some sort of marketing role and you have to scrape a bunch of um competitor ads. Well, you can actually set a loop to scrape the

[1:07:04] competitor ads every morning at like 5:00 a.m. before you get into the office or before you start working on your own business. And that way, the second that you show up, you will have a list of all of the competitor ads. All you need to do logically is just figure out, okay, which ones am I going to try and duplicate and swoop. You could build loops like this across virtually every part of your marketing stack. And I will show you guys how to do so. But obviously, the first step here is you need you need the skill. And in order to

[1:07:26] build the skill, you need the prompt. Okay? And then once you are done with the loop, sort of takes us to the last step, which is the routine. Now, loops are great. The only issue with loops is that they need to be run on your computer. And so that that decreases your ability to share the loops with other people in your organization. And it also means that the loops are kind of brittle. If something changes in your computer, if maybe you change your I don't know login or whatever, um you change the plugins that your clot has access to your skills and MCP servers and whatnot, the loops can break and

[1:07:58] they cannot work. But if you really want to like expand things and make them you know institutional and actually add things to the organization not just your own work space uh you need to take them out of your computer and then put them on somebody else's computer. So if you guys remember we covered routines earlier and uh you know like the difference between local and cloud environments well routines just run in cloud environments and the idea is you will pay you know anthropic or another cloud provider some amount of money usually very little to be clear. I mean, we're just spending money on tokens at this point. And then it will do the task

[1:08:30] and then it will output some deliverable, which is, you know, a lot more money. And so, maybe this is the simplest ROI of your life. You spend, you know, $1 in tokens, then you get the equivalent in deliverables of um, I don't know, let's just say like $100 or something like that output. Um, it is entirely self-managed. Uh, you don't have to touch anything on it. It just occurs without your involvement. And so this is eventually where we are going to

[1:08:55] go with all of our systems to be clear. Um although you do need some sort of um input to that like a schedule, API, web hook, trigger, whatever the heck you want. Um but I'm just going to be approaching every single one of our steps through this four-step pipeline. So you guys are going to see that in our very first u automation, which is going to be using Cloud Code to create ads essentially wholesale. Uh, we're going to start by building out a step-by-step prompt where I walk Claude manually through a process to go and do what I want to do. I'm then going to convert

[1:09:25] that into standardized instruction sets and then show you guys how that runs over and over and over again. We're then going to run that a few times to make it perfect, make it very low likelihood of failing. Then we're going to convert that into a loop that works it without us. And then finally, we're also going to convert that into a routine. And I'll show you guys pricing and and stuff like that. A couple of simple rules. Uh, in general, the moment that you find yourself saying the same thing more than once, okay? I would turn it from a prompt into a skill. So, you guys are

[1:09:51] going to see me do that as we build all of these and get really, really hands-on and practical rather than you having to do the same thing over and over and over again, like saying act as an expert social media marketer and generate five LinkedIn post ideas about AI automation. You know, we're just going to save that as a skill and then just, I don't know, type ad batch or something like that and then do it. But um in general, you can

[1:10:12] also just do this like across the stack. Anytime that you find yourself asking AI to do something more than once uh and if it is like virtually the exact same ask, you should at least expend a tiny bit of time and effort to see if it is even turnable into a skill because you know

[1:10:25] not all things are turnable into skills. Um some things do require a human being in the driver's seat. I mean if they didn't, thank god we'd be essentially useless tomorrow. Uh so the idea is what you want to do is attempt to turn everything that you do often into a skill. see if it passes that, you know, totally autonomous barrier or not

[1:10:42] threshold. If it does, now it's a skill. It doesn't really concern you. You just run it. And then if not, you know, you're still in the driver's seat. In this way, you can very quickly and easily determine all of the tasks that you yourself are not necessary for in the business or in your marketing role and then ascend up through the ranking of abstraction and leverage and get to the point where, you know, you're overseeing a fleet of agents that do a bunch of economically valuable work for you as opposed to actually having to sit down and do everything yourself. Okay, so hopefully you guys are as excited as

[1:11:10] I am. Um, this is just a final review of the concepts. Prompts are triggered by you. Skills are triggered by you. Loops are a timer on your computer. And routines are on the cloud. Uh, prompts run when you type. Skills run when you press run. Loops run while your laptop is on. And routines work regardless of whether or not your laptop is on. Uh, prompts are great for quick

[1:11:32] experimentation and building things out. skills are great for repeat tasks that you still need to be in the loop for. Uh, speaking of loops, loops are best for monitoring things that automatically scrape a resource and stuff like that. And then routines are best for daily production. Um, and I think we've now done more than enough sort of explaining about these concepts, we can actually go build something. So, I'm going to head back up here and then just mark that puppy as done. And now we can build some

[1:11:58] actual creatives with Claude code. Okay. So, how are we actually going to put this creative generator together? Well, the first thing I want to dispel you of is the notion that you can just say, "Make me some great ads," and then it'll automatically work. Um, you need to be a lot clearer about what it is that you want the model to do. You also need to be a lot clearer about how you want the model to do things. Like, if you just say, "Make me great ads," you'll see that the model will instinctively just try asking you a bunch of questions to narrow down what it is that you want to

[1:12:24] do. Um, but that's not really enough. even if you answered all these questions, it take you on a wild goose chase and it probably wouldn't get anywhere despite the fact that it sounds like it knows what it's doing. You know, it really doesn't. So, my recommendation is you need a process that allows you to create ads already. Uh, and ideally, you

[1:12:39] would already have a process, right? Like you're not trying to come up with a new process today. What we are doing is we are taking a pre-existing process and then we are turning it into or we are basically process automating it. And sometimes we'll have to shift a few of the steps in a process in order to make it more amendable to automation. like we'll need to provide some sort of input data and so on and so forth. But um in general uh you know don't come to AI without a process. Try and sit down and

[1:13:03] think about the process yourself. Actually like whiteboard the process cuz if AI just picks the process for you um not only is it picking the process itself, it's also carrying out that process and then maybe it's even evaluating the results of that process. What you're doing is you're multiplying probabilities. And in general, if you multiply probabilities where somebody else is in control of the quality of the

[1:13:23] output, the end result goes way down. Um, just to, you know, demonstrate for you guys what I mean by that. You know, if I go back to, um, my little drawing pipeline here, if I had a couple of processes that were 0.8 aka, you know, 80% likely to work. Let's say I had three steps in my pipeline. Okay, that's 0.8 time 0.8

[1:13:48] 8 * 0.8. Um, let's say we had, you know, step one, step two, step three. Sure, the probability that step one resolves correctly might be 0.8. The probability that step two resolves correctly might be 0.8. And the probability that step three resolves correctly might be 0.8. But if you do the math on all of these, okay, the total result is not 0.8. the total result is 0.5 or basically 51.2% success rate. So what you want to do is you basically want to minimize the total number of steps involved in the creation of a good prompt, a good skill or a good process. And the way you do that is, you

[1:14:30] know, rather than have AI come up with the process for you, then rather have AI actually maybe judge the quality of the outputs for you. You want to do both of these yourself. All you want is for AI to do the thing in the middle, which is actually like carry out the process. And in that way, you can go from like a 51.2% success rate, maybe up to an 80% success rate. Okay. Also worth noting, you're probably not going to achieve 100% success rate with any sort of AI agent process, at least not of as of the time of this recording. Um, really, the way that AI works is it allows you to

[1:15:01] take a lot more shots at the net in any process, and then you can apply your own human intellect to just picking a bunch of the correct shots, if that makes sense. So, for instance, with our ad generator, what we're going to do is we're not going to aim to like make this thing just pop out perfect ads every single time. What we're going to do is we're going to have it generate dozens, hundreds, thousands of ads, okay? And then all you're going to do is you're just going to review them and then run them. And in that way, uh we're going to be able to generate a lot more top

[1:15:26] offunnel ads, which will alleviate the creative bottleneck, assuming that that is the bottleneck. So, anyway, this is the process right here that we're going to be um going through. We're going to save a bunch of high performing ad formats as templates. Then we're going to feed in a variable. So a niche, some sort of offer, some sort of angle,

[1:15:41] whatever it is for whatever product. I'll just come up with a hypothetical one. Claude will generate a bunch of different combinations of niche times ad format, offer time ad format, angle time ad format. And then we're just going to review, pick, and ship. Um, staying in the loop, uh, basically all the way up

[1:15:56] until the end where the taste matters. But actually before we even get there, I I hope you guys know just based off of that bottleneck thinking, um the only situation in which this would make sense to apply to a business is if creative is your bottleneck. Like a lot of companies are big on ideas and they also have the means to distribute those ideas via paid or organic. They just can't actually make the ads themselves. So this is this is the situation in which that would make sense. If your company has lots of ideas, if it has lots of distribution, but it doesn't have a lot of creative,

[1:16:24] um you could actually meaningfully improve the business significantly. um you know so if you're currently bottlenecked on designing and stuff like that like this the system will genuinely improve things but if you already have the ability to uh I don't know like make a bunch of ads or maybe you're really short on ideas or maybe you're really short on distribution you having the ability to create more ads won't necessarily improve things really all that much. Okay, but let's carry out this process. So this is something I sat down and I thought a little bit about because obviously I wanted to spend our time effectively. I didn't want to just

[1:16:54] like do a bunch of whiteboarding on ad processes. A lot of people have way better ad formats um and and ad systems than I do. Uh and I'm not like a professional, you know, advertiser by any means. I used to run a PPC company, but it wasn't super successful. I mean, the PPC company probably capped out at like 200,000 in a year. Um you know, I run ads for my own products now, my information products, and then my

[1:17:14] education products, but um I don't know. I mean, like 5 to maybe a 10x rowass was considered okay, but it's probably not considered the best. So, I just want to give you guys like a cookie cutter process that is probably above average in terms of performance, but I also want you guys to know that you guys all have your own internal and successful marketing mix. So, you guys should do whatever the heck you want there. Just map out the process ahead of time and follow the series of steps that I'm going to show you to convert this into, you know, something that is replicable because the first thing is we need to

[1:17:41] find a bunch of high performing ad formats and then save them as templates. And so, ideally, you know, in your company, you would already have high performing ad formats. Okay? Okay, so you'd actually already have a bunch of things that are very good. They've worked before. You have high rorowass, high uh I don't know, LTV to CAC ratios. You know, you'd already have a bunch of these. We don't. So, I'm just going to go and I'm going to find a bunch. And I don't know what the performance of these are, but you know, I need to basically show you guys the process of getting

[1:18:08] data into the system. So, bear with me. What I'll do is I'm just going to open up like Facebook ads library, and then I'm just going to look through a bunch of ads that people are running for a variety of products. So, I'm in Canada, which is why it says that I don't want to stay in Canada, so I'm just going to go all. And then I'm also going to go all over here. Uh, we don't really care about the political ads, so we're just

[1:18:27] going to avoid those. Uh, I don't know. Keyword. I'm just going to do um let's see. Looks like I searched out Stripe, Maker School, Maker School, Survive, Nick Surive. That's kind of funny. Um, I I I like Stripe. Why don't we take a look at how Stripe is running their ads? Maybe show you guys a cool pipeline to do some poaching. Um, okay, cool. And so you can see that they actually have a pretty solid format or a pretty solid template set up already. Uh looks like there's a variety of different languages that they're advertising in. They're doing things in French. They're doing

[1:18:54] things in a variety of different ones. But you can see that they have a pretty straightforward format. It looks like what they have is they have like a highquality almost like gradient image that fits their brand colors. They have like their logo and then they have text that's sort of placed just like this. Uh make your you know company in America or or whatever. my French ain't so good, but something like that. Um, they also have another format here where they have like a cool gradient sort of background and then they have simple text right over here with a start now button. Looks like they have some videos as well. If I

[1:19:23] just play this without the audio, looks like they have a very simple video. Offer your customers your preferred way to check out. Okay. And then they also have a bunch of other languages as well as well as people that are advertising with Stripe. And so I actually did a video with Stripe a while back. Um, I don't know where that is, but anyway, clearly it's not here. Uh but anyway, this is a pretty good I think like first step just to show you guys how to automate really simple like image creative like it's it's super

[1:19:48] straightforward to automate this stuff. So maybe we'll start with that and then uh we'll get progressively more complicated. That probably makes sense. Okay, so we you know obviously need to grab the ad itself. So how am I going to do that? Well, I'm going to try um you know opening this ad detail and just seeing if I can get it as big as possible and I can right here. Okay, and you can see the images. This is an image ad. That is basically I don't know if I open up my little drawing tool. It's basically this is the bounds of the image, right? Uh probably up to about

[1:20:15] here, then up here. And so this is sort of what we need to feed into our model. Um if I go to the next ad here, like we'll just do why don't we do three to start. Um it looks like this is actually opening this little JavaScript segment. So why don't we actually go to Finder and I'll just make a folder and I'll start dumping all of them in now. And I'm full screen, so I just got to unfull here. Okay, we'll go downloads. I'm just going to make a folder. And why don't I call it um why don't I call it add formats for claude. Okay. And what I'm

[1:20:48] going to do is I'm just going to zoom in because I want this creative to be as high quality as possible. Obviously, it can't get super perfectly high quality, but we can get pretty close. And then I'm just going to take a screenshot and bump down to maybe here or so. This is going to be an ad format that I'm going to attempt to replicate. It's not going to be perfect as mentioned, but uh we will attempt to replicate it and then

[1:21:07] spin it. Kind of do our own style. Okay. And then next up, I want to see the actual ads themselves. So, I'm just going to click on this one here. And then you can see, same idea. I'm going to zoom way in. And it looks like this one is specifically for mobile. So, it's

[1:21:22] a little bit taller, but that's okay. This should natively kind of figure out that I am attempting to run a little bit taller of an ad. Going to paste that in as well. Why don't we do just one more to start? and then we'll get more complicated later on. Uh, we need a good format. Yeah, there I am. We need a good

[1:21:40] format. Create a link. Sell anywhere. Okay, cool. I like this one. This is pretty solid. I mean, these guys obviously they crush it with ads. They're um and they're they're really really pretty, too, which I like. Then why don't I take this little photo? Okay. And then go back to my folder and then stick that in there. Cool. So, you would fill this with whatever your high quality ad formats are. Um, you know, in my case, I'm just I'm assuming that Stripe is performing well with these ads. I don't know for sure. Uh, obviously, you don't want to make a bunch of shitty ads, but the idea is

[1:22:08] fill this with like the best formats ever. And then we're just going to spin these formats as much as humanly possible. Uh, okay. So, once we're done with that, we can go back to draw. And I'm just going to pretend now that I'm done with this. The next thing we need to do is we need some sort of variable angle to feed in. So, here's where, you know, you need a bunch of information about your business. So in my case, you know, uh I'm running uh a couple of different products. I'm running like a um Claro, which is a software as a service product. Uh Clarvo is basically

[1:22:36] a dialer, which improves your ability to connect and then have people on the other end of the dial actually pick up. Also automates the process of dialing. So you can dial multiple people simultaneously, call 10 people, and then just have the one that actually picks up routed to an agent. All of these minor optimizations, it significantly improve the total number of people you can dial and the total number of connects you get. Um, I also run a product called Maker School. And Maker School is an AI automation program that basically

[1:23:01] guarantees you a customer in 90 days. And so these are things that I can market. And uh, you know, I always like making courses where I'm marketing like my own products because I know the most about my own products and I can, you know, actually talk about the benefits therein and stuff. So why don't we try doing some ads for Maker School? I like that idea. Pretty straightforward and pretty simple. Maybe I'll do Claro ads later. Claro obviously being a software as a service is going to be different from like an e-commerce product that people can hold in their hands, but that's okay. Um, so what I'm going to do

[1:23:28] next is I'm just going to voice dump a bunch of stuff into a cloud and then I'm going to have it come up with, you know, niches, offers, and angles. And actually, just to get ahead of things, I'm also going to take a screenshot of this picture right here. Okay. Then feed that directly into Claude. So what I'm going to do here is I'm going to see if I can clear this. So we're back to a base conversation. I'm going to paste this in. And then I'm going to press this little voice transcript button and then just talk. So I'm going to press

[1:23:53] and hold to record. [gasps] Hey, my goal is to create a creative pipeline skill. Before we create a skill, I'm going to prompt with you back and forth to determine how best to structure the skill and then uh you know eventually arrive at a set of highquality outputs that we like. So here's the situation. Um I have a bunch of ads in a folder called ad formats that I'm going to feed you. What I want you to do is I want you to apply a bunch of information about Maker School, which is my AI automation program that helps people get their first client in 90 days

[1:24:27] or their money back for a service that sells artificial intelligence or is enabled by artificial intelligence. You can think of this as like systems that help create ads, the very system that I'm getting you to make, uh, for instance. So, the way Maker School works is people will come in and then I will coach them day by day in order to help them get their first client. They also watch a bunch of videos on a daily basis, uh, usually three to four videos where I actually show people how to do the thing that they're supposed to do that day. So, I don't know, setting up their cold email boxes. I'll actually

[1:24:59] record a video walking them through that whole process and then it's their turn to do it. Uh, it's also a big community. We give a bunch of discounts and free resources, around $21,000 in discounts at the time of signing up. And uh yeah, if you need any more information from me to, you know, produce niches, offers, angles, and that sort of thing, just let me know. My goal is I want to recreate the ads that um are in the ad formats library. So, I'm just going to send all of that to you and then I'd like you to take that information and then use it to

[1:25:26] scaffold out virtually the exact same ads. Um I know that it's not going to be perfect to start because we're going to do so procedurally, but u we're going to go back and forth until we get like the highest quality um ad version or variant as possible. Um, last thing is we can't just use, you know, Stripe's logo or whatever company's logo that we're feeding in as an ad format inspo. So, uh, because this is a demo, I just went

[1:25:48] and I found like a random set of ads. I'm not using my own ads. Uh, so we're going to need to replace that with Maker School. But aside from that, I want very similar styled ads. I want them to look quite similar. Um, we just need to change sort of the image up at the top. Instead of the stripe gradient, I want something different. And then we'll turn this into a repeatable engine to generate dozens if not hundreds of these ads uh very quickly. Okay. And then what I'm going to do is I'm actually going to go and feed in this folder too. If you guys remember, you have the ability to

[1:26:14] feed in a folder. So now I've given it the folder. I've also given it a big prompt. And at the very end of it, I've also given it an image. And hopefully you guys can see I'm just talking to this thing. Right? So what it's doing is uh you know it just did some thinking. It's now found the screenshots. It's

[1:26:29] actually going to take a look at them. And uh we'll go through and see sort of where we land. Okay, so it's breaking down the pattern. Bright abstract gradient, wave graphic, bold short benefit driven headline, tiny wordmark logo, single CTA button or text, high contrast, very buildable. Answering my question for free, do you have an actual Maker School logo file or wordark I should use or should I type set Maker School in a clean senset here? And I'll say yes, type set for niches. Okay, what I'm going to do is

[1:26:58] continue answering these questions. Yeah, in general it's people that want to quit their nineto-5, freelancer agency owners that want to add AI services, uh, and a lot of marketers and ex-corporate professionals. So, just create a bunch of different niche sets based off of this this audience, and uh, we'll vary them. The whole idea is you're going to help me generate way more ad variants than I could ever possibly do myself, and then I'm just going to select the ones that work. uh for offers, 90-day money back guarantee to land a first client, $21,000 in bundle discounts and resources on sign up, and then also include the fact that

[1:27:31] I'm there. Like, I actually show up in the community every day and respond to almost 100% of every post. Cool. So, now I'm, you know, giving it a bunch of information about that. And now it's actually going through and building a template spec. Now, you guys should know that Claude does not have an inherent ability to generate images. So, what we're doing this time is we're actually just building out a a series of templates. Okay. And these templates are basically rules that allow Claude to generate um these things on demand using very cheap old school image processing platforms like image magic and stuff like that. So, these aren't going to be

[1:28:04] like the highest quality things right off the bat. We're going to have to massage them a little bit. But, the idea is at the end of it, we're going to have a pretty repeatable template and then anytime we want to come up with a different ad, boom, we're just going to like say, "Hey, you know, come up with 400 ads." And it'll do. So the benefit to that as well is it's very cheap. You know there are image generation platforms out there which I am going to show you guys how to do after this. Um but these image generation platforms also cost a fair amount of money. Uh if

[1:28:25] it costs me $2 let's say to generate an ad and I want to generate a 100 of those. That's $200 every time. And while $200 every time for something that is a top offunnel growth lever like ads might not make that much difference to like a $10,000 a day business. You know it's a fair amount I think for most people that are probably watching this. And yeah, I wanted to make sure that it was about as accessible as humanly possible. Okay, so what it's doing is it's actually generating an HTML template to show me what the ads are going to look like. Um, and then, you know, after this is done

[1:28:54] in presumably another 20 seconds or so, I'll show you guys the results. Okay. And so it's actually come up with an example here, and it's done so underneath test stacked.png. And you can see that it did an okay job. I mean, you know, it came up with its own gradient here, which I think we're going to have to significantly improve. Um, but you

[1:29:10] know, it did kind of space out the font. Well, it does have a little start now button. Um, I would say we're just losing a couple things on the spacing. Um, you can also see there's this little stacked overlay where we have a much smoother gradient down here with a similar sort of card. And, uh, yeah, you know, this isn't perfect, but it's going to be pretty clean. So, what I'm going to do next is I'm just going to give it a bunch of feedback on this and see how well it did. I'm also going to have it generate me a web page that I can very

[1:29:34] quickly use to uh, make changes and alter things. So, first thing I'm going to do is I'm just going to hold command D because that's how you record. Hey, this is an okay job. I thought we got maybe 70% of the way there, but we do need to significantly improve the design and make it much closer to the initial stripe one that I showed you guys. Um, so we do need to push the CSS further with a variety of different gradient styles. Um, we could actually generate high quality gradient background images as well. Uh, and then, you know, do a bunch of flips and stuff like that. I'll show you or I will use

[1:30:03] an image model later. For now, I want to do things entirely procedurally. I also think the text spacing is a little bit off and uh yeah, I just want to significantly improve the quality of this first. So, in order to do so, could you create me a web page that I could use to adjust various sliders to change things until the settings are exactly how I want them to be? Uh, think about it as like me placing something on a on a page, let's say. Uh, build me the ability to, I don't know, change the font width and the different styles and and and where they're placed on the page

[1:30:28] and so on and so forth. Um, after we're done, we'll save those and make those concrete. And then once they're concrete, then we can spin, you know, hundreds if not thousands of these um to, you know, generate a bunch of ads. You see it's now actually building an SVG filter pipeline to try and reproduce what we're seeing here, uh, which is some sort of smooth meshed design that obviously Stripe used their own internal image model or something like that to generate. So, we're going to try doing this procedurally and programmatically just to save a bunch of time and energy and money. Okay. And now we have a

[1:30:57] clawed browser, which it's going to open just to allow me to actually tune the ads. So, it just opened it over here on the right hand side as you guys could see. It doesn't look like I'm actually capable of seeing one of these ads just yet. So, I think it's probably going to have to do a little bit more work to tighten that up. Now, one thing you'll find when you just have Claude start a task and then, you know, give you the results is it'll self-manage a lot of errors and debugs. In the past, this didn't happen. In the past, you typically had to do a lot of this

[1:31:24] debugging yourself. But you'll see that it's just passing its own outputs back in as an input to the next step and then progressively getting closer and closer to what it is that you want. Um, we're also coming up on like 3 minutes here. Uh, there are a lot of prompts that I run that now genuinely take somewhere between 15 to 20 minutes. And so I will fire off a task by dumping in a bunch of my demands and want and then I'll just step away or, you know, I'll work on something else. Or what I'll do is I'll just spawn a bunch of these. And then

[1:31:50] eventually this little icon here will go from gray to yellow. And the second that it's yellow, I know that I, you know, it's done and I need to come back in and actually check in with it. I think sometimes that also turns green. We can actually see the filter debug right here. It's actually coming up with a variety of different filters. Um, I think what we could do is we could probably just apply noise to this, some sort of procedural noise. That would probably look even better. Uh, that's fairly easy to do. And yeah, as you can see, it's even figuring out like the streaks, which is neat. So, it'll just

[1:32:18] do this hundreds, maybe thousands of times until it comes up with something that looks pretty clean. And then once we're done, we'll actually be able to generate those demand uh those ads on demand. And keep in mind, um if anybody from Stripe is watching this, I'm more than happy to build you guys procedural ad generators in 5 minutes. But you could also ask literally anybody that's watched one of my courses, and I'm sure they'd be able to do the same. Not exactly rocket science. And as you can

[1:32:41] see, we're getting much closer to that. We now just have to deal with a couple of little spacing issues. So, I'm actually going to jump into the ad tuner and do that myself. Okay. And it just opened it for me. Um, looks like we can land your first AI client in 90 days. Looks like we can also change the angle of the gradient applied to this, which is kind of neat. Um, I don't actually need to do any of that cuz I'm I'm quite satisfied. While it's working, I'm just going to say add a bit of noise or the ability to add it so I can test. And the

[1:33:09] reason why I'm going to do that is because I think noise just makes it a lot cleaner. So, organic streak textures and stuff like that. We can come up with about as many of these as we want now, which is nice. Streak win. That's kind of cool. You know, it it looks like stripe. It looks like stripe. It's obviously pretty sexy. Okay. So, I'm just going to change the word mark a little bit. Um why don't we make the size bigger? And then what we need to do if think if you think about it is actually change the spacing of some of these because um it didn't it doesn't

[1:33:36] really look like you know as clean I would say. I don't know. I'm just playing around with this a couple of different ways but yeah I think we should also change the line height a little bit maybe. I don't know. Content pad X. I don't exactly know how all of this works but I'm just going to change this around a little bit. Block gap. Oh

[1:33:53] yeah. Yeah. I like that. I like that. Cool. Awesome. And then I think what we just need to do now is just I don't know, maybe find a way to make the font at Maker School just a tiny bit different. Then we can just rip this over and over and over again. Also allows you to use plain text or a pill button. That's kind of cool. CTA text is kind of neat. Maybe I'll say join now instead of start now. And then maybe we'll do something like maker school. Okay, great. So I really like this. I'm actually just going to export

[1:34:20] the settings which is right over here. This will allow me then to copy and paste it into Claude. Uh, and then I'm just going to give it one final piece of advice. I'll say, "Okay, let's find a way to make the Maker School logo more interesting." Right now, it's uh I don't know why I'm

[1:34:40] not using voice transcription here. Just a little too spread out, aka the letter spacing makes it look not as professional. Um, but the fact that we're using a default font also makes it sort of bleed into the rest of the design. So, come up with like 20 different fonts that I could use that

[1:34:58] are all very high-end, new, and clean. And you can see while I was doing that, and this is something that you guys should um do uh pretty often. While I was doing that, it also allowed me to adjust the grain. So, it came up with a new way basically of doing this. So, I can now change the grain. It's a little bit difficult to see here, but if I change some of the colors just to maybe I don't know, let's see here. Maybe instead of white, I want it to be kind of like reddish or something. Maybe pinkish. Um, you guys could see that I'm now adding a little bit of grain. I can

[1:35:27] also change the size of the grain. Maybe I'll make it a little bit smaller like this. Then I'll just make it just pop out a tad. And we can even change what's called the seed here. Maybe we'll do soft light. I don't know. Multiply. Multiply would look a little too strong, I think. So, we'll stick with overlay.

[1:35:42] and then maybe do some color RGB static. Cool. So, I I really like this. I mean, I don't know if you guys could tell, but it just adds a tiny bit of additional finesse to it. Um, and then, you know, I'm just going to wait until it finishes with this, and then we can actually generate a thousand. It's also added um the ability to use radial bursts, which is kind of neat. So, you can even see like I don't know, you could add the gradient right over here or something like that. Um, which probably spices it up even more. Still, I think I'm going to keep it simple with just linear. And

[1:36:08] then I can just like adjust the colors and stuff like that as they will. Uh, nothing super special. Looks like we also now have the ability to change the fonts. So, this is what I wanted the first time. I think they tend to use inter for everything. So, I would I would just want mine to look a little bit different. Basically, we need to do is we just need to wait for this to stand out so it doesn't look like, you know, the rest of this. Um, I don't know. Man rope. Man's such a funny name for a Who comes up with that? Maybe we'll do this. I would say change the

[1:36:36] letter spacing so it's tight. I don't know. Could be distinctive. Sora, that's a little too big. Bricklage grotesque. I actually really like Red Hat display. I think that's probably the best we could get here. So, I'm actually just going to make it a tiny bit smaller. Maybe I'll add a period at the end of it. Okay. Land your first client in 90 days or it's free. So, we'll just change the letter spacing a little bit. Make it nice and tight. And then now I'm just going to zoom out. And h maybe I'll actually just make everything just a tad bigger. You know, one of my friends once

[1:37:05] told me that literally like one of the biggest problems with ads is people just can't see them. Okay, zooming out now. You know, people looking at it on their phones. Maybe we'll do even a tiny bit bigger. Cool. That looks pretty nice to me. I like this. Um, cool. So, what I'm going to do now is just export this as JSON. Copy. And then go back here and

[1:37:23] I'll say, "Great. I really like this. Now that we've created a format that looks pretty good, uh, I want you to apply these settings and then I just want you to generate like 50 different ads for me. uh all with different CTAs, all with different looks and and so on and so forth. What I'll do is I'll just pick the best ones. Then I want you to turn this into an artifact and a reusable resource. So anybody on my team can just quickly jump in and then generate these ads at scale. Uh eventually we'll scale it up and then make it a little bit different where uh

[1:37:53] you know, we have the ability to change which ads we want to generate as well. But yeah, for now this is this is pretty good. First generate me 50 so I can quickly look through and then pick and then do work on generating the artifact afterwards. Cool. So, as you guys can see, we are slowly alleviating bottlenecks and we're making an internal tool to build these sorts of artifacts in just a few minutes. Um, really hasn't taken me that long to do. I'm just going to cut to, you know, when we actually have it. And, uh, yeah, let's take a look. Now, under the hood, what it's doing is it's

[1:38:19] actually creating a script to do it for us. Like, we've, um, used intelligence, okay, this really flexible sort of intelligence to, you know, create a script that is now multiplying and iterating through all the different procedural things we could do. Um, so you know, I hope you guys know like what this has done is it's actually gone through and then already generated all of these things. It's now just going to iterate through all of them. So we can take a look at what that looks like um now because it's gone through and done it. And I'm just going to take a look at it now. Cool. Taking a look here. You

[1:38:47] can see we've now generated uh something like 50 in total. It's even then kind of given me some information on the angles and stuff like that. So land your first AI client 90 days or it's free. Add AI services. Get 21k in tools free. I'm in your inbox every day until you land a client. Your first day a client

[1:39:05] guaranteed in 90 days. I like this one. Sounds pretty cool. So, you know, I could just note like I want to keep four for instance. Daily coaching day one 05. Okay. So, we're starting to get a little bit more impressive here with the various gradients and and looks and stuff. Um, I am finding that we should probably vary like I think we should

[1:39:22] vary the angle of whatever is going on. So, we're not doing that right now. Um, I think it's probably just not actually doing this with uh I don't know its settings. Like it's not actually varying the angle. So, I'm just going to go back. Going to hold command D. This looks great. You should also be varying the angle of that cool background gradient thing so that it doesn't all look vertical. Right now, the vast majority of it looks vertical. Also, ideally, we'd have a quick and easy way for me just to select all the winners and then download them, too. So, hopefully that gives it some more

[1:39:52] context. Um, but yeah, I mean, it did did a pretty good job. Okay, so what have we done now? We've basically built like a procedural ad generator just based off of a scooped template if you think about it. Um, and you know, the idea is not that you just build systems that scoop other people's templates and then just like duplicate the hell out of them. Obviously, the idea is that you get inspired by winning performers um either that your own team has created or that you have run in the past that do quite well and then you kind of scoop your own work, right? And then you just

[1:40:19] generate variants of that really really quickly. You know, I think these variants are pretty neat, but obviously there's only so much we could do with procedural looks, right? We can only do so many things just by varying uh I don't know, changing little procedural variables in the background with color and angle and stuff like that. Even if we fixed the angle and we made that look a little bit better. Obviously, there's only so much we could do. Uh so what I want to do next is I want to take this from a like procedural image generator and then I want to turn it into an AI

[1:40:45] based image generator flow where instead of us just using um you know like this sort of gradient background or whatever we're actually using an AI model to generate imagery and maybe we combine it with our procedural generator to get the best of both worlds and also keep cost pretty low. And then after that, what I want to do is show you guys how you guys could go one step further and you could make video ads with, you know, AI. Have AI orchestrate all of that. And then, um, now, you know, your creative, while still costing money, does not cost, I don't know, let's say $500 per video ad

[1:41:16] that you're testing. Maybe you can get it down to $2 per video ad that you're testing. In doing so, test 10x as many, 100x as many video ads and then have the market actually tell you which ones are the best. Because I find in general that tends to be the approach these days. You don't really just want um you know to like spend a bunch of time and energy on your opinionated idea about what performs the best. You know the cost of production is now so cheap as you guys could see here that it makes sense to build systems to produce a lot and then just have the market actually select

[1:41:43] things for you. Okay. But first I just want to make sure that we do have a repeatable process for this. Okay. And now we have what looks to be um a promptbased way of doing this. And what we can do now is we can also select maybe the ones that we like. Obviously I'm just doing this with like two formats. Imagine if you had 50 and afterwards it'll actually download them all onto your computer. So, for instance, I've now downloaded all of these which hopefully you guys could see are now pretty straightforward and easy to like I don't know upload to uh Google uh Facebook ads or whatever ad library

[1:42:12] manager that you were using. Um but I I just want to turn this into a skill really quick uh before we proceed with the next step and then show you guys how also to loop it. And then we can also have like cloud routines do this for us too. So, you know, now that we've basically accomplished the um prompt done, would you run me through the skill? Well, first thing is if I go maker school ads, you can actually see it um now populating down here. So, no

[1:42:35] longer do I need to manually do this. You know what I can do is now next time I can just write maker school uh ads. But if we read the skill.md, this is the file that it actually went ahead and generated for us based off of my conversation with it. So, you know, I prompted it back and forth. I figured out a prompt that works and then I said, "Hey, I want you to consolidate this into a skill, right?" So, it went through and it did. So, generate stripe style minimalistic ad creative with an organic gradient header and bold headline plus CTA for Maker School, Nick's AI automation and coaching

[1:43:01] program. Use this whenever anyone asks to create ad creative, add variants, add copy with visuals, or run a creative pipeline or ad batch. And you can see that, you know, we have a bunch of layers to this. We even have the scripts involved and so on and so forth. So, it'll go through and generate copy. You

[1:43:16] know, it it's doing quite a bit here. What we've done is we basically taken our whole prompt, which was a series of distributed messages, and we've codified it as a instruction set that we can now feed in and just have it generate over and over and over again. You know, for instance, what I could do is I could actually just open up a new chat, one that has no context, run maker school ads, and now, you know, this model is now going to run based off of this entire long strip. So, you know, this has zero context of our prior conversation, right? And now it's asking, what do you want to do? Generate

[1:43:43] a single ad, run a batch, adjust the visual profile? Um, no. I want to run let's Gen 10 ads and then I don't know maybe instead of maker school ads I want to do this for another product but do it for Clarvo. My startup is clarvo.io. Once done send me contact sheet. So now for instance what I've done is and it's asking me about Clarvo. I'll say

[1:44:07] research it claro.io. Um I'm going to have it research clarvo.io for me. So, it's going to go find a bunch of information about my dialer and stuff like that. And now I can actually spin this and I can use the skill regardless of whatever, you know, product I'm selling. It can go and it can find that information for me, apply it to the skill, and then it can spin it, mutate it in a variety of different ways. So, what it's going to do now is it's actually going to generate a bunch of its own, gone through and just ran all of the different ways you could

[1:44:33] pitch things. And now I even have a contact sheet, uh, which I'm going to open the folder of, and then double click on to show you guys. So now, you know, I've done this, but I've done it for a different software platform entirely. How cool is that? Uh, and the reason why is because I have that reusable instruction set that I can change very slightly. So this one's cool. Dial leads in seconds. Hit 85% pickup rates. See how you paid for the lead. Talk to them. Start now. Call every lead while they're still warm. Get started. I mean, these are all pretty cool. Um, so you know, based off

[1:45:01] whatever background gradient I have, you guys could see how you could extend this. How do I turn this from a skill, which is that instruction instruction set that is reusable, into a loop? Well, I mean, this situation is a little bit different because we've already built in like an ad format, uh, for instance. And, you know, that obviously required some manual work, but hopefully you could see if instead of this set of two ad formats that were duplicating, we had 500 ad formats that were duplicating, we could genuinely create some novel things on a daily basis. maybe once every morning, you know, I would come in here and

[1:45:36] because we're spinning 500 different ad variants times, I don't know, 10 different brands times 10 different formats, 10 different niches or whatever, every time I come in here, you know, I can choose from like one of like 50,000 options. Maybe every morning we could generate, you know, a big batch of them and then have some creative specialist that just goes through, finds the best ones, and then sends them off to the media buy team. Um, that's actually fairly easy to do. So, what I'll do now is I'm going to go back to my main chat, which is general coding session, and I'll say, "This sounds great. I want you to um put this on a

[1:46:07] loop." The idea is we want to run this every morning at 5:59 a.m. So that when our creative specialist gets in at 6:00 a.m., they could take a look through the ads and then identify and pick the winning um combinations. So now what we're going to do is we're going to turn this skill into said loop. It's going to run unattended every

[1:46:27] morning and ask me for some scope size. So batch size, why don't we do 20 copy pool, we will generate new copy every time. And you can see, you know, it's also asking me questions here. I'm just going to answer it manually because I think that's easier. Output location delivery. Um, we're going to generate locally. And then old batches. Um, don't clean up for now. Maybe later. Cool. So now what we're doing is we're converting this from that program that we've run into a selfgenerating or running program that does it all without me. Now in the process of converting from a skill to a loop, you know, some things will

[1:47:04] probably have to change because we're no longer here. If it needs to ask us some questions about, hey, you know, how should I generate the ads? Um, you know, obviously I I I I'm not around to actually like guide it at 5:59 a.m. Well, knowing me, I probably am, but uh we need to give it even more autonomy than it already has. And so what we did is rather than it using a bunch of fixed like pieces of copy that it had created and just ran through and iterated through a script, now it'll actually pass through itself every time when it runs. So it'll actually basically say

[1:47:34] like, "Hey, here are all the ads that we've generated so far. You I want you to make this one different." and it's going to come through and then make something new, something timebased so that it actually runs and varies on a day-to-day basis. So, it's done a little bit of testing and now it's going to do a real batch render to show me how it works before actually instituting it on my computer uh via a platform called cron which you know actually goes through and then makes it happen. Here

[1:47:56] it's setting up the daily automation. First, it's going to check a scheduling skill because cloud actually has a built-in skill called schedule which outlines how to do all of this. And now it's going to actually run this. Okay, now it's asking me some final confirmation info like does this run every day or does it do it, you know, Monday to Friday instead. Um, just gave it that information and now it looks like it's good to go. It's going to run tomorrow morning at 559. Uh, just cuz right now we're already at 8:52. But as we know that is not enough, right? We went from prompts to skills, skills to

[1:48:26] loops, and now we need to go from loops to routines. The major difference between loops and routines being that just because this thing runs at 5:59 a.m., it's still doing so on my computer, meaning I'd need to have like an office computer, you know, uh, at my HQ that somebody signs into every day. That's kind of a pain in the ass. Rather than doing that, what I want to do now is convert this into basically a cloud-based skill that runs autonomously on a schedule and then dumps the information not into a local folder on my computer, but some company accessible folder that anybody in my company can

[1:48:55] use. The question is obviously, how the heck do I do that? You know, everything up until now has been done entirely locally on my computer, and now I'm considering giving it access to something else. Well, like most things in life, you can just talk to Claude about it. Hey, now that I have done the prompt to skill and skill to loop step, what I want to do is convert this from a loop into a Claude routine. My idea here is rather than generate these locally at 559, what I want to do is generate these in a Google Drive folder and then make that Google Drive folder accessible so

[1:49:25] that anybody that has the link can look at it. I'm then just going to give the link to a couple of team members so that when they come in at 6 a.m. they'll all have access to the exact same ad pool. Uh, walk me through how to do this step by step and then set it up for me. Um, I know we'll have to give some access to

[1:49:39] some platforms. That's totally okay. Okay, cool. Now that it has everything that it needs, we're going to convert this again up to uh the highest quality version of this, which is the architecture shift. Obviously, since I'm doing this, it's saying, "Hey, okay, we should probably kill the local one that we're doing and then turn it into something on the internet. Um, does that sound okay with you? How should we do it?" Well, now what I'm going to do is because we're on cloud, I need to figure out the time zone. It's also going to ask about GitHub repos. So, what I'm going to do is I will create a new

[1:50:07] GitHub repo. um that for people that don't know is just like an online place to put code so that code can run without my involvement. What should we do? New dated subfolder and uh you know anything else here it's asking about like some routine information on one time setup and so on. So I need to give it some information in my Google Drive. I'm going to start with a private GitHub repo. I'm actually already authenticated but if I was not authenticated it would ask me to sign in. Uh as you guys could see it's not very big ask. It would just basically say hey do you want to sign

[1:50:36] in? I'd then click yes, sign in, and then I'd be in. Uh, if you don't have an account, it would then prompt you to create an account, which is fairly straightforward, not that big of a deal. The cool thing is Cloud will just walk you through most of this now, assuming that you um have a clear defined step. Uh, this is now finding the Clarvo branded variant of the pipeline that I created before in Maker School ads conversation. So, it's just kind of

[1:50:58] removing that. And now, it's asking me some questions. Yeah, I did this earlier. Not a big deal. can remove. Eventually, it'll ask me about Google Drive. So, what I'm going to do is open cloud.ai/customize connectors in a browser, connect Google Drive, authorize with the right Google account, and then come back here. Um, so in order to do that, I'm just going to go to the top right hand corner, go settings, and underneath connectors,

[1:51:24] which is all the way down at the bottom. If you guys remember, I'm now going to connect to my Google Drive. So, just click that button. It'll now say, "Hey, can we connect?" So, it'll then open up a browser. I'm then going to give it access to my Google. I'll just say use all uh for now. Okay, now we can go back. So, we should be good. It says connected to Google Drive. So, I'm going

[1:51:44] to go back here and I'll say connected. And it should actually be able to set up the list and then create the folder. Cool. And it's gone through and actually added that information. And we now even have a folder called Maker School Ads, which I can click and open right over here. Uh worth noting we don't have information in there yet. The reason why is because it's saying the connector can't change sharing settings. So you might have to do this once and uh that's fine. Uh so I'm just going to go to the top right hand corner, go to share, click share, and then change this to

[1:52:14] anyone with the link can now view the ads. So now I'm going to go back here and then I'll say good to go. So now team members will have access to this folder cuz I'm just going to share it, you know, within my company or whatever. And then it'll actually build the routine that generates the batch. And the last thing we need to do is just give it access to the Google Drive itself. Um, and I say Google Drive itself, not as we haven't already done that because we have done that, but we've done that locally. If you think about it, what it needs to do is it now

[1:52:42] needs to connect Google Drive in the cloud routine instead of here. And it looks like it also connected the Google Drive connector too, which means we now have the routine created. It even kicked off a test run. It's going to check on the status um you know sort of as time goes on. If you go back to routines in the top lefthand corner you'll see that we now actually have maker school daily ad batch. So I'm going to give that button a quick click and it's actually testing it right now which is why the status is still paused. So as you guys

[1:53:11] can see very similar um to the skill. It's just sort of done almost in like this I don't know cloud way. Generate today's maker school ad badge and upload it to Google Drive context. This repo is a stripe style ad creative generator for Maker School. It is a tuned visual profile and a batch script that generates ads with procedurally varied copy each run. Okay, there's a bunch of instructions down here where it runs it through how to do the thing. We then have cloud code remote Google Drive as connectors. Then we even have the GitHub repository which again is just like a it's like a Google Drive but for code

[1:53:41] and that allows it to do things. You can see that we've actually started running it. If I click on that run button at the background, it there's literally like a cloud chat going on with Claude. So, it's doing all of the work, but it's just not doing it on my computer. This is now in the cloud. And anytime you see this little cloud logo, you know that it's not occurring on your computer, which means that it also just doesn't depend on you. So, it's going through this whole process right now. And uh, you know, I'll check back in a minute to confirm that the routine is done. And

[1:54:06] once it's done, we can move on. Gone through and generated the ad successfully. Now, it's just going to verify the output directory contents before uploading it to the Google Drive. So, it's creating a dated subfolder in Google Drive, which I'm going to take a look at right now. Show you guys what's going on. You can see it's now created that. Okay. Now, if I double click on this, it's in the process of uploading the ads. So, we'll see them pop up as um you know, the advertisements get uploaded. Obviously, these are images,

[1:54:32] so they'll take a little bit of time. Okay. And you can see the ads are starting to populate. Now, if I give this one a click, we got some good Maker School ads. Uh same ones that we were looking at before. So, nothing super special or crazy here, but hopefully you guys see how easy it is to build even like a simple procedural ad generator using the the systems that I I talked about. Okay, what I want to do next is I want to show you guys how to go from developing really straightforward um kind of simple image ads to using AI to generate some of the images. Then, after

[1:55:02] we've generated these images with AI, what I want to do is I want to go to video generation. And you know, as mentioned, none of these are going to like knock your socks off. you guys are probably going to be better at PPC than I am if you guys are established marketers that have been working on your own companies. Um, but you know, we can get pretty far with with where we're at and I'm going to run you guys through all of that now. Okay. So, how are we going to do this in practice? Well, first I just want to make it really clear what we've done so far. This is a

[1:55:26] Claude logo in case you guys didn't know. U basically what we've done so far is we have one had Claude, you know, generate I'm just going to handr write here so I'm a little bit faster. um our assets. And for those of you guys that don't know, Claude doesn't actually have any image or or video generation built in. What we're doing is we're actually having Claude generate our assets by using kind of compositing tools. So, for instance, um there are variety of different tools that allow us to, you know, make things on the internet. One of them, for instance, might be, I don't know, Photoshop, right? Another that you

[1:56:04] guys might be familiar with is Canva. I don't know if we have any old heads in the building, but uh I don't know. How about Microsoft Paint? Well, what we had Claude do in that last example is we basically had Claude drive, you know, these tools or equivalents of these tools in order to generate um you know, our assets. And so Claude literally baked in a set of instructions similar to how you know you tweak your lighting and your your settings in Photoshop or whatever. And then it just controlled or steered these until eventually we got, you know, a bunch of outputs, which is what we

[1:56:39] wanted. And so that is how we got those cool stripe images. It literally like locked in and hardcoded. Every single pixel on the page, right? Every single, you know, part of the the headline and the text and stuff like that was just hardcoded. And this is cool obviously, but, you know, we can't really extend far past that. um you know obviously we can do okay things but they're going to be procedurally generated at the end of the uh day and you know they're just not

[1:57:05] always going to be of the best quality. So, what we're going to do instead is we're going to go from this sort of like, you know, version one with really simple ads, which you guys can obviously do and I've done this and it works quite well instead to um kind of version number two where instead of claw generating our assets, okay, what's happening is we're using direct image and video models steered by Claude just like we were steering our Photoshop

[1:57:38] and our canvas and our Microsoft Paints. Turns out we can have Claude steer direct image and video models too and you can do a pretty good job. But there's a big difference between you know uh Photoshop let's say which is a compositing tool and then um you know something like I'm just going to say GPT Image 2. For those of you guys that don't know you know Photoshop composits a bunch of different elements. You kind of control them. GPT image 2 literally produces the pixels on the page directly. Like when you prompt a GPT image 2, you are legitimately turning your text into an array of pixels on the

[1:58:12] screen. Likewise, um you know there's also a bunch of video models out there, right? So I don't know how many people here do video editing, but I mean I do. So Premiere Pro is my chosen tool of choice. Well, that is again a compositing tool for video. It's a video compositing tool. That's pretty different from like one of these next generation video gen models like you know Cance 2. I'm just going to say N here, but really I think the um model as the time of this recording is like a

[1:58:36] 2.5. And so like let me show you guys some examples here. Um these are a bunch of images that I generated in just one of the many tools that uh is available to me. This one's called Higsfield. And what I had was I had to generate a bunch of product photography. So product photography of a frosted glass jar of blue clay overnight face mask minimalist

[1:58:54] label labeling reading midnight clay. Sorry. And so this is actually just like every single pixel on this page has actively been generated by an artificial intelligence. So that little like glint here, you know, the shadows, everything like that. It's not like this was a drag and drop sort of tool. Every single pixel on this page was like independently generated. And it's simply because of the intelligence of the model

[1:59:17] that this is capable of of occurring. just been trained on a variety of stuff. And you can see that idea extends to virtually whatever sort of medium you want. So, you know, it for some reason flagged me when I tried generating images of horses, maybe. [snorts] Anyway, I'm going to talk about that. Um, but uh, you know, you can see, you know, I'm doing this for like painting

[1:59:36] style stuff, for old school astrolabes. I'm doing this for like jellyfish. I'm doing this for cartoons. I'm doing this for some futuristic building with drones. I animated a bunch of videos of a friend and I drinking coffee in Italy uh just to send to him. That's a place that I visited called Rio Major, by the way. That's me making a mocha pot. No idea why it made me look so feminine, but hey, what are you going to do? Um so yeah, all of these are just pixels that are generated directly. And what we're going to do is we're actually going to have AI sort of drive the generation of

[2:00:06] this stuff. And it doesn't just stop at, you know, like image models. You can also generate videos. And so for instance, this video here is generated by a model called Sea Dance 2.5, which is pretty badass. Like here we have a bunch of, you know, ships lying in weight. Obviously, none of this is real, so to speak. It's all a simulation. But the video models are pretty cool because it's actually like uh it's a world that it is generating. It knows the rules of physics and stuff like that. It knows consistent characters, shadows, lighting, how human beings look, and so on. Um, the thing to know about, you

[2:00:37] know, these sorts of direct image models and video models is there's a much higher likelihood that they're not going to work the first time. So, with compositing tools, okay, most of the time once you build a template, [sighs and gasps] uh, compositing tools, maybe you have like an 80% chance of it being okay. And I don't mean okay here as like this is perfect and I'm going to choose it. just okay is like stuff is in the right place. There's no weird elements in the image. Uh the shadows make sense. There's no odd motions. But with direct, you know, gen tools, at least right now, and I'd encourage you

[2:01:18] guys probably to think about this even later in the future, you have closer to maybe a 25% chance. And um the idea here is because of the comparative differences, you know, you should use compositing tools for simple stuff like we did earlier and you should use direct gen tools for more complicated stuff. In addition, compositing tools are virtually free because number one, you get okay-ish outputs every single time or almost every single time. And we're also just controlling like procedural software like programs, right? But with direct gen tools, we are controlling these mammoth intelligences. And so compositing tools tend to be quite cheap as well as accurate um for simple

[2:01:55] things. And then direct gen tools uh can actually be quite expensive. And it's expensive both more expensive per gen, but it's also more expensive because it'll take an average of like four gens before you get a good one, right? So I would just keep all of this stuff in mind as we proceed. This is not going to be free. And realistically, what we're going to do is we're going to generate around four times the number of images that we want because odds are a lot of them are not going to be, you know, make sense and whatever. Okay, so I just say this to you guys because a lot of

[2:02:25] creators will show like the one shot on gens and they'll be like, "Oh, it's perfect. You know, this is an amazing video model. Changes everything or whatever, but it doesn't really do that." And uh it's good for us just to, you know, dial our expectations back a little bit and be more realistic. Despite that, you can obviously do incredible things. And uh I'm going to show you guys a couple of really cool examples uh right now. Okay, so the very first thing we need to do if you think about it, um, is we need to go and grab our creative pipeline. And so, same thing that we did last time, very, very

[2:02:52] similar. We're going to save our best performing ad formats as templates. We're then going to feed in a variable, which in my case, I'm just going to be one of those product images that I generated earlier. Then Claude is going to generate a bunch of combinations of all the different types of ways the product can look within the confines of

[2:03:08] the template. Then we're going to pass it off to nanobanana or GPT image 2. I think I'll probably do GPT image 2 to start just because um it's easier for me and I have the API key. And then uh finally we'll review pick and then you know ship just like we did before. And the human is still in the loop. We're still the people that are ultimately making decisions whether we're creative director or whatever. Like we're going to be the people that are in the driver's seat. Okay. So how how do we do that? Well, I'm going to go to the ad library here. You can see I was kind of

[2:03:34] peeping um another image, but uh what I want really is I don't know. I just want let's just say socks. I don't know. I just want to like see static image ads of socks. Why? Cuz I'm probably going to try advertising some socks or it's like some consumer product or something which is fairly straightforward. So, just zooming in here, what do we

[2:03:56] got? We have uh it's an image of a belt. Okay, that is literally an image of socks. So you can see like Amazon's doing a lot of these and Amazon has quite the style. Uh they typically don't like provide any text or anything like that. So I just want some text cuz I think it'll probably be a little bit more representative. That is that is I don't know that just does not seem like a very good ad. This also doesn't seem like a very good ad. So I don't really

[2:04:18] want to use that as like my my template. Let me find something nice. All right. This looks pretty good. Nordic socks. Right. So the logo of the company in the middle and then we just have like a four um sort of panels each that have their own socks. So, I'm just going to go over

[2:04:31] here and then see add details for one. And then it looks like I can just open image a new tab, I would hope, which now gives me the ad. Cool. So, I'm just going to save this. We'll just call it socks one. Okay. And uh you know, this time I just want to do a little bit more than that. So, this is cool. It's Marvel and then some other brand with kids wearing a bunch of socks presumably. And you know, if this is like some sock company, then I think as long as I have some notorious business that I could like claim affiliation with or something,

[2:05:02] then that'd be okay. Maybe we have a new line that's coming out or something. I don't know, man. I definitely do not sell socks if it's not clear. [sighs and gasps] This is interesting. Uh, it's like some dude putting the socks on, moving around. Let me just see if we could find another really good

[2:05:16] one. I like this one. Sock well feel better in style with marino wool. Very similar to what I am aiming for here. The only issue is the sock images are pretty difficult to see. Still, I think that's going to be okay. It's just we're going to have to like, you know, if I open

[2:05:34] this in a new tab. Yeah, like they're pretty small still. I'm going to save it. I'll call it socks 3. Cool. And now that I have a bunch of socks, and you know, these are kind of my templates. What I'm going to want to do is I'm going to open up Finder and then I'm just going to grab all these. And then what I want to do is I just want to pump them into the same sort of folder that we had before. And I have one called ad formats for image generator. So I'm just going to go up, select these three. And then I don't

[2:06:00] know why the images are so big here. Let's just zoom out. It's a picture Jack sent me for our podcast. You can see we are we are really ramping up the podcast here. the stacked podcast, anybody that was wondering, is uh is my life's work and I'd highly recommend you guys check it out. Okay. Um and now we have actually everything in the folder, which is nice. Cool. Right. All right. So, now that we've done this, okay, the next logical step is we obviously need to have Claude sort of steer and then Okay, so now that we've done this, the next logical step is we need to feed in

[2:06:32] the variable. And in our case, the variable is just going to be the product. Um, I do recommend that you guys have products. You know, if you guys want to advertise something, you should have something that is like very visually distinct and it's a product. So, I got a couple of products here. One's called Cloud Socks. They're marino wool sleep socks. Another one here is called Linen Tails, which is a waffle

[2:06:51] towel bath sheet. 100% organic linen. You know, it's kind of cool, right? I like the look of that. And uh what I want to do is I basically want to take my product and then I want to use that proven format that is obviously known to work really well. Assuming you're taking it from your own company or something like that, right? Um and then just generate a bunch of variants. So the first thing I'm going to do is I'm actually just going to go to an image generator directly and I'm just going to verify that it can do the sorts of things that I want it to do. And if it

[2:07:15] can do the sorts of things that I want it to do, great. Um that's fine. We can move on. If it can't do the sort of things that I want it to do, I'm happy that I'm testing it at the beginning because then I'll know, okay, I probably shouldn't pursue this because even if Claude prompted, I wouldn't be able to do a good job because the model itself just is insufficient for my task. So that's typically how I treat these things. And um as mentioned, the model I want to do is called GPT image 2. Uh it's a pretty cool model. You know, it has a lot of I say highquality detail

[2:07:40] orientedness to it. So you can actually like write things. Um, the reason why every single one of these uh has text in it is because for a while images could not be generated with text with most image models. Um, now, you know, this was like the first time that we had coherent image gen uh with like a bunch

[2:07:57] of different topics and stuff like that. So, anyway, now I'm right over here and what I want to do is I want to create an image and we'll see if I could do this with the free model. If not, I'll actually have to go and uh pay a little bit of money for it. But what I want to do is I want to take my little finder here. And I don't know, I like sockwell as like a demo. So I'm going to feed this in. Okay. And then I'm also going to grab this these cloud socks. Actually, maybe we should do the lens. No. No. We'll do

[2:08:22] the socks. Call these sock real. Okay. They're totally real. I'm g feed this in. Looks like I can't just because I think the file type is a webp technically. So I'm actually have to do like webp. Let's do uh WEBP to PNG converter. Kind of annoying to have to do this step, but sometimes, you know, as you are scraping and taking ads from various parts of the internet, you'll find that the um formats of the ads are not what we want. So, [gasps] let me just get a real PNG here. Okay, cool. We do have it. I'm just going to go back here and drag this up. Okay. Anyway, it

[2:08:57] turns out that I cannot upload more than one image in a free plan. So, I did have to go to my business plan, which is kind of unfortunate. uh just kind of how it is. So, no major deal here. Uh let's put that away. Um yeah, so what we need to do is, you know, go and get a paid plan. For those of you guys that didn't know, ChatGpt has a variety of plans. My recommendation for um you know, pricing model here is I would probably get at least as of the time of this recording like the Go plan or the plus plan. um you can't upload

[2:09:29] more than one image with the free plan and it'll also interfere with your ability to communicate with the via API. Uh any who so I went through and I got the you know chatb plus plan and what I'm going to do here is I'm going to upload socks real okay which is that real image of the socks and then I also wanted to upload socks 3 which is um I think not findable right now because I hid it away. Cool. And I'm just going to say, I want you to use the add format in the second image, but apply the product in the first image to it. I then want

[2:10:00] you to add, you know, value props that make sense for what the product is, which is a light, fluffy cloud sock that's cozy and helps you go to bed. You don't need to do the exact same um background or setting. I just want you to apply the template in so far that you know the main focus of the uh static ad should be on enumerating the benefits of the product in sort of that cute quirky way. Uh also make sure the name of the company is cloud socks. Then just going to put it in. It's going to analyze both of these images and then it's actually

[2:10:36] going to try and generate for me. Okay. And so here we have our cloud socks ad. Softer, cozy, cloud-like, your new bedtime essential. We have the ad itself feels like a cloud. Lightweight warmth, cozy bedtime comfort. Helps you wind down. The actual image of the product all rolled up. Looks like some of the text is a little cut off there. I would say this is actually like reasonably good. It's not the best ad ever, obviously. Okay, we've now generated another one. And yeah, I mean, I'm I'm pretty confident this is okay enough for

[2:11:06] me to uh want to move forward with it. You know, I think obviously the bottom is kind of weird and not all of these are perfect. Um, like for instance, what's going on over here? How is she grabbing this thing? And what exactly is she wearing if her if her thumb is sort of popping out underneath? So, there a couple of idiosyncrasies, but it's not terrible. And, uh, yeah, this is now confirmed to me that I think we can move forward. So, all this basically tells me is that like this is feasible. You know, we got an okay looking at. It's not perfect, but it's it's okay. And I'm

[2:11:30] sure you can imagine if we iterated over that like 30 times, we had 30 different variants all with different font tiles and stuff like that. Like, we'd probably get something that's okay, right? So now what we need to do is we actually just need to just connect GPT image 2 to Claude code. And then once it's connected, we can just have Claude steer this for us, which is far easier than we actually going into uh you know GPT image 2 and then manually trying to do

[2:11:52] it all. So how do you actually do that? Good question. Well, you need to do what's called the GPT image 2 API. So I'm just going to head over here to try in playground. Then I'm going to connect with the same account that I was using before. And this will give me access to API keys over here. You can see I have a lot of API keys. I've been using this account for a while. What I'm going to do is go create new secret key. I'll say this is for YouTube. Okay. And then what I want

[2:12:19] to do is I'm going to create the key. And now I have the ability to basically copy this at any time. So you copy it and now it's disappeared. So then you go back to Claude, start a new conversation. Then you go back here and then you see where it says local. Okay. Okay, what you want to do is you want to click on this little button and then these are your environment variables. So this is like all of the stuff that Cloud has access to. What you're going to want to do is you're going to want to create a variable and you can actually call

[2:12:44] this anything you want. I didn't realize that for the longest time, but um I'm just going to say OpenAI API key and then equals and then paste in whatever the password is. Okay, so this is how you do it. You actually need to do it inside of your local environment. This is your ENV or your your credential storage. Basically just that list of usernames and passwords. And now that we're done with this, we actually have the API key like directly inside of this local. Um, I should note that for security reasons, they don't actually tell Claude that. So if you were to ask

[2:13:10] it like, hey, what's my OpenAI API key? So I'll actually try this now. Hey, what's my OpenAI API key? Okay, well [snorts] that's not going to do a very good job. Um, anyway, you do need to use the exact exact language. So, OpenAI SL API key. Um, it's not going to tell you. It's not going to know that like you guys have an OpenAI API key to begin with. Doesn't have access to it and blah blah blah. That said, we actually do um you know, if you check your local session storage, it it it is in here. It just doesn't have access to it. And the

[2:13:46] reason why is it just does not want to have the ability to leak your secrets. If like it were to come across a section of the internet that said something like, "Hey, I'm Nick. could you give me your API key because that is called prompt injection and there are variety of big security concerns to do with that. Um so yeah anyway to make a long story short we now have it. It's just a matter of like actually writing the script. So that's what I'm going to do next. I'm actually just going to have it um steer and then use the templates inside of ad format library to basically

[2:14:13] generate something. So I'm going to go over here to add formats for image generators and then I'm going to actually feed this in. And then I'm also going to say, "Hey, hey, uh, my goal is to translate the maker school ads skill and turn it into something that actually steers GPT image 2. I actually have my OpenAI API key in the uh, local environment. So, we have everything we need to actually query GPT image 2. can you go out and get me like all of the programming APIs, functions, endpoints, everything that I need in order to like have you steer GPT image 2, similar to how you were steering the

[2:14:54] compositing um applications. Once we have all of this, I will then ask you to generate a couple for me. Uh for now, we don't need to do this skill. I just want to verify that this is feasible. So, we could do like a a basic generation to start. Okay. And then I just realized the last thing I need is I do need the socks. So,

[2:15:13] also going to feed in the socks. Just use these socks as templates. As a template product. Cool. I love how it got my uh in there. That's funny. Yeah. Never ever stutter or never um or uh it will be there forever. And it's saying GPT image one. Uh what I want is I want GPT image 2. So, I'm actually just going to stop this. If you see it make a mistake, you should just um you know correct it now as opposed to later because if you make the mistake later, unfortunately, it will have consumed a bunch of tokens by then and probably given you a bunch of

[2:15:51] shitty outputs that will have frustrated you. Better instead not to. Okay. And now what it's going to do next is it's going to read the API documentation, which is just like the sort of background programming um SOP on like how to actually call the API, how to actually generate an image without the front-end um you know, user interface. Okay. And now you will eventually get some sort of notification like this asking, hey, you know, can you give me access to the directory. That's fine. Now it's listing a bunch of files in there. Just let this thing steer all the way until the end. You can see it's

[2:16:24] actually now running a live test image edit, which was part of what I asked it to do. Um, it looks like it's going to be using the socks photo as an input image. Okay, so this thing went nuts. It generated five different variants as you guys could see here. Maker School cloud socks. Uh, that's sort of based off of

[2:16:40] that, I guess, like Marvel Bombas one. Same thing over here with the logo up top. I I like this one personally. Um, I mean, you know, I just gave it maker school. Obviously, it's not actually related or whatever, but yeah, I like how it put it in that box and it it does look quite clean. Same thing over here with the product. And then same thing over here with the product. I mean, the products are a little bit different from image to image and that's just because I gave it a rolled up version of the product. If you guys wanted to have a lot more, you know, product adherence,

[2:17:07] then you would actually just feed it like a unrolled photo, maybe something like this, and then it would actually be able to get the curves and the folds and stuff. But yeah, hopefully you guys see like I, you know, we basically have an image generator. However, um, the probability that this will be exactly what I'm looking for has gone way down simply because it now has a lot more uh, freedom and flexibility. And that, you know, takes me to an important point, which is that the more freedom and flexibility that you give a model, typically the more rope you're giving it to hang itself. And that's good in many

[2:17:35] ways because it's the freedom and the flexibility of these models that ultimately allow us to do cool things with them. Stuff that you could not have imagined doing just a few years ago with a Photoshop. Um, but it's also the the main downfall at least for business because with business you tend to want to have like very repeatable consistent processes as talked about that you know convert maybe a stranger into somebody that actually has paid you money for a product or service. Hard to do that when you have total crazy pixel um variants like this. you know, a photo that's on like a linen cloth to one where both of

[2:18:03] these are like hanging in the air. If you think about it, like these are very, very different pictures. The pixels are all over the place, very very fundamentally opposed. So, we, you know, what we want to do is we actually want to exert a little bit more top- down focus. Um, we want it to be a little bit clearer and closer to, you know, the the sample image. Uh, we don't want to copy the sample image, obviously, but we do, you know, we want to we want to extract the elements in the sample image that make sense to vary and then the ones that don't. And so what I'm going to do

[2:18:29] is I'm actually going to talk to it and ask it to do that. Hey, I thought this was a pretty solid job. However, I'm noticing wide variance over the input image. For instance, it looks to me like the crux of what you regenerated was the logo uh rather the two logos being opposed, Maker School and Cloud Socks. And that's fine, but it varies tremendously from that image ad that I gave you as an example. I think a better way of doing this would be you digest the image ad, then identify a bunch of different areas that we could tweak things, sort of like levers to pull, similar to the compositing

[2:19:07] workflow. Then you template out a bunch of changes and then we generate them that way. However, we do need more conformity with the main image. The idea is we just want a template that we can hot swap any product into and then be approximately 80% sure that what comes out is going to be okay. Okay, now I'm going to feed that in. We're going to go back and forth with the model until we get something reasonable. And to be clear, this is the prompting stage, right? This is the whole idea. We are um we're not automating this yet because we don't have the process in place. If we

[2:19:37] were to try and automate this like build a skill essentially, we would not be doing a very good job. Um now, what is this doing? It's actually coming out with the various things that it can change. So, the first thing is somewhere between 0 to 20% of the height, you know, we want some sort of a logo band. Okay. Then we need transition or

[2:19:55] negative space. Okay, that's important. Then we need uh the product band. Okay, and then finally at the bottom we need some other thing. So what this this is doing now is it's actually logically breaking down the template into its components and then it's you know actually going through and generating the stuff. Now it's asking me what would you do? Should I do a model plus some sort of overlay? Um generally speaking that's a bad idea because you are then adding you know a procedural compositing approach to an AI based generating pixel by pixel approach. So, I'm actually just going to use a pure prompt template basically every time. And you can see it

[2:20:26] even says I lean product option number two since text and image is the single biggest variant source and you already have a proven compositing pipeline for exactly that layer. Uh, I would agree with that. And so, yeah, it's just going to continue creating this stuff and uh I'll take a look at its results afterwards and then circle back. Okay, looks like it has finished one and now it is just waiting on the second. Both are done. So, I'm going to get to look at both of these images now. Let's take

[2:20:50] a look here at these pictures, shall we? Okay, you can see we're starting to get a lot closer in conformity. Uh, mainly we're now kind of comparing uh I think the photo was like kids wearing Spider-Man socks or something like that. Now we have something that's a lot closer to what it is that we wanted initially. Now we have people basically with their feet up wearing these socks showing off how cozy they are, which is kind of cool. Um, so this is more more the idea. I didn't want just a bunch of people doing whatever the hell they were doing back over here, which looks to me

[2:21:19] drinking coffee sort of solo because the only real element that was taken was the the logo in opposition. What we have now is we have something a lot closer to that initial, which is, you know, the feet on a couch or something like that comparing the two. Presumably, you know, in advertising, this gets closer to the core of what the product is about, which is, you know, getting cozy with somebody

[2:21:38] that you love or something, right? That's the whole idea. Just doing a little bit more tweaking here. Uh, we got a reasonable looking variance. Uh, this is more close to I think what I was looking for, which is I just want to take the product cloud socks and then apply them to a template. Uh, it's not one for one, but it's pretty good. And what I like is we basically have like a rainbow of different skin colors. Um, we have really cute looking socks with our logos and everything like that. Um, this this looks good. It's like it really took the horizontal idea, which is valuable. So, I'm still iterating on the

[2:22:09] generator, of course. Um, but you know, the generator is getting pretty close to where I want it to be. This is going to take some back and forth no matter what it is that you are doing. Hopefully you guys see now just how much more variable AI can be versus, let's say, a compositing sort of flow. Cool. And now we are building out the rest of the generator, which is just ensuring that this can work across the board. Um, this one looks like it kind of put this in tennis. I don't know if that's necessarily the way to do things. We'll see. Allight comfort and you're out

[2:22:32] playing tennis. This one over here though is basically exactly what we wanted. We wanted some sort of like collage based flow where we have like a bunch of different things. So we can see the texture of the sock. We can see a couple of the socks laid out. And then yeah, we have that other one. So what I'm going to do now is I'm just going to turn this from um this like built-in flow here which is occurring locally to a flow very similar to our HTML one where you could actually like click on the ones that you want. So that's what I did right here. just gave it a prompt

[2:22:58] with an abundance of ums and o's. Basically saying, now that we have the flow working inside of cloud code, turn it into that HTML flow we had with the little check boxes and the ability to mark the ones that we want. Also, instead of just generating three, generate a lot more variants. So, uh, kind of misunderstood me there, but every variant should or each template

[2:23:16] should generate at least five variants. So, that means if we're going to do three, we should have 15. And then we should actually like feed very different instructions into each so that you know a taste person, some sort of creative director can just select the winners. So that's what we're doing. Um we're adding a bunch of different like variation pools per template so that we can adjust uh the prompt that we're sending. Maybe some of them will be warm, some of them will be cool, some of them will be

[2:23:39] inside, some of them will be outside. But again, you know, we're just taking that template and adhering to it quite a bit. So you can imagine how for this template we'd probably have four distinct sections. This one here would be some sort of feet arrayed horizontally most likely. This one here maybe would show it in action in some way. Uh and then you know we just add as many of these ad variant inspir

[2:23:58] inspiration ads as humanly possible. We're done generating 12 out of 15 of the first batch. Okay, now we have our own little generator very similar to what we had before. It's like uh like another mini app. And the idea behind these mini apps of course is you know I want to show you guys how to build cloud codebased flows but the cool thing about cloud code is you can use them to build apps. And uh this could be a really cool internal tool for team members obviously based off of you offering the correct

[2:24:20] formatted um you know ad initially. Yeah, you can see how we have maker school cloud socks. We have these sort of brands uh opposed to each other. Now all of them are slightly different in their own way. We have a background colors. We have people's different skin colors. We have uh the socks themselves that look a little bit different from one to one. We have our little like ad collage here. If this was a really high performing template and we were just like ripping that template, you can imagine how we can just generate a thousand of these really easily and really straightforwardly. Uh none of these are going to win any awards. I

[2:24:51] think in the way that I've shown them here or done them, but suffice to say, you guys understand that I don't have like really high performing ad templates selling socks. You guys might for whatever your own company is. So this is how you would rip that format and then really expand the top of funnel uh to be able to select, you know, the the the the best variants for you. So, you know, this app allows you to very quickly and easily scaffold out, let's say, 20 and then you can just download selected. Um, maybe you you fill your template folder with like 500 ads or something and then

[2:25:19] just consistently rip the best ones until you get something that's nice. And you can change adherence, too. I mean, like in our case, we have it pretty strongly adhering to the same look, but we could also decrease that if we wanted to add diversity. So, I don't know, maybe instead of having these on like a uh the top right hand corner every time, maybe we just change it so that we can actually experiment with maybe triple collages instead of quadruples and so on and so forth. All right, so where do we go from here? Well, now we need to turn it into a skill. This looks great. Um,

[2:25:47] let's turn this into a skill. Looks like I was voice transcribing without realizing it. This into a skill now. Um, the whole idea being that we're going to go from prompt to skill, skill to loop, and then loop to routine. Skill's just about done. What I'm going to do while this is um wrapping up is I'm also going to turn it into a loop. So, I'm going to press and hold this to record. Hey, now that you've created the skill, I'd like you to turn this into a loop. I want the

[2:26:08] loop to fire at 5:59 a.m. every morning. Very similar flow to our Maker School ads flow. The idea is being uh we will have a link by you here that anybody can use to access a page that includes let's say five renders per variant in the ad format. Sorry, I don't know where I got 10 from. If we had 10 images, it' be time five or 50 in total. Generate three * 5 or 15 in total. Cool. And uh now that it's, you know, running this on a single image smoke test, I'm just going to feed this in. Uh what happens when you feed in something while it's working

[2:26:43] on something else is it finishes this thing before it starts this thing. So uh you know, now that this is sort of in the background here, it can start on the other task. Basically, just fire that off and it's now going to turn it into a loop. So that's kind of neat. Um obviously you could do the exact same thing now again by telling it to turn it into a routine once we've verified that the loop works. Now for the routine we

[2:27:03] have to go just one step further here. Um the reason why is because we are now going to be running a cloud environment that includes an API key. We don't uh currently have the ability to provide to cloud. So uh if you go back to your API keys here I'm just going to create a new one just out of simplicity and so that you know I forget I don't forget. This is going to be routine API key inside of GPT image. Just going to copy this. What I'm going to do is I'm going to go back to cloud, open up that routines tab, and then um as you see, there's a local

[2:27:34] routine, which is kind of a skill. There's also a cloud routine, which is kind of the one that we want. [gasps] And what we need in order to make this work is we need to set the environment such that it includes our OpenAI API key just like this. Okay, so that's in ENV format. This is our default environment. Uh this default environment will apply to all new sessions. So I'm just going to save changes. Now we actually have all that. And then if you think about it, what we need to do in each session is we just need to uh basically do the same thing that we did before just in

[2:28:06] the maker school ads. It's just we need to do this in um the GPT image non-compositing workflow. So you know it has the key. It has everything that it needs in the default environment. If I go back to routines then try creating a new one on cloud and then check the default environment you'll notice the key is still in there. So, it's now like inside of our thing, and now we can actually go back and like ask it to create the create the actual flow. So, I'm just going to jump back in here and then say uh well, uh what kind of ads did we have? We had I think product ad

[2:28:34] genen. Yeah, we did right over here. So, turn product ad genen into a routine. um for the routine mirror the maker school ads system that generates a Google drive folder and then a dated subfolder and then populates it with the gens do so using this structure for every ad in the template ad folder which is currently

[2:29:01] hosted at generate five variants. First take the ad formats for image generator and put it somewhere on my Google drive. Then uh build the structure that I talk about mirroring the maker school ad system. Um the maker school ad system is a routine. So query that routine first, get all the information and use it to scaffold out the new gen. I don't know why I can't type to save my life today, but hopefully this is going to be more than enough. Um oh and you know I should probably give it some context. I have also added the API key in the same

[2:29:33] format as our original skill. It's in the default environment under OpenAI API key default cloud environment. There you go. Um I'm just giving this it just in case it knows sort of how we're going to structure things. uh we need to be 100% clear because as mentioned it will not have access to the API key inside of the cloud environment simply by simply because rather of um you know security doesn't want to like show you the API keys everywhere because every single push it's not going to make any sense I think it's probably misunderstanding me too check the routine first there is a routine called you know maker school I

[2:30:14] don't know let me actually double check this what this thing called maker school daily ad batch daily ad batch This is what I want you to replicate just for our product aden skill. Yeah. And then we're just going to turn it from, you know, a local flow into a routine. Cool. And now it has the full spec. And it's just going to continue going down the top to bottom. And just like we did last time, we're going to create a new GitHub. And then did I

[2:30:39] already put a file on my Google Drive? Uh, no I do not. No, I don't. So you will create one. And voila. If I move my fat head and super cool Clairvo hat out of the way, you guys can see we now have both the product aden daily batch and we also have the Maker School daily batch. So, next run tomorrow at 5:59 a.m. If I jump in here, you guys will see we have more or less the exact same layout. You know, install the dependencies, run the script. All the stuff does not really require us. We don't have to be pros at programming to know what's going on. Um,

[2:31:08] has some scheduled runs. You can also trigger it manually. API web hook. I'll cover more about that later when we talk about weaving it into a pre-existing automation flow. But yeah, it has the Google Drive. It also has the clock code remote. And then it um has our API key again stored inside of the environment itself. So that's important. It's tough to get stuff in the environment unless you know how all of that works. Um I don't think they make the environment feature as straightforward as it could be. But anyway, that's that for both um you know, compositing workflows, which hopefully you see are pretty

[2:31:36] straightforward if you use a simple ad. Um image based workflows, which hopefully you guys can also see is pretty straightforward if you guys use a good image model. Why don't we talk a little bit now about video models? Now, video models need to be done quite differently. If I go back over here and then open up something like uh Higsfield, which is where the vast majority of all video models are currently hosted and uh you know people sort of use them, you guys could see that there is just tremendous tremendous um possibility for really high quality like adgen. The one issue I would say with most of this stuff is there is

[2:32:07] currently no like there's no captions on here. There's also no text and the reason why is because text just tends to perform kind of poorly. Um, also this model is like kind of an advanced model. It's not technically here yet and so on and so forth. Um, it's starting tomorrow as of the time of this recording, so I'm sure by the time you guys use it, it'll be really good. They're offering free C dance 2.0 4K, which is kind of neat. Um, core things to know before we get into this is, you know, like video models, just like image models, there's a lot more frames, and because there's a lot

[2:32:34] more frames, there's a lot more opportunity for it to screw up. And so, you just need to understand that what you are going to generate is not going to be perfect. And so long as you're okay with that, that's fine. Um, generally speaking, the flows that tend to work really well are like kind of UGC style, uh, user generated content style, uh, either testimonials or product reviews or that sort of thing. What I'm going to do is I'm going to show you guys how to use this particular vendor, Higsfield, which is quite a high quality vendor. It allows you to basically streamline the process of putting in

[2:33:00] like e-commerce style ads. You can uh add a product as well as a consistent character um to generate like UGC style advertisements. And I want you guys to know that like you can make this better. You could make this better by hard coding captions. You could make this better by, I don't know, like hard coding some sort of visuals and and having them pop up on the page. You could automate the process of like the dubbing and stuff if you wanted to be better. You could add noise layers to the background to make it more realistic. I have a few friends that spend several tens of thousands of

[2:33:26] dollars a day right now running like totally AI generated ads using the flows that I'm going to show you. They just touch them up a little bit more. And for time purposes, I'm not going to get into all of that, but it is very straightforward. You literally just talk cloud code through what it is that you want and then yeah, that's that's more or less it. So, let me show you guys more or less how how this is going to work. Um, yeah, for those of you guys that don't know, Higsfield is just like a major model aggregator. Basically, it just combines a variety of other models

[2:33:50] that you guys could use. So, all the video models, all the image models, everything that we've already touched at uh touched on, as well as even like text models and stuff into one place. And it kind of makes sense in an environment where there's just more models than most human beings know what to do with that a point of value would actually be in the organization and um making available of those models to the general public. Like think about it, there's like a bajillion

[2:34:13] different models you could choose. There's a bajillion different platforms you could sign up to. Um you know, in an environment where it's just really confusing for people to get started, what they do is they basically just combine all that onto their platform and then you just need one API key and then it does it all. So this is what I'm going to be using. Uh you guys don't

[2:34:27] have to use it, but I really like them. It's a cool, you know, set of of people. Um it's also just uh yeah, it's a pretty straightforward platform. And I think Claude Code is really really good at driving this puppy um because they have like a really simple and easy sort of API endpoint structure called a model context protocol server. So if you guys don't end up using this, then just know that you'll probably have to do a little bit more work there if you want it to work like very similarly to what I'm showing you guys. But that's okay, right? Um, and you know, hopefully we've

[2:34:55] established now that video costs more money than photo. And, uh, I haven't made you guys feel like this is super unrealistic or anything like that. [gasps] Okay, cool. So, what do you do? Uh, well, first things first, you got to like make an account. So, I'm just going to open up an incognito window here and then go to higsfield.ai. What you have to do is you have to click sign up. They give you guys like a bunch of discounts and stuff right now. So, I'm just going to let's exit out of that. Sign up. I'll go continue with Google and I'm just going to add my email address. So, I've

[2:35:19] signed in. It's going to say verify that you are a human. So, I just did that. Keep in mind the interface is going to look very very different um by the time that you guys are using it probably. But anyway, we want to do UGC videos and we're maybe beginners so we want oneclick presets and I don't know uh what do we want to try? Probably MCP and then I'll just leave it at that because yeah, again this is just like a just an onboarding and there's many different

[2:35:40] ways you could do this. Okay, cool, cool, cool, cool. So yeah, we've now received some big fat offer and they'll give us a bunch of money off. That's cool. Um I'm just going to you know proceed without that account understanding that you do have to purchase some credits but um yeah anyway so what we want is basically we want two things. The first thing we want is uh we want to generate this locally just so you guys understand what is going on under the hood all the different parameters and stuff like that that you could use to tweak in order to generate your like high quality UGC. And then

[2:36:11] after that we're going to do this via program uh you know claude code via what's called MCP model context protocol. And this is where Claude can actually just like drive this whole thing. So we can actually just copy over a prompt here and basically say, "Hey, install this whole thing." And it'll work perfectly fine with no issues at all. So yeah, pretty cool. Uh I'm going to head over to video first and then I'm

[2:36:30] going to generate something pretty cool.

[2:36:39] You can see I've already done a couple of different gens here before. Um you know, one of the gens that I did was uh uh ash falling in a city, which is kind of cool. show you guys what that looks like. You can see it's slowly sort of going down to the ground. That was for a website that I made. This is sort of like a cool

[2:36:56] wheel that's zooming in and out. That's kind of neat, right? It's like a like an old church style theme. I've done a couple of other pretty badass ones, too. Here's a bunch of like people charging in like a more medieval battleground. You'll notice that some things are kind of weird, right? Like why are the horses going backwards? Well, it's because this model isn't perfect. And uh you know every now and then one of the gens will be sort of trash. So you just got to keep that in mind. Uh anyway, so what do we want to do? We basically want to create a video using some media over here, some user

[2:37:25] generated content stuff um of somebody talking about a product. And I think what a lot of people don't understand is that you can't just do this in one shot and expect it to work really well. What you need to do is you actually need to like vary it and have multiple different shots. And so what I'm going to do is I'm going to feed it an image of a product and then I'm also going to feed an image of like an influencer and then I'm going to get the influencer to like talk about the product. But if you think about that that's like that's actually multiple steps. Um so what I'm going to

[2:37:50] do first is I'm going to go actually not over to video but to image and this allows you to generate things using um image models. So as you can see that midnight clay thing that I generated is right over here. So, I'm actually going to reference this in an image flow. And then I'm actually going to go and I'm going to find some like UGC style influencer on Pinterest. So, I don't know. I'm just going to go like, as you can see, there's a lot of like UGC style influencers here talking about products and stuff. Um, I don't know. Let's use this one. You know, obviously most of

[2:38:21] the time, and I I don't want to stress this too much, but most of the time it's like a woman that is demoing a product of some kind. And that's just because mathematically they found that when they have women demo products, they do a lot better. So a lot of these UGC's are going to be um you know images and videos and stuff like that of women. But anyway, I'm just going to feed in this woman. And then what I want is I actually want to go to uh what do we have? GBT image 2. I don't know. I think we had GBT image somewhere here. So let

[2:38:47] me find it. There you go. 4K images. And uh we have a bunch of settings here. We can go high, best visual fidelity, 2K, 1K, 4K, auto size, which is important. So this is going to depend on how we how we're going to generate our videos. For the purposes of simplicity, I'm just going to go one one and assume that we're generating everything one. And then batch size. Batch size is just the total number of gens that I am going to create here with these two images. And so what I want to do first is I want to generate a woman that's kind of like this woman. Okay. Actually, so let me

[2:39:13] just go back. Generate a woman that contains all of the features that this woman has, but is ultimately different. Um I don't know, change her hair a little bit. Uh and then instead of outdoors, put her indoors. Uh the reason for that is I'm going to use her in like a product shoot later. So I want to generate I don't know three options which is going to cost me 21 credits. And the workflow is you know we're just going to take this woman who is I don't know if she's real. Maybe she's already an AI generated image or something. She might be. But we're going to take her and then we're going to

[2:39:43] stick her in a fake environment that I've made and it's going to be indoors and we're going to use this woman who is now fake. It's not like a real woman to advertise our products by turning that into a video. So that's sort of the three or fourstep flow. And uh yeah, I always like doing this manually at least once because it's important for you guys to understand how it works manually before you guys move to fully automating it with cloud code. If you guys don't know how to do the process manually, I find a lot of the time you're just misunderstanding the the different features and the different options that

[2:40:08] are available. Uh I generated this in 2K. You could just generate this in 1K, which is just 1024 pixels. Uh, but I find that, you know, if we want really high quality input images to be used for the purposes of video, it's better to, uh, you know, do this with like a high 2K. Um, yeah, and then, you know, there's a bunch of different models you could choose here, too. There's Higsfield Soul, Higsfield Soul Cinema, Cream 5.0 Pro. This one's actually pretty good. Um, they're also offering, at least as the time of this recording, a bunch of free usage. So, that's pretty neat. Uh, it's just like an advertising

[2:40:38] play where they're trying to, you know, make as much money as humanly possible by incentivizing models that are last gen, pretty cheap inference-wise to run. And it's kind of like a Costco chicken. You come in through the door, buy yourself a Costco chicken, and then you're like, well, I'm at Costco, I might as well buy myself some tomatoes. Yeah, that's their whole idea. This is not occurring instantly because it is quite computationally intensive. We are using GBT image 2 and generating, you

[2:40:58] know, 20 48 pixel x 20848 pixel images. So, I'm just going to circle back when this is done. Okay. And so, we now have a bunch. Uh, I like probably this one the best. I think it's like it's a little different from the original, right? It's a little different. I mean, like, yeah, the lips are a little bit different. She's wearing the same outfit. So, I'm looking at this one more and like this is obviously an AI image as well. So, we are just putting AI on

[2:41:19] AI on AI. But, uh, yeah, this is fine. So, I'm just going to download this now. Well, actually, you don't even need to download it to be honest. Uh, and then what I want to do is I want to use this as a reference. So, I'm going to reference this here. And then over here

[2:41:34] is going to be my generated woman. Okay. And then what I want is um make it so that the woman is holding my product. Uh maybe in like a bathroom or something like that where it would make sense to put on a midnight clay mask. She should be enthusiastic about using the product. Uh but don't like oversell it. The reason why last time we generated three is just because I wanted to provide myself a lot of um and I'm just going to do this 2K again is because I wanted to give myself a bunch of different options just to make sure that like you know we had diversity, but I'm

[2:42:05] finding that I don't really need that diversity. It does a pretty good job. So, I just one-shoted it and I'm going to see how it looks now when she's holding the element. Um yeah, and this is like a little bit different of an influencer than this chick. You can see like the features are slightly

[2:42:16] different. Very very slightly different. obviously hyper hyper exaggerated and stuff like that. Uh but yeah, we'll we'll we'll check out when see what the gen looks like. While we're waiting for that, I'm actually just going to go back to Claude. I'll go to new and then I will give it um a picture of this product and then I'll have it generate me a script for my VSSL. So, just scrolling down here to uh this midnight clay mask. Okay, actually I don't think I have the midnight clay mask set up

[2:42:44] yet. So, let me go and grab that. We're actually going to go here, download this. So now we've downloaded it. Go back to Claude Code. Going to go to Finder. And then this midnight clay mask is jumping in here. This is a product that I am about to record a UGC style ad for. I want you to generate me three scripts similar to what a real influencer would say talking about this product. Pretend the influencer is a young woman that's found a lot of value in our midnight clay mask. and then make

[2:43:16] it somewhere between 30 seconds or so. Okay, so now uh it's going to read the image and then generate me a bunch of different options. Okay, so I need to talk about this cuz I've been sitting on it for like 2 weeks and I'm mad I didn't try it sooner. This is the Midnight Clay Overnight Mask and I have combo skin that's oily in my T-zone but flaky everywhere else. So most clay masks either dry me out or don't do anything. This one's different cuz you can literally sleep in it. I put it on like a thin layer before bed, wake up, rinse it off, my skin just looks

[2:43:43] calm, like less redness, my pores look smaller, and it's not stripped at all. So, I mean, like, this is, you know, it's not really all that good. Yeah, I think I'm just going to like edit it slightly. So, I need to talk about this cuz I've been sitting on it for like 2 weeks and I'm mad I didn't try it sooner. This is the midnight clay overnight mask. I have combo skin that's oily in my T-zone but flake everywhere else. So, most clay zones either dry me out or don't do anything. But, this one changed the game for me. I can literally sleep in it. Or maybe you can literally

[2:44:09] sleep in it. You just put on a thin layer before bed like this. Then wake up and rinse it off. Your skin just looks really calm. Way less redness. My pores look smaller and it's not stripped at all. This is going to be my little script for the video. Uh let's go back and see how the gens are looking. Cool. So, she's now holding this, you know, in a in a bathroom supposedly. So, that's nice. I'm now going to turn it into a

[2:44:32] video just like this. Okay. All right. Okay. And what I want to do, there are a variety of different models you could choose for this. Uh, let's see. I think I'm probably going to do the Well, we can actually just go test a bunch of this to be honest. Uh, I'm going to try C dance 2.0 fast, but yeah, I'm just I'm going to test a bunch. So, this is a woman doing, let's say, a woman doing a UGC style. Let's say a woman influencer doing a UGC style ad talking about a product. Okay. And I think just for starters, I'm probably going to make it

[2:45:10] really small. And there's also a bunch of elements that you can reuse. So, what I'm going to do is I'm just going to upload that media from before. The reason why is because um if you don't upload media, sometimes it will like not actually keep the midnight clay mask in the right place. So, I'm just going to

[2:45:28] add this to my element. Here it is. Midnight clay mask. This is a clay mask used for product videography. Okay, so now it now it actually has access to the midnight clay mask for the video. So, we're just going to use this as our prop. It's actually added to the prompt box. Uh, and then cool. I mean, I I see cling here. I'm just going to try a bunch of different models. We'll see which ones work best. Okay. And now what I'm going to do is I'm going to turn on sound and then I'm just going to generate multiple. So, I'm going to do, you know, generate 10 with cling. What

[2:45:56] I'm going to do here just so I can test out a couple of different approaches. So I'll do the same thing with Cling Turbo and then I'm going to do the same thing with Cance 2.0. Okay. And after a little bit of playing around, uh, we now have a couple of different ads. So I'm going to turn this on. Hopefully you guys can

[2:46:08] hear this. Okay. So I need to talk about this because I've been sitting on it for like 2 weeks and I'm mad I didn't try it sooner. This is the midnight. I have combo. So as you guys can see, we generated a pretty short one here and I did that just as a test. I wanted to make sure that we could actually hear, you know,

[2:46:20] audio and it didn't seem super crazy. Um, but suffice to say, you can actually do Yeah, you can do a fair amount with this. assuming the woman's moving her mouth. She's like promoing the product. There's even a little zoom in which is kind of neat. Um, so I'm going to try this on a variety of different models and just see what the best one is, at least as the time of this recording. Uh, you can expect to spend maybe like $10

[2:46:37] to $20 doing this initial test round. The whole idea is you just want like a reusable pipeline, right? Okay, so I need to talk about this cuz I've been sitting on it for like 2 weeks and I'm mad I didn't try it sooner. This is the Midnight Clay overnight mask. I have combo skin that's oily and Okay, so this is good. I mean, I like what we what we had here. This is probably a little bit better than Cling 3.0 Turbo. The issue is obviously the time here. Um, you know, that's just not enough time to actually do the ad. So, I think what I'm probably going to do is

[2:47:02] I'm just going to max it out at like 15 seconds. What was the Gemini Omni Flash good for? I think it's only 10 seconds. Yeah, it's only 10 seconds. So, basically, we need to like we need to make this a little bit shorter. Maybe instead it'll just be like you need to try this. This is the Midnight Clay Overnight Mask. It fixes oily skin without flakes. 10 out of 10. would

[2:47:24] recommend change the game for me. Sure. I don't know. Something like that. And looks like I'm doing 16 by9 which is kind of unfortunate, but I'm just going to do a couple of these now. Um, and then see if we can actually make it through the entire script. Then once we have that, we'll have the rule which we can use to generate using Gemini Omni Flash. All right. So, while that's working, let's scaffold out the skill that'll be used to generate these on demand. If you think about it, the idea

[2:47:45] is I'm going to feed it in a product. That's going to be like my initial step. It's then going to, I don't know, use some influencer that I've already sort of chosen um to represent that product. And then all we really need to do is we just need to come up with a bunch of, you know, scripts. And those scripts need to be pretty short. Um, you know, probably somewhere around, I don't know, like 150, 200 characters. So, just checking back on, you know, the gens here. It looks like you need to try this. This is the Midnight Clay Overnight Mask. It fixes oily skin without flakes. 1010 would

[2:48:16] recommend. Change the game for me. Okay. And this is just a little bit too long. You need to try this. This is She's looking all like, but I don't really need to just wait three hours. You need to try this. This is the Midnight Clay Overnight Mask. It fixes oily skin without flakes. 1010 would

[2:48:32] recommend. Change the game for me. Okay, so that one's much better. I think we could probably just rip this ad over and over and over again. Um, so yeah, this is this is good. I think what we need to do is we just need to time this. So, let's check out word counter and then figure out exactly how many words are here. [sighs and gasps] Looks like it's about 146 characters. I think we could probably go to like 170 characters. So, I think that's what I'll do. I'll say 170 character. Uh, well, why don't we just do words instead? 27 to Yeah, 27 to maybe like 30 words or

[2:49:01] so. So, I'm going to go back to Claude and I'm just going to voice dump all of this in as the process. Hey, I'm going to give you an image of a product. I then want you to use the provided image of the influencer to generate a picture or a starting frame of a woman holding the product uh somewhere in an environment that it would make sense. Then I want you to generate 10 script candidates. A script candidate is a script that espouses the virtues of the product, talks about the benefits in an influencer UGC style fashion. Uh, the script needs to be under 35 words or so in order to fit our

[2:49:40] time limit. After you're done, I want you to show me the uh, generated product image with GPT- image-2. And then I also want you to show me the scripts so that I can verify everything's good. After that, I'm going to add a couple more steps to the flow and we'll turn that into a skill. Cool. So, that is what we're going to do. Clearly, we verified that this works if we do it manually. So, let's make this work um digitally as well. Here's

[2:50:08] the influencer and here's the product. I'm going to feed both of these in. Voila. And you can call GPT image 2 using open AI API key. It's in the local environment. Okay. Uh and then going to circle back and just double check that now that we can, you know, we verified we can do this through like manual usage, we need to verify that we can also do this through claude. And then we we're just going to scale the hell out of this thing. I'm sure you guys can imagine we could take this further. Um, as this is building it, uh, just some brainstorming out loud. Rather than just using one woman,

[2:50:43] you could have, let's say, 20 or 30 people, men, women alike, people from different geographies, different ages, different ICPS, different like personas all all over the place. You would do this by using AI to pre-generate a bunch of images of or videos of influencers that you like. uh and then you would also kind of multiply or cross that by all the different products that they're holding in a bunch of different locations. So now what we're doing is we're just multiplying uh a bunch of different options. you then multiply

[2:51:09] those by a bunch of different scripts. And in that way, what you could do and what you know I I I like thinking about AI as what you could do is um you could basically choose let's say 10 influencers, okay? Multiply them by I don't know 10 settings, multiply them by 10 different scripts. Uh the resulting thing is you'll have 1,000 possible

[2:51:31] options. Now, it kind of sucks to look and watch through each, right? So, what you could also do, and I'm just brainstorming. I'm not going to do anything as complicated as this because I don't want to kind of put the cart before the horse. What you could do is you could have some sort of automated image check where every single one of these videos, which are all 10 seconds or so, you break down into 1 second frames. So, now it's like 1 second frame, 1 second frame, 1 second frame. You take screenshots of each. You feed that into an AI model like, you know, GPT uh or Claude Code. Have Claude

[2:52:01] Code take a look at the image and tell you, is there anything malformed that doesn't make sense? If so, just delete it from that problem set or from the option set entirely. After this process is done, maybe you have like I don't know, let's say 250 to 500 that are remaining and each of these are 10 seconds. So, what you can do is you can

[2:52:17] have a human being watch them all. That's 2500 to 5,000 seconds of watch time. And then, you know, if you do the math on two 2,500 seconds, I think that's like 40 minutes to, I don't know, 80 minutes. If you have somebody that then chooses the top, let's say 20 performers of those in the 40 to 80 minutes, you've effectively replaced an entire creative department for, you know, a thousand times the cost of generating this, which is not trivial, right? Uh, but now you have literally 20 new pieces of creative that can make it to the next step. So, I say all this just to show you guys, uh, this is

[2:52:50] extraordinarily possible right now. It's very economical and, you know, you can then test 20 with some small little ad budget, figure out what makes the most sense to scale. So, just going to take a look at this. Yeah, this is her now holding the product, which is more or less exactly what I wanted. Cool. Midnight clay overnight face mask. Fully

[2:53:06] kind of comprehensible and digestible. Everything here is good. What we want to do now is we want to take all this and then just like we did with Higsfield earlier, uh we just want to have that generated through uh I think like an initial start frame uh which we did primarily using Gemini Omni Flash. So the question is how do we you know take uh Gemini Omni Flash and then actually make it kind of accessible within Cloud Code. Well, it involves the MCP and CLI feature that we did before. So if I go to MCP and CLI, you'll see here we have Hixfield MCP and CLI for an AI. I'm just

[2:53:38] going to go to cloud code. Then what I'm going to do is I'm going to copy this and then go back to my cloud code instance here. And then I'm just going to paste this in. Okay. And so I think I actually already have, you know, some form of Higsfield MCP on. So that's okay. I'm just going to say run it. What it's going to do now is install the

[2:53:56] Higsfield CLI globally. And what you have to do after you install a command line interface or a CLI and make it accessible to cloud codes. You have to log in. And that's just how we're going to authenticate. If you think about it, previously we did so in kind of an inefficient way. We had to get the API key from Google uh from GPT uh image 2 directly, right? And then we

[2:54:14] had to like add that to our environment. Well, this is just a simpler and easier way of adding stuff to our environment. So, anytime there is like a CLI option, I would highly recommend it. CLI options are awesome. They also save you just a tremendous amount of time. And uh, you know, it's a very easy and straightforward way of doing this. And so, one thing you guys don't see is it just asked me to sign into Higsfield right over here. So, here's where I'm going to do the sign in. I'm just going to go click um, you know, sign in and then go from there. Now, it's saying,

[2:54:39] you know, hey, we want to access this. Is this authorized? Yes, it's authorized. I can actually close this out. And now it's saying authenticated. Now, the skills can all install. So now we've authenticated inside of Higsfield and we have everything that we need in order to actually use it. Um so yeah, this is fantastic. Now we can actually run this. So what is the next step after this? Well, the next step is we're just going to cross them. So what I'm going to do, okay, is I'm just going to voice

[2:55:01] transcribe. I want you now to take the provided image, the hero frame, and then I want you to use this as a start frame for the Gemini Omni Flash model. Set it to 10second total gen on a 16x9 aspect ratio. The prompt I want you to provide is as follows. Eg. A woman influencer talking about a product. Ensure the provided scripts are reasonably short so we can get them in the whole 10-second gen. Um, every time that I ask you to do this, I'd like you to generate three candidate options. And for those of you guys that are probably wondering what's going on here with the voice

[2:55:42] transcription tool, I like using Claude's built-in voice transcription tool, but um I also have my own. If I just hold Fn on my computer, it'll pop up and it allows me to paste text into any box. So I keep on trying to use the Claude one because I'm trying to showcase Claude code features, but I also just oftent times forget and then use my own. Um same deal. So it's going to go through this whole process and it's actually going to select things for me. It's going to build them and so on and so forth. Um, you know, it's double checking with me that, you know, I'm

[2:56:09] okay with getting build for this. Obviously, I am okay getting build with this. The first time I think it'll probably realistically take uh a little bit more, you know, trial than you'll think. It'll take more credits. It'll just not be necessarily perfect. But over time, you'll be able to make this cheaper. And then once you consolidate it as a skill, uh, you can go. So, I'm going to say yes. And now it'll actually do the generation. And yeah, I mean, like check out these scripts. You need to try this. This is the midnight clay mask. Fixes oily skin without any flakes. 10 out of 10. Would recommend a

[2:56:38] change of game for me. Okay, I have to talk about this. Midnight Clay Overnight Mask. You put it on before bed. You wake up glowing. No effort, just results. Genuinely obsessed with this thing. Like for instance, I don't like that tone of voice, but we could tweak it. [gasps] Uh, if you have oily skin, you need this. Midnight clay works overnight. So, you wake up matte, not greasy. It's the best skincare purchase I've ever made all year. Hands down. I could absolutely see, you know, some woman talk about it

[2:56:59] on like a UGC UGC style video at Crush. And what's cool is um you can parallelize these. So what we've done here is we've actually sent three jobs simultaneously. So we have job A in the queue at the same time as job B in the queue and job C. And we're just going to grab all three of these once they're all done and then combine them back into the output. Uh it's then going to present the output to me just to verify that

[2:57:20] like this did what I wanted it to do. And once I have verified that it has done what I wanted to do, I'm going to turn this into a skill. Anyway, we now have the jobs. So we have the videos right over here. You need to try this. This is the midnight clay mask. Fixes oily skin without any flakes. 10 out of 10. Would recommend. 10 out of 10. Would

[2:57:37] recommend. It changed the game for me. Okay, so we could see we had some weird things there cuz uh the script was a lot shorter than the 10-second video length that tried to like come up with stuff. Okay, I have to talk about this midnight clay overnight mask. Put it on before bed. Wake up glowing. No effort, just results. Genuinely obsessed with this

[2:57:55] thing. Cool. So, that looks pretty solid. If you have oily skin, you need this. Midnight clay works overnight so you wake up matte, not greasy. Best skinare purchase I have made. Okay, so see this is the sort of stuff that I'm talking about. That's like kind of weird. So we just went from having the product like physically like what

[2:58:12] happened there? So you wake up matte, not greasy. Best it just disappeared like in purchase vanishes. So I mean like this is the sort of thing you don't want to happen which is why you're going to need some sort of Q&A. Um that said, I mean people still rip it, right? Like it's not that crazy. Um, maybe she's just a a magician

[2:58:28] that can summon the midnight clay mask. Your hands like I'm more interested, girl, in how you summon the midnight clay mask to your hands. Anyway, um, cool. So, now we we we basically have the skill, right? So, we're pretty good. Um, this is fantastic. We now have everything that we need in order to create a skill. Um, only one tiny change. I'm finding the script for the first video was just a little bit too short. So, the woman had to repeat herself a couple of times. Uh, we need to make shorter scripts just a tiny bit

[2:59:03] shorter on the video side as well. Eg. candidate A could have been done in 8 seconds instead of 10. This will also make us a lot more credit efficient as well. Oh, and then once I'm done with that, I'm going to say, let's now turn this into a repeatable skill. I should be able to simply say, "Generate me uh videos of this using the AI video generator product ad skill." And then you should be able to take the products that I supply you, the influencer that I supply you, and then turn that into a set. Okay. And now it's just going to

[2:59:35] reconfirm the product pipeline with me. So what is the pipeline? We're going to take product images and influencer images as input. Prep resize each. Generate hero frames. Generate 10 script candidates. Show the hero frames and scripts for approval before spending video credits on confirmation. Generate three video candidates per approved script via this model using the hero

[2:59:53] frame um 16x9 8second duration not 10. And then download and deliver all video candidates. So it's not actually correct. Uh we don't want it to uh we don't actually care about this step. So we'll say no step five self select the winner scripts no fixed 8second duration. Sometimes if the uh script is a little longer use 10 seconds if it's a little shorter use 8 seconds. And then if I supply multiple influencers in a folder plus multiple products cross them. eg have influencer A on both product A and product B and influencer B on both product A and product B. Then generate them autonomously and deliver to me in a

[3:00:52] folder. No human oversight until the last step and parallelize where possible. Okay. So, what I'm now going to do is I'm just going to make a folder here. And I'm just going to call this um I don't know, video add gens video add assets. Okay. And then we're just going to combine both of these in here for now. And then the duration rule. What's the actual word count threshold? So, less than 22 words, 8 seconds. More than 22 words, 10 seconds. Script A was 28 words and worked fine at tech 10 seconds, but felt short. Okay. So, as part of me now testing the skill on multiple gens, um

[3:01:30] we have two products. We have the woman with the midnight clay, but we also have a man with uh what looks to be some rosemary scalp oil. So, the idea is we're going to cross them. Um she's going to do this and she's also going to do this and then he's going to do this and he's also going to do this. This should allow us to test what the kind of cross functionality is like and whether or not that works. So, I'm just going to feed this into the skill that we just generated and make sure that this is fine, that we have the skill. It's a

[3:01:53] simple product uh byproduct of just checking to see, hey, you know, can you do this on the video add assets folder and it's just going to look at each image and classify them. I could probably actually have created a subfolder structure that does this. So yeah, we'll give that a try and then have it run. It looks like it misunderstood me the first time. It wanted to generate nine videos per person instead of three, which would have been like,00 credits or something like that. So we've uh cut that down to 360. And it's actually going to be less than 360. I think it's probably going to be like 300

[3:02:23] cuz we're going to do a bunch at 8 seconds instead of 10 seconds. So, we'll see how all that goes. But, yeah, we have woman and mask, woman and scalp, man and mask, man and scalp. And we're going to do three of each of these hero frames. It's now just going to run this through top to bottom. Um, and then also generate uh the scripts, too. Sorry

[3:02:39] about that. My fire alarm just went off. That was pretty fun. Anyway, where we at? Woman x mask. You need this. Midnight clay pulls out all the gunk out overnight. No effort. I just wake up glowing. Generally, the best thing in my routine right now. When in scalp oil, my hair stylist asked what I've been doing differently. It's this rosemary and mint scalp oil. A few drops for a bed and my scalp actually feels alive again. So, this is kind of neat. You can see we're actually implementing that logic where the shorter word ones are 8 seconds, the longer word ones are 10. So, yeah, we're

[3:03:04] we're going through and generating. It looks like there was some issues with some sort of concurrency limit. Looks like you can only have eight jobs at the same time, whereas it is firing 12 at once cuz we did four pairings times three candidates. So, that's okay. We're just going to kind of update the skill to reflect that. And it looks like that's what it is doing. Um, candidates within a script should submit and wait in parallel for speed capped at three

[3:03:28] concurrent. So, we're already six done. We're just going to retry the two failed man pairings one at a time to avoid the concurrency cap, which is kind of neat. [gasps] Um, I'm just going to take a look at this right now. Cool. We actually have the videos up here now that I could test. So, you need this midnight clay pulls all the gunk out overnight. No effort. just

[3:03:47] wake up glowing. Cool. Thank you. You need this. Midnight clay pulls all the gunk out. And see how we're testing different angles. I should move my fat head around. See how we're testing out different angles now? Uh we have different like product shots. You know, this time we're actually showing this. Her hair looks a little bit different. She's actually putting this on in this one. This is the sort of creativity that

[3:04:08] we want. You need this. Midnight clay pulls all the gunk out. Cool. And I like how we're having a lot more adherence to the thing. Uh I think Do we have anyone with men? My hair stylist asked My hair stylist asked what I've been. It's this rosemary and mint scalp oil. A few drops before bed and my

[3:04:23] scalp actually feels alive again. That's cool. Yeah. Yeah. And the timing is basically perfect. My hair stylist asked what I've been doing different. Okay. And so some of these are just going to sound weird, kind of more staccatoy, but that's okay. We're just frontloading again the the gens. Um so I'm just going to go and find the ones of the dudes. Looks like we got them

[3:04:41] right over here. I was skeptical about clay masks, but this one's different. Midnight clay works overnight. There is no mess. No mess at all. And I wake No mess. No mess at all. Wake up looking like I Anyway, obviously that I was skeptical about clay masks, but this one's different. Midnight clay works overnight. No mess. And I wake up looking like I actually slept well and startardy. I was don't know what the hell that was. You know what's really interesting? I think there's just less training data out there on men. Um and so as a result, the the male ads tend to always I was skeptical about clay

[3:05:10] kind of suck more. This audio is really good masks, but this one's different. Midnight clay works overnight, no mess, and I wake up looking like I actually slept well. That's cool. Yeah, I like this one the best. So, you can see how, you know, we need to generate multiple, otherwise we're not going to get like good good uh shots. And who knows, maybe this works better. Maybe having some handsome dude, you know, show off the clay mask is going to work better for a specific core

[3:05:33] audience. You never really know, right? Anyway, so what we have now, we actually have a reproducible skill uh that does this. And uh as mentioned, you know, I think you could realistically run this, I don't know how many times it was. I don't know what our math was, but a lot. Uh you could have run this realistically like every morning, let's say, on something like 10 gens, and then by the end of the month, you would have significantly alleviated the bottleneck in a lot of ad uh production flows. And yeah, you know, well, you can experiment with making this better. You could have like some slight audio track in the

[3:06:02] background. You could make it a little bit noisy. You could add like grain to the videos to maybe remove some of the more characteristic AI components. Like I'd say his skin is just a tiny bit too smooth. It's almost like a little too real. Um yeah, you could have like automated processes that check to see, hey, you know, is the text totally legible on the midnight clay mask? Hey,

[3:06:20] anything weird going on with the person? What's the start frame here? Right? Like cuz it looks like for whatever reason it screws up the start frame on this one. Maybe it's like, oh, every video cut in like three frames so that we always start at a good one. Or I don't know, like if a person doesn't have anything to say, just kind of smile and like

[3:06:35] wave, right? [sighs and gasps] I'll leave it there because I think you guys probably know where to go. Um, all I'm going to do now is do the same thing. Transform it from a loop uh from a skill into a loop and then from a loop into a routine. Exact same idea just with the Higsfield connector instead. And um yeah, then when we're done uh I will have successfully walked you guys through the process end to end of generating a very in-depth series of

[3:06:57] prompts, skills, loops, and routines. Okay. Now, in order to connect HGfield, what you have to do is go up here to MCP and CLI, copy this URL. Okay. Then go back to Claude. What we'll do is we'll scroll all the way down to connectors. Under add, we'll go add custom connector. Add this as the URL and call it Higsfield. And then you will now essentially automatically connect to Higsfield via their OOTH flow. That just means it's going to open up a little window, which it just did in my other screen. So I'm just going to bring that puppy over here. Then you can actually proceed with the login. And now this is

[3:07:31] just another version of what we did earlier. It's just now we actually have the ability to connect to this through um our cloud routines as well. Okay, great. So now that we have this uh if we go back to routines and then at the top right hand corner go new routine can then go cloud. And what's cool is we now have the ability to connect our Higsfield connector. It'll actually automatically populate um every time we do this. And so then I can go down here to our prompt and then I can just ask it hey make this a routine. Awesome. And

[3:07:58] now I just told it make me the routine. What it's doing is creating a tople Google Drive folder for me. This is where all of this information is going to be stored. Now, it's actually uploading the raw source images. Um, the ones that I was showing you guys earlier of the men and the woman and then the clay mask and then the rosemary oil. So, what it's doing is it's going to upload via some tool. Uh, this is I guess some base 64 representation of my file. I don't know. You can turn images into like big long strings and then you can send them over over the wire. Um, and

[3:08:27] that's what this tried to do. It didn't do it a very good job though. Uh, and now it's actually downsizing the images, sending them over, and then it's going to, you know, base 64 encode, upload each. All right. And now we have generated both the uh, loop and the routine. And so you guys could see that um, right over here. If we go back to routines, we have UGC video ads daily batch. So we can connect, you know, the platforms that we need via custom connectors right over here. We also have cloud code remote, which I think is just what Cloud Code always inserts now if

[3:08:56] you want to be able to contact it. Um, and you can see the step-by-step file. So, list the files in the asset drive folder. Uh, for simplicity, that is this folder right here. So, you can see that like they're images that we are providing of the rosemary oil of the product and the people. So, maybe we could make it easier by subdividing that. I don't know. Determine which assets are new since the last run. So, it actually checks to see is there a dated uh subfolder that exists within the AI video gens in drive. So, it's not this folder, it's another folder. If there is no dated subfolder, it'll

[3:09:24] actually go through and then build the cross of every influencer times product. So two influencers, two products means four pairings. For each pairing, it'll generate a hero frame, write 10 adcript candidates, select the best ones, determine the video duration by word count using a rule, generate three video candidates for the pairing, then actually insert all of that into the Google Drive link. So this is how you develop a small granular application in your company using cloud code. Uh this is going to run for cents on the dollar every day. The credits that we're spending on Higsfield gens are going to cost more than that obviously. Um but

[3:10:00] that's okay. Uh you know this is still way cheaper than actually having like a production studio go find influencers to do this with, right? And uh yeah you know you guys now have essentially a full endto-end pipeline allowing you guys to do this for whatever product you guys want. So that is a end to end step-by-step walkthrough of how to basically go through that full like prompt to routine process for you know creative ad generation. I want you guys to know probably looking at the time and you're like looking at the skills that we still have to build and and so on. Um I decided to go really in depth on that

[3:10:32] first one cuz I wanted you guys to see everything involved in the actual production process. Future ones are going to be a lot faster because now that you guys understand how to go from a prompt to a skill, a skill to a loop, and a loop to a routine, we don't actually need to do this whole process every single time. Instead, what I'll do is I'll math it out, make it a skill, and then just automate the process of turning it into a routine, just using a little prompt template. I'm also going to just proceed a little bit faster, and maybe explain things a little bit less,

[3:10:56] and just show you guys what the process of building it looks like. Um, just so that you know, you guys aren't sitting here forever waiting for an agent to finish. Uh, and now that you guys understand how kind of realistic the weight times are and stuff like that, I think we'll all be able to proceed a little bit faster. Hopefully you guys have appreciated this so far. Let's move on to our next major step. What is that

[3:11:13] next major step, you might be asking? Well, now that we're done with creative generation, which is that top offunnel sort of step, I want to work our way down just one more substep before we get to middleunnel with the generation of personalized copy. So, text, templates, and really outreach. And uh I'm going to show this to you guys in the context of two major you know assets. The first is going to be like a newsletter campaign where we're using AI to automat automate

[3:11:37] the process of filling out variables. And the second is going to be like an outbound email campaign where we're using AI to automate the process of entering in little icebreaker snippets and like custom assets and stuff like that. So all of this is in the pursuit of making copy sound like you wrote it yourself even though you probably didn't. And so if you think about it, okay, these newsletters that we're usually sending out now for um flash sales and, you know, multi-step sort of long-term campaigns, nurturing sequences and whatnot. These are pretty great from a marketing perspective, of course, but they suffer from the major problem that

[3:12:14] usually the only real personalization data you have is you have like a first name. So it's like, hey, Nick, sometimes it's not even hey Nick, sometimes just like hello. um you know we're running a last second flash sale from this time to that time and you know you need to sign up now to get 20% off [gasps] and this works okay but what if I told you you could immediately make all of these campaigns work like 5% better right off the bat just through some very basic AI personalization obviously you'd probably do it right the cost of the tokens used to you know personalize these emails is

[3:12:43] way lower than whatever you're currently missing by not doing so so uh yeah you know I'm going to show you guys how to do that and I'm going to show you guys that as an example on my own data which I collect through Maker School and Maker Zero, which is my free community where I host all my resources. So, I have a big newsletter and sort of like mailing list set up for those people. And I'm going to show you guys how to automate that

[3:13:02] just in the context of my own business. The second big thing is cold outreach. So, you know, instead of somebody signing up and entering a bunch of information into some form, right, to sign up to your newsletter. With cold outreach, typically what happens is you'll scrape databases, you'll scrape places like LinkedIn Sales Navigator and so on and so forth. And in doing so, you'll get a bunch of information on uh you know, people that they may or may not have actually given you and that you may have just like stolen from the internet. And so, this is a good example of that. Um here's a big list of people

[3:13:30] with email addresses that I'm not going to leak. They're on the right hand side. But I've basically gone through and I've and I've scraped, you know, about a hundred of these. And all of them have email addresses, all the first names, last names, and everything. And there's a really important column right over here. And the column that I want to point out is this company description column. You see how it says Waltz is an e-commerce company that helps brands grow revenue on Amazon. Advertwise is a digital marketing agency which empowers local businesses to do blank. Should you hire an SEM agency or an in-house person? What if you could get the best

[3:13:57] of both? Lean SEM offers full-time SEM managers. ROI works, Net Booster, Flexi Webb, and so on and so forth. These are all PPC companies that I scraped really simply and easily. And imagine if you will a moment that you wanted to reach out to these businesses at scale with personalized outreach. You didn't just want to say, "Hello, I would like to sell you something." Instead, you wanted to say like, "Hey, Walt, I love Walt go and I've been following you guys for the better part of the last like couple of months. I uh think I could probably solve a major problem for you. Here is the problem that I can solve and here's

[3:14:29] how I would solve it. Let me know if there's any interest in your end. I'd be happy to do it, you know, upfront for the first 30 days, no charge." Which one do you think would do better? Obviously, the latter, right? So, AI affords you the ability to do that. And sorry, my camera just died. But um yeah, I'll I'll loop back around after I've spent some time charging this thing. And that's that's what I want to do here. I just want to show you guys how to essentially use AI both to customize not only, you know, outbound, um not only inbound, but both of them as well, because they're

[3:14:56] both very powerful levers. Right? So, let's start with the very first thing we'll be doing today, which is automating newsletter customization. I'm going to be using a platform called Kit for this, which is a very straightforward email newsletter platform that, as you guys can see, a lot of cool people use. What's going on, Ali? Um, essentially what this does is the same thing that most other email newsletter platforms do. Just allows you to, you know, add your subscribers, get a bunch of information from them. You know, first name, email address, last name, why they signed up, all this stuff, and then send blasts and email campaigns. So, yes, that's right. I am

[3:15:33] among Matthew McConna, Ali Abdal, and the rest of the greats. Uh, it's the platform I'm going to be using today. I'm not affiliated with them at all. Just um, you know, it's one of the many you guys could use obviously for marketing. So, yeah, I'm signed in right over here and I don't want to scroll down anymore because if I do, you guys will be able to see actual live people's email addresses. Um, so the goal is obviously I don't want to show you guys that, but um, it's this is a real account and we're going to be doing something that I'm legitimately doing in my business. Just before we dive super

[3:16:00] deep, I just wanted to give you guys some context. Um, we have 38,000 almost 38,100 people on the mail list. We started this, I don't know, 6 months ago or so. So, we've grown at like an okay pace. The way that we currently do our email newsletter growth for the most part is through LinkedIn and through Instagram. On Instagram, I'll publish videos that are uh basically like lists of tools that you could use. So, like, hey, here's a free AI tool that attacks your vibecoded app and does all this cool crazy stuff. And then up here I have a typical many chat flow where I say comment code to get this free and

[3:16:32] open source to secure your vibecoded app. You then scroll down, you can see a bunch of people are commenting things. Uh I think we might have changed the name in this particular case because we're not getting open codes. We're getting um I don't know like hands or something like that. [gasps] Uh yeah, so this you know agent and then what we'll do is we'll actually comment just message you to the person immediately after. And then um we then start a DM flow with them where we will give them the asset usually in exchange for like their email address or something to my mailing list. Okay, so that's that's

[3:17:01] kind of the context. And then um we also have a bunch of other things that occur. Uh you know the second that they sign up, what we'll typically do is we'll send them an email blast. And the email blast is like a Maker School welcome for instance, which you know Maker School is my product. So uh we'll do like a Maker School welcome. A Maker School welcome just looks like this. Um, hi. Welcome to Maker School. This won't be one of those bloated stuffy onboarding emails. Just the essentials. First, thanks for joining. Make sure to read through our onboarding guide. You'll see them in Maker School every day. Just the

[3:17:31] classics, right? Sort of like the post purchase email that you would see in any major big business. Okay, great. And I I show all you guys this because I find it annoying when I'm watching a course or whatever and we're just doing isolated cherrypicked tasks that don't really kind of weave into a real business. You know, this is my real business. We do uh I think we did $280,000 selling this exact product um last month, which is like pretty damn chunky if I do say so myself. We did so with an extremely lean team. Um and the vast majority of it was just through content and marketing like

[3:18:00] I'm showing you guys how to do right now with cloud code and stuff like that optimizing it. Okay, so what are we actually going to do today? Do you guys see this email here? Hi-ash, welcome to Maker School. Extremely excited to have you with us. Is there anything personalized or customized about this email? No, there's nothing personalized or customized about this email. This is essentially just a mass email that we are sending people. It's the same goddamn thing every time. And you know, somebody that is at the forefront of the AI revolution, um that's kind of annoying because with AI, we don't need to suffer, you know, crappy templated

[3:18:30] emails anymore. We actually collect a fair amount of information about the people that we work with. Um this isn't uh exactly accurate, but you can imagine how if somebody signs up to your mail list, you would probably have information like this from them. You'll have their first name. You obviously have their email address. Maybe you'd have their company name. Maybe you'd have reasons why they signed up. And so I just added a bunch of like fake data here with fake people. So none of these people are real. At least not that I know of. If they are, then uh it's a bunch of monkeys on a typewriter making

[3:19:00] Shakespeare. But you know, I just wanted to give you guys an example of like all the data that you that you could collect and what you could do with some of that data. So, what you could do with some of that difference is instead of just saying, "Hi, welcome to Maker School," you could obviously go, "Hi, Marcus.

[3:19:13] Welcome to Maker School. That's neat." Instead of just saying, you know, you'll really like it here, it's like, you know, I know you want to grow your agency to 20K a month plus. Um, I will help you with this through Maker School or I don't know, hey Sally. um you'll love Maker School specifically because especially because you mention um tech stuff being a bottleneck right

[3:19:38] now. I will solve that for you, right? Like Maker School will deal with that. We're like, "Hey, you know, no, you use HubSpot. Want to make sure you have XYZ cool asset for HubSpot." So this is like basic email segmentation, which most marketers will know, but then we go one level up because rather than just inserting the variables verbatim like most people will do, we're going to use AI to customize the variables so that it makes sense in the context of our mail

[3:19:59] campaign. [sighs] And so this is essentially what I am showing you guys how to do today. I'm showing you guys how to use fuzzy variables. That is sort of the the term. So if I go back to my little draw app here, um the whole idea is, you know, if you guys use hard variables, it's stuff like first name, company, city. These are usually copied straight from the data. And you know, this is stuff that you already see in a lot of like email campaigns and stuff like that. That's fine, but we don't have to suffer that anymore because hard variables are just sort of like they're they're like a

[3:20:30] byproduct of just how difficult it is logistically to make high quality personalized emails to scale. We have a solution for that. So what we do is we're just going to pass all these hard variables through AI basically. And then in doing so can actually just use my damn pen tool. We're going to pass this through AI. And in doing so, it'll turn into I mean that probably wouldn't actually because it's just first name, but it'll turn into a much more enriched variable like why their company is interesting, right? So rather than like taking the verbatim hobbies, Pottery and Indie films, it it'll be like, "Hey Pete, saw you love snowboarding and

[3:21:03] drone flying. I live uh in X location a couple of minutes away from like a massive mountain and I go snowboarding really often." So warmed my heart to see that. Anyway, wanted to talk to you about X today. like you can actually do stuff like that um at scale and that sort of personalization is leverage. So yeah, that's what we're going to be doing. We're going to be turning all these hard variables into fuzzy variables and then inserting them in an email. Um and I have a couple other kind of things I want to show you here um that'll help drive home the point of value. Okay, so here's what I want my

[3:21:35] email to look like after all is said and done. And this is different from how it looks right right now. So just bear with me. What I want to do first is I'm just going to actually call them by their name. So instead of just hi, it's going to be like hey Nick. Uh second, I'm going to remove the n dash because it's kind of AIE and I don't want anything that that's even remotely AI here. So now it'll say hi Nick, welcome to Maker School. Extremely excited to have you with us, especially because short

[3:21:59] plausible reason why they signed up. What this is is a fuzzy variable. You'll see that I use what's called uh camel case where the first letter is lower, right? And then eventually all the other ones are high. It's kind of like a camel in that the head is low and then the there's a bunch of humps underneath after that. Kind of interesting. It's a programming convention. Um and then I typically make them pretty long and they're also quite descriptive. So short

[3:22:25] plausible reason why they signed up. That's actually for the AI agent. That's not for me. Anyway, so extremely excited to have you with us, especially because I know you struggled with client uh or like lead genen in the past or especially because you mentioned you struggled with lead genen in the past. Boom. Now it's like, whoa, whoa, hold on a second. What? You guys are reading all these customized like whoa. Obviously, it's like a it's a strong effect. First, this won't be one of those blood bloated

[3:22:49] stuffy onboarding emails, etc. Um, okay. It covers what to do first, how to get the most out of your group, and our event structure. particularly useful in your case because of short custom reason why it's useful. So what what are we doing here? We're doing the same thing. We're just providing a short custom reason why the onboarding guide will be useful to you. Uh what we're going to do is we're going to take those variables that I just showed you earlier. Okay, pretend that this is just all the data we're getting from people. You know, it's a little much. We're probably not actually going to get that much data,

[3:23:16] but I could see you get like why you signed up and first name, last name probably or maybe biggest challenge, one of the two to minimize friction. They're probably not going to add like 10 questions to your thing. And then down over here, paraphrase challenge. I know a big challenge for you is paraphrase challenge. It's like, whoa, damn. Is

[3:23:30] that Nick speaking to me seriously? Hm. They got their AI stuff unlock, right? So, yeah, this is more or less what we're going to be doing. The question is, how do you do it? Um, so first things first, you know, get an email template and then make the email template really high quality like this. Customize it with variables that you could realistically fill based off of whatever you are currently collecting as part of your email flow. That's sort of step one. Okay. But I'm assuming that you guys are with me here and we've

[3:23:54] already done this. So what's step two? So I'm going to copy this and I'm going to paste this up over here at the top of a prompt. And then let me read out the rest of this prompt to you. I'm not going to read out the entire thing, but I'm going to read out like a fair chunk of it so you see how I'm going to do this. First, I need you to personalize a Google sheet of leads for an email campaign. Here it is. The sheet has six The sheet has columns. first name, last name, email, company name, job title, city, hobbies, why signed up, favorite

[3:24:19] tool, and biggest challenge. This is context for an email template that uses these merge variables. That's just another way to say what this is. Short plausible reason why they signed up, short custom reason why it's useful, and paraphrased challenge. This email is a welcome email for Maker School, Nixar's AI automation coaching community, which is school-based. It teaches people how to get their first AI automation agency client. Their email's tone is casual, direct, no fluff, and first person from neck. Task. Do this in one step without

[3:24:50] pausing to ask clarifying questions. First, read every row of the source Google sheet. Use whatever drive or sheets read tool you have to pull the full row data. Do not guess or sample. Second, for each row individually, write three new column values that are custom to that specific person, not pulled from a shared lookup table or, you know,

[3:25:09] keyed on some sort of repeated thing. Even when two rows share the same text, the generated copy should read differently because you're allowed to draw on their other fields to make each one feel individually written. Vary sentence structure and word choice row to row. Avoid falling into a small set of repeated template sentences. Then it

[3:25:25] actually talks about the uh variables. So short custom reason, short plausible reason, paraphrase challenge. Do not modify the original sheet. Instead, create a new Google sheet. Upload via drive. Do not attempt to write directly. Title it this. include all original columns plus the three new ones appended at the end in this exact order. So this is highly specific. I don't know if you guys could tell but obviously AI helped me write this and then before uploading validate everything and then give me the link to the sheet. Cool. So what I'm going to do is I'm just going to give this to the model. And if you guys

[3:25:56] remember in a previous module we um actually gave AI a Google Drive and so it has access to our Google Drive and it can do all of that stuff. It does not have the ability to um you know create new Google Sheets I don't think but what it can do is it can use drive and drive is you know a way that drive essentially allows you to communicate with Google sheets. Um so what we can do is we can actually like we can spin up a Google sheet despite not having access to edit

[3:26:21] a Google sheet if that makes sense. Yeah. To make a long story short the Google API is highly confusing but um we have the ability to make a new one. We just can't edit it. And that's fine. So, I picked this specific example and I'm doing in this specific way so that I don't have to walk you guys through authorizing Google Sheets, which can

[3:26:38] also be kind of a pain in the ass. Anyway, what's kind of going on under here? Well, it's writing the full data set with handcrafted per row copy. Given 100 rows and three fields, it's going to build this in a Python file as a list of dict writing genuinely distinct copy for each row using their specific job title,

[3:26:52] city, hobbies, favorite tool, context. So, whatever that means, uh, basically what it's going to do is it's not going to like do 100 sets. let's say um one at a time. It's just going to rate like a big huge database um just like regular text by reading the job title, the city, the hobbies, the favor tool, and then just like doing it as a as a little variable. So, yeah, this is probably going to take a couple minutes. I'm not going to waste your guys time. So, let me cut when it's done. Okay. And it's looking like it's now done all of the rows that we need. Um just going to

[3:27:23] build and validate the CSV itself. And now it's doing a Google Drive create a file. So, I want you guys to see this is taking a fair amount of time. It's taking a fair amount of time because I'm doing this in in a session. Um, we haven't actually built a skill for this yet. What we're doing is we're just doing this like the first time. And you can imagine the first time would take

[3:27:39] way way longer than any other time. Eventually, the idea is uh we would just run daily based off of the data that's in our kit. And so, you know, people fill things out, they're in our kit once a day. You know, AI would look through, see all the new signups in the kit, you know, generate all these new variables for them, and then send the send the batch blast. That's a much simpler and easier way of doing it. And then what's cool is we don't actually have to wait and do it in a session because it's happening via a cloud routine. Um, we'll get to that soon. But what's really

[3:28:07] important, and I figure I'll take a second just to talk about this, is you should verify that you can actually do the thing you want to do first. Like always, like don't just trust that you'll be able to do the thing that you want to do. Like actually start at the end, do the thing, and then work your

[3:28:20] way backwards and systematize it after. I don't know if you guys have noticed, but in this course, that's basically what we've done every time. We start with the prompt, then we do the skill, then we do the loop, then we finally do the routine. And every time we verify, is this thing actually doing what we want it to do. And the reason why is so that we don't end up in the shitty situation, which I used to end up in all the time when I started automating stuff, where you have this grandiose idea about all this amazing stuff you want to do and all these great systems

[3:28:43] and then you build them, but then the output just isn't very good. [snorts] And that's because you didn't start at the end. You started at the beginning. So, best just to start at the end, work your way backwards after you've verified that you can get the sort of output that you want. like with our ad generator and then worry about automating it. If you can't get the output that you want, then there's actually no point to automate in the first place since you're automating a process that sucks. So, you can see we actually have the Google sheet right over here or the the CSV rather not a

[3:29:08] Google sheet. So, Marcus Hayes at marcus northparters.com. If we go to short plausible reason why they signed up, you're trying to push north partners past that 20k ceiling and you're betting a automation is the lever. Short custom reason why it's useful. We spend a lot of time in repeatable atband systems, which is exactly what fixes lead flow

[3:29:24] that drives up between projects. Paraphrase challenge. Marcus' biggest challenge is that leads show up in burst instead of a steady monthly rhythm. Okay. And I think that actually probably doesn't work. Blake's biggest challenge. Blake's biggest challenge. Um, let me just go through and read the thing. I know a big challenge for you is paraphrase. Yeah, that doesn't actually make sense. So, clearly there was something lost in translation. I'll have to edit that. Um, I would say these are okay variables right now. They're not the best. They're a little too long. I think um realistically we would we would have written this like 1 2 3 4 5 6 7 8 9

[3:29:56] 10. Probably 10 words or so. So I think I'm just going to like double back. Hey, these look fantastic. Um we just need to make them shorter. So short plausible reason why they signed up should be closer to 10 words. Short custom reason why it's useful should also be closer to 10 words. And paraphrase challenge

[3:30:12] should be maybe five words or so. [sighs] However, in paraphrase challenge, instead of saying the person's name, you need to write it in such a way that I'll be able to insert it into the email copy itself. As a reminder, here is the actual email section that that text will be inserted into. I'm just going to go and copy this. Then I will go back to pasting this. I'm just going to add a slash here and then paste it. So now it'll just redo that last gen part and make it a lot shorter. So I'll

[3:30:42] circle back when that's done, too. What's cool too is this allows you to essentially send a single template and then turn it into a thousand private letters. [sighs] So for instance, hi Sarah, some templated stuff up here. Saw you finished the cold email module. This week's issue builds right on top of that. Hi Mike, some templated message right over here. I know you mentioned on your call you wanted to leave your day

[3:31:04] job by June. This one is for you. Imagine if you had quality like that. only this text is ultimately what's going to change too, which allows you to um build the rest of the template out such that it's very high converting. It's like a high quality offer. And this is different from just having AI write your entire email, right? AI writing your entire email usually sucks. So instead of that, what we're doing is we're having a human being write the

[3:31:28] entire email using practices that work. Then we're just inserting a couple of these variables here or there. So this is a pretty solid strategy and it doesn't really matter where you get your data from. Like for instance, um I'm going to show you guys how to do this in cold outreach in a second. But basically, like the only difference between a newsletter and then cold outreach is, you know, our first step is not data collected by a newsletter signup form. It's data collected or scraped by usually on LinkedIn Sales Navigator or an alternative. So, wherever way you're getting your data, um you guys can do quite a lot with this

[3:31:58] method and uh hopefully you guys will see it as valuably as I do. Okay, so looking over here at the update. So you guys could see now a short plausible reason is you're pushing North P partners past 20K a month. You want to run lead genen for LuminOps that runs while you sleep. You and your co-founder want a lead genen engine. You're launching Harbor Digital's automation agency. You got curious how Claude could sharpen your cold emails. These are all

[3:32:19] far better. They're also way shorter. And typically the fewer AI generated words you have in your email as a proportion of total words, the lower the probability is. Anybody would be like, "Wait a second, did AI write that?" Like your biggest challenge is keeping up with how fast AI changes. That's really cool, right? If you consider that in the context of our U template, it's like, I know a big challenge for you is keeping up with how fast AI moves. So, treat this as an opportunity to solve it. Work hard, put the energy in, and let's just see how high you can climb. That's pretty cool, right? And I mean,

[3:32:46] obviously that's pretty custom. Okay. So, uh yeah, now we are now we actually have a prompt that can reproduce this. Um what I am going to do is I'm just going to um basically upload this list and then like pre preview this so that we could see, you know, whether or not this is what we want it to look like. So, I'm just going to go over to subscribers and I'm going to hide the rest of this from you. Now, in order to do that in Kit, you have to add what's called a custom field. And I think the same might be true for Mailchimp. Um,

[3:33:12] kind of annoying, but anyway, we're going to go to the subscribers tab here. Okay. Now, I'm going to go and add a custom field here. And if you guys think about it, we needed how many did we need? Okay. And then eventually, you will have all of these fields. So, first name, just going to be first name. Uh, last name. We don't actually have a last name, which is unfortunate. Probably shouldn't have added that. There's a bunch of this stuff down here. What we really want is we just want these, right? Like we want short plausible reason why they signed up. We want a short custom reason why it's useful. And

[3:33:38] then paraphrase challenge right here. Then I'm going to go next. And you can see that we can identify this import with a new tag. Sorry for zooming in so much. I just don't want this to, you know, leak a bunch of personal information. Okay. And now what I'm going to do is I just want to preview this so I can make sure that we're actually getting the email. It looks like there's just some issue here with the variable. Well, it did get the first name, but it didn't get any of those expanded ones. So, I'm just going to double check why. Probably the variable syntax is a little bit different. Uh, I

[3:34:06] think if I go here. Yeah. So, short plausible reason why they signed up is almost certainly different. I think they just happen to use a different one. So, that's fine. There you go. Looks like there's just a space between the two and it's uh subscriber and then all of it is lowercase. This is short custom reason why it's useful. We'll add that in. And then this one here is paraphrase challenge. Cool. So now if I go preview and now if I preview as Marcus at North Partnersh you guys could see it says hey Marcus welcome to Maker School extremely excited to have you with us especially

[3:34:38] because you're pushing North Partners past 20K a month with automation. This down here is a second one particularly useful in your case because we build repeatable outbound systems that fix inconsistent lead flow. So we should remove the of. Then finally I know a big challenge for you is getting consistent leads every month. So treat this as an opportunity to solve it. Boom. So, pretty solid. We obviously just have one thing we need to fix it because, right, short custom reason goes here. Just going to exit out of that. And then then we'll do one more test just to make sure that this does indeed work. Uh, where's

[3:35:11] our newsletter? Rachel at LuminOps. So, we'll go back to kit and then I'll go to the preview button. Then, always good to test these things, right? Especially because you want lead genen for Lumops that runs while you sleep. particularly useful in your case because our day-to-day structure turns lessons into shipped automations. I know a big challenge for you is finding time to implement what you learn. Cool. That looks fantastic to me. Awesome. So, what we've done is we've done sort of the proof of concept and we verified that we can do this through a prompt. So, we have to do now is we have to systematize

[3:35:41] it. How do you systematize it? Well, there are like a majillion different ways you could. Um, but a simple one that I think would make sense is if we connect the kit API, okay, so whatever uh their API key is in the back end to Claude and what the skill does is every day, okay, we pointed towards the new email subscribers and like the previous, I don't know, let's just say like 10 hours or something like that, 24 hours, maybe once a day, uh, 24 hours in a loop, I think would make sense. And then

[3:36:07] it goes through and actually enriches. So, it adds the custom fields, these why they signed up, the biggest challenge, the, you know, reasons why it's useful. And then it is now in the campaign set that publishes that broadcasted message. So, I don't know entirely if this is possible, but we're going to have Claude help us. So, I'm actually just going to go back here, close this out, clear this out. Okay. Actually, no. Instead of clearing this out, I think I'll keep this. And then I'm just going to voice

[3:36:34] dump. Hey, my goal is now to turn this into a skill. However, instead of doing this in a Google sheet, I actually want to query the kit API directly. Can you go and find if there are all the endpoints that I would need to do the following? One, every 24 hours, check and see all of the

[3:36:50] new subscribers to the mailing list. Two, for each subscriber that is new, run the query that we just did to create those three variables. Then update the custom fields. Three, um add them to the, you know, welcome maker school broadcast. Is this possible? So now uh what I wanted to do is just go out and do a little bit of research for me. So it's come up with a couple of ways to get what I asked for done. First it'll create a tag called maker school welcome pending. It'll then tag new

[3:37:18] subscribers with it via some API call. Then it'll update the draft broadcast to filter on that tag and then send the set the send at to now and then in doing so it'll send after it's done it'll just remove the maker school welcome pending tat. That's kind of neat and I think that like prevents double messages and stuff like that. So here's what it is doing uh every single day generating these short fields from the profile data, same logic as the sheets exercise and so on and so forth. It will require my kit API key and you know a bunch of other information about it. So I'm just

[3:37:46] going to wait until it's done and then give it all that info. So that's fairly easy to do. You just go to settings and then I think we could probably do this via MCP too now that I think about it. Okay, never mind. I'm just taking a look at this and it's saying that the MCP does not actually cover what we need it for which is kind of annoying. So, we are going to have to do that um you know API key thing. Unfortunate. Just one of those things we tried. Uh we'll go back to the dev tools. And as you guys can see, I have an API key here. I'm just

[3:38:10] going to go V4. We're going to call this for YouTube. And I'm calling it that because I am going to change this immediately after uh the video. Okay. So, first we need to disconnect the kit API. Unfortunate, but you just have to make sure that it's not part of your flow if you want to use the API key route. Then, uh I'm going to go add new conversation. Go to local. I can click on this little thing and I'll call it kit API key equals that. Save it. I'll say okay. Added kit. Let's just say what did I actually call this thing for

[3:38:40] Christ sake. Kit API key. Added kit API key as a local credential. See if you can use it for something. Um we'll see whether or not it actually has access to that. So it's going to try hitting a harmless readonly kit endpoint. If it is not in this shell, the probability of this occurring uh sorry, the situation in which it does not have access to these things is probably just because it's in the same conversation thread as it was before and we just added this in technically a new conversation thread. And then it will say, can you verify if kit API key exists? Make a random query using

[3:39:21] that variable. I added it to local environment. you won't natively be able to see it. So, just make a kit query. And now what it's going to do is it's going to just try to call the kit API key and then um you know it'll essentially allow us to verify in our new thread whether or not that's fine. So now it's testing a query against the

[3:39:44] kit API to confirm that it is valid. Cool. There you go. So it just actually did so and everything is working fine. What I'm going to do now is I'm just going to paste our entire conversation history from the last one which I can find right over here all the way down to the bottom. Then go all the way back up here and paste that in. And in that way I have continued the conversation. I've just sort of taken it out of one thread

[3:40:05] and then pumped it into another thread. So uh what are we going to do here? Um code review first then a readonly verification pass. Yeah, sure. That sounds good. Okay. And we got it. Um, we now have the system that automates the process of adding these paraphrase challenges, short custom reasons why it's useful, and short plausible reasons why they signed up to emails. Um, I will note that I did add a bunch of fake emails in order to do this and I don't want to blast a bunch of fake emails cuz that is how you hurt your deliverability. Uh, but this is now actually pulling real data based off

[3:40:35] challenges from my list and then it is enriching it in this way. Um, so I have that set up right now as a skill, which is cool, but obviously I want to go past just the skill and then I want to eventually turn this into like a fully automated system that just fires once a day on my command. Um, and so, you know, obviously we know how to do that via loops because I just showed you guys how to do that. Um, you could very easily extend what I did via skill that runs locally to a loop that runs locally as well, just on a schedule. But, uh, now

[3:40:59] we just need to turn it into a routine. And if we head over to the routines tab, you can see that we now have the routine created. Run today's Maker School welcome email personalization batch against Kit contacts. This finds new Maker School subscribers, generates three personalized merge fields per subscriber, writes them into Kit custom fields, and sends the existing welcome broadcast to exactly that day's new signups. Uh, you have no local file system access, no blah blah blah. And

[3:41:22] then it gives me a bunch of information. It gives it a bunch of information rather about the kit API_key which I just provisioned it access to by heading over to you know a new routine going into the cloud environment and then um actually like manually provisioning it that access. So you can do so just by clicking on cloud and then

[3:41:39] same thing that little like gear icon. Hold on a second so I can move my head. That little gear icon just to the right of that and that'll open up a thing that allows you to select your API key. Add them in just like we did locally and then uh yeah you're you're good to go. So yeah, I mean like we we've done it we've done this now I think a couple of times and most people here are probably significantly more familiar with the workflow than they were when they started. This is how I would recommend you build these automated systems. You have to hammer it out time and time

[3:42:06] again manually over and over and over until you eventually get what you want. But when you do get what you want, turning it into like a repeatable process is quite simple. Start locally, eventually you'll have it on the cloud. But that's newsletter based. How about cold outreachbased? Well, as mentioned, the only major difference in the pipelines that I built here between uh warm newsletter and cold outreach is it just changes where we're getting the data. So, you know, over here in our newsletter example, we got the data through just people signing up on our website, right? Or, you know, me sending them a many chat link and them opting

[3:42:34] into an offer. Well, over here on our cold outreachbased flow, which is another form of marketing, um what we do is we just scrape their data from services like LinkedIn Sales Navigator. Then, you know, we connect them to a bunch of these different services. Vein is one of them. Air scale is another one. Amplify is another one. You get a bunch of raw rows and then cloud will just read each row and write the fuzzy variables before inserting into some sort of templated platform. Big difference is the templated platforms are just different. Right? So um as opposed to me using kit for instance today I'm going to use a cold email

[3:43:04] platform called instantly. Okay. And this is just the exact same thing that kit does. It's just instead of sending it to people that have already opted in uh to a mailing list and they want to get and receive email from me. This is for people that have never met me in their entire lives and they don't know who I am and I'm trying to sell them regardless. And so I used to do this way back in the day um quite a bit. You know, I would go doortodoor and then when that eventually failed, I would pick up the phone and I would dial and then when that would eventually fail, I

[3:43:31] would start sending like cold outreach via emails and DMs on X and Instagram and all this stuff. And um the degree to which you personalized your comps was like the number one determining variable to how likely people were to reply. You know, if I sent a 100 DMs or emails and they were totally templated and there was nothing personalized. It was literally just like, I don't know, even like, "Hey, Nick, I have something that could help you or something," I would

[3:43:55] get maybe 1 to 2% of people replying. But if I said, "Hey, Nick, love your last vid, man." Or, "Big fan of XYZ business, man. I want to help you with this thing, and I think it makes sense given what you're currently focusing on," I would get like a 5% reply rate. And so, that's like 200 to 250% of the revenue if you just zoom out. uh because every dollar I made was entirely contingent on cold outreach. If I got 2.5 times better at cold outreach, I would make 2.5 times the money. And you think about that proposition with a lot of companies that are doing doortodoor

[3:44:24] cold emails, cold DMs, cold videos, whatever. Um by weaving some form of personalization into your process, you can massively increase the profitability of that business. And it is very much an R or reachbased activity, too. So that's what I'm going to show you guys how to do here. Um, to make a long story short,

[3:44:41] you basically start with the same flow. You will usually have their name almost certainly. Um, we'll usually have some information about their company. Sometimes we'll have scraped bios and posts on social media platforms like LinkedIn, but I'll show you guys how to do this without it. And then we'll have, you know, the icebreer icebreaker actually written by. So, I don't know, let's say this is some, you know, dental firm and there are a bunch of posts that uh talk about hiring hygienists or something like that and how important it is to do X, Y, and Z. Will you hire a hygienist? Is that how you say that? Or

[3:45:11] hygienists. Okay. Anyway, um what you can do is you can have AI take that information and then write a short punchy icebreaker that's literally like, "Hey, loved your post on hiring hygienists." Boom. And then it just goes right into the rest of your copy. Or in this case with Omar at Peak Reefer, uh Peak Roofing. Congrats on the second crew. Imagine you see that immediately after hiring a second crew and making a post about it. You're like, "Hey, this person knows who I am. This person's

[3:45:32] actually watched and read my content. This isn't totally cold outrage. This is somebody that like meaningfully has a connection with me." your open rates and your reply rates are going to be through the roof. Ain't no way you're getting an X on that message. Uh I mean obviously you still will because you know people are quite discerning these days, but it'll certainly improve the odds with which you see success. My best email copy on planet Earth so far has had over a 20% reply rate. And uh you know I was using a method like this combined with conference leads which are leads that

[3:46:02] are going to the same conference as you. I had a bunch of information about their companies. I was able to turn that into a very convincing, very personalized flow. And yeah, I had like a 20 something% reply, a positive reply rate, which means like one in four people that I emailed would send me back something like, "Hey, Nick. Yeah, sure. That sounds great, man. Let's chat. Do you have any idea how much money I made from that list?" It's it's insane. I've also had cold email campaigns that are much more conservative at a 2 to 4% reply rate, and I've still made hundreds of thousands of dollars from them. So, it's

[3:46:27] not that you can't at a sufficient scale make money from cold email, even if your reply rates are kind of poor. It's just you have to be smart about how you do this sort of fuzzy variable generation. Okay. So, yeah, that's what we're going to do and it's going to be sweet. So, first things first, I'm just going to exit out of these kit pages because uh Oh, and I should probably resend my API key now that I think about it. So, I'll do that after. Um, first things first, you know, we got to get leads from somewhere. A really simple and easy place to get leads from is this service

[3:46:53] here on uh a scraping platform called Amplify. I use these ones all the time. It's Pipeline Labs Leadfinder with emails, Apollo, Zoom Info. Basically, it's like a big database of leads and then you just look for the leads based off your filters. So, you'll have filters like job title, I don't know, I'm looking for founders. Seniority,

[3:47:10] okay, I'm looking for seuite people. Email status, you know, needs to be high quality or deliverable. Phone status, okay, this needs a phone so I can also cold call them and so on and so forth. And then you end up just paying for um I guess a $1.50 to $180 per 1,000 leads in order to, you know, actually like connect them and talk to them. So, pretty cool stuff. I'm not going to walk through actually using it. Uh so if I say pretty easy, you just try it and then you enter in some of these filters yourself and then boom, you know, you're you're good. Now we're scraping a bunch

[3:47:39] of leads. What I've done here in this example lead list is I actually went through and I did scrape real leads and all of these are well not all of these but a lot of these are actual real leads and uh you know I just anonymized their information so it's not actually real people. So what you can expect with a scrape like this is you can expect something like this. So you'll have fields like annual revenue, company audience, company city, company country, company description, company domain, company industry, uh LinkedIn URLs. This is like not real LinkedIn URLs but fake ones. Company name here, company specialties, how cool is that? And then I think all

[3:48:14] the way on the right you even get uh their location. So their states, full names of the person, functions, their cities, and even the technologies that they use in their company, which is pretty wild. [gasps] So imagine, you know, that you'd put all of this information into a super smart personalization sequence. I mean, you really think you can't make something that'll at least have the person on either end raise an eyebrow and be like, do I know this guy? Like, of course you can. And that's the value in doing this for marketing uh at scale. So the flow here is going to be really, really

[3:48:46] similar. nothing different whatsoever. Um, just head back to this and then I'll make a new a new chat. [gasps] Um, yeah, very very similar. All we're going to do is instead of doing this via, you know, some sort of like warm list, we're just going to do this via cold list. And usually the only thing we need in this email to be honest is just like a an icebreaker. So, I'm just going to whip up a really quick email for you guys, showing you guys the sort of thing that you can send that can get people to click on it. Um, and then we'll kind of

[3:49:18] go from there. And to be clear, this is more sales than marketing, but I'm still just going to write it out for you guys cuz I think it's valuable for you to see. So, usually it'll be like, hey, first name or maybe yo, first name or something like that. Um, saw um thing in common. I'm thing in common too. Okay. So, okay. Email campaign here. Uh, yo, Nick, saw you liked Labradors. I have one myself and wanted to say hi. I run B2B outbound for agencies that do 5 mil a year and up. We work with Meta and just help them close 500K in the last 45 days. Real talk,

[3:49:57] I've done my research on you and I think Leftclick is fairly close to that number. If so, please hear me out. I want to guarantee you an additional 20 book meetings every month with qualified uh MSP vendors. It is probably another one to two million a year if we play our cards right. Would cost you $0 upfront and I would do all the work myself. And if I didn't hit that number, it would be free, no cost whatsoever. And past that, I get paid mostly a commission. I can guarantee you 20 appointments a month in perpetuity. Are you open to this? If so, let me know and I can start generating

[3:50:23] them for you right away. It would take no more than 30 minutes over the phone. As as mentioned, I'm serious and would guarantee you results. Thanks, Nick. So, is this going to, you know, win me any awards? No, but this sort of email is actually far better than the current state of affairs. Uh maybe by the time that you're watching this, you know, the trends will have changed again. But right now, um like lowercase, very casual language tends to do well. And the reason for that is because people are really used to AI spam where all of the wording and terminology and and

[3:50:50] vocabulary in the email is like perfect. They're used to like lots of M dashes. They're used to perfect grammar. So when they see stuff that's like intentionally or unintentionally not perfect, they start thinking like, okay, this is probably a person, you know? Sure, man. Maybe they suck at spelling, but it's probably a person. And so, they'll actually like read through your email and not just think it's AI garbage. Uh, which is kind of cool. So, believe it or not, what they don't know is we actually used AI for this. Their thing in common, my thing in common is very valuable here. So, uh, you know, there's a lot of

[3:51:18] data in this example lead list, right? I [snorts] don't know, man. I saw you guys use Google Ads or something or like I saw you guys have double click or I saw you guys are, I don't know, in Chicago, right? And like I spent some time in Chicago. Like like you could talk about whatever the heck you want. Obviously, make it relevant. Make it real. Just give AI like a big blurb about all the things about your life and let it do the matching. Uh, but yeah, that's that's about it. So, what we do is the exact same thing as before. I'm just going to copy this and then I'm

[3:51:47] going to dump this in. Then I'm going to say [gasps] above is a cold email template that I've been sending. It's been working quite well. However, I want to kick it up a notch with AI based variables. Anything in the two curly brackets here is an AI based variable. And what these fuzzy variables do is they naturally weave into the context of the email. They don't seem like variables. They just seem like real human personalized segments. What I want to do here is I want to feed you a bunch of information about five leads and I want to see how well you can weave in uh

[3:52:24] their information into an email like this. Okay, you'll be generating these fuzzy variables based off of a bunch of data basically. Okay, so now what I'm going to do is I'm going to go and actually grab like a bunch of our leads. So [sighs and gasps] I don't know, let's say we'll do one

[3:52:40] here. Okay, fine. Why don't we do one here? Then I'll do one here. Then I'll do one here. and we'll do another one here. Okay, so I'm just going to paste this in. Just space this out a bit. Um, here are the five leads that I want you to test this out on. Um, create five email variants so that I can take a look at them and uh I will let you know if the quality is okay. [gasps] You already know everything you need to know about me in order to make this happen. Nick Sarif, feel free to do a little bit of research on me if needed. These are hypothetical

[3:53:18] demo prospects, so no amount of research on them will yield any benefit. So, don't worry about that. just because I've noticed sometimes it does additional research on them too. Okay, so what are we doing? We are having it start doing the search. The leads are fine to test the fuzzy variable concept. Quick flag. The templates numeric claims are the kind of thing that gets outbound account suspended and can cross into deceptive claims territory if the underlying numbers aren't real or substantiated. Yes, yes, yes. Of course, the idea is you can't just pull numbers out of your ass. Like this is just a demo, right? But um you need to make

[3:53:47] sure that you actually have case studies and and proven uh growth and stuff like that. [gasps] Okay. Yo, Luke, saw you're running performance marketing out of Tampa. I've spent a lot of time in the Florida agency scene, too, and wanted to say hi. I run B2B Outbound for agencies that do run $5 million a year. We work with XCO and just helping close 500K in the last 45 days. Real talk. I've done my research on you and the Haven Group is fairly close to that number. If so, please hear me out. I want to book you an additional blah blah blah blah and blah blah blah blah. So, this looks

[3:54:14] pretty solid. Um, let me just see if there were any additional variables here that I didn't notice. Yeah, qualified prospects, right? Qualified western brands entering China. Qualified fintech wellness brands wanting SEO and PPC. Qualified agencies looking to automate their PPC management. Qualified hiring managers looking to build out search and social media teams. And qualified SEO and PPC clients ready to buy. Cool. So, do you see kind of the idea here? Um, we just have AI personalize it, but do so in a very fuzzy way. And I'm sure you guys can imagine if I fed in a big block of information about who I was, all of my

[3:54:47] life experiences, hell, my LinkedIn profile, a bunch of other stuff, and I said, "Here's data that you can use to pull similarities between me and the person, it would do quite quite well. Uh, I'm just not because I don't really want to showcase all of that right now." And, uh, hopefully it's clear. You know, you'd be sending this to people, but you wouldn't really be publishing on the internet, which I think is a little bit different. So, yeah. I mean, um, pretty cool, huh? Pretty cool. So, where do we go from here? Well, we actually need to weave this into, as mentioned, the Google sheet itself. So, what I'm going

[3:55:14] to do is I'm going to go back and I'm going to say, "Okay, great. This is the Google sheet." I'm going to paste it in. This is exactly what I wanted. However, I now need you to actually and actively update a Google sheet uh with those variables. Add these columns to this Google sheet and then customize them for every uh row. The idea is I should be able to insert those variables directly into the email via merge variables and have them be 100% readable and contextually relevant. It's the exact same thing that you just did. I'm just giving this to you in a more structured

[3:55:51] format. Cool. And now it's going to run through a flow very similar to what we were doing before with the newsletter um and the Google sheet except now we're doing it via outbound. Okay. So, it just actually wrapped that up and I ended up proceeding saying great work on the lead enrichment flow. It was quite nice. Uh what I want to do now is set up a skill where when I provide you a list of leads like this, you'll go through and do the entire process automatically. Basically,

[3:56:13] input to the skill is a lead list. Output of the skill is, you know, I want a totally enriched lead list. Enriched meaning um you know, a lead list that has these four additional columns. So, additionally, there were a lot of fields in the CSV. I don't know if you guys can tell, but like man, are these a lot of fields. And I don't really like going

[3:56:30] through all these fields all the time. So, in addition, I said we currently have a lot of fields in the CSV that make it a pain in my butt to upload. So, I'd like you to cut it down. The fields we need are the four customization fields as well as their first name, last name, email address, company name, company size, and so on. So, um yeah, I mean, I sent that out. Uh I did it about an hour ago. I was actually recording the video and for whatever reason, my audio cut off. Um so, I'll just give you

[3:56:53] guys sort of a transcript playbyplay. And [snorts] then it ended up asking me a bunch of questions like, "Hey, what sort of fields do I want? Location fields? what's the input format that I should do? Should it be a Google Sheets link or should it be CSV or pasted text? I ended up walking through it all the way up until we actually ended up getting um you know a trimmed version of this which is a lot shorter and a lot higher quality. And so this is what that trimmed version looks like that now runs via skill. First name Marcus, last name Hayes, Marcus and North partners. As we

[3:57:22] scroll through to the right, you'll see that most of the superfluous fields have now been removed and now we just have, you know, my thing in common, their thing in common, paraphrase, company name, prospects, but their specific terminology. It's very straightforward. Okay. So, you know, we have the skill, I guess, is what I'm trying to say. Um, but I I wanted to extend it and kind of take it a little bit further than just that. And the reason why is because I don't think it's sufficient for me personally just to have like an input output where I give it a lead list and then, you know, it does all the work on

[3:57:49] the lead list and returns it to me. I'd actually like to go further. And I want to go so far that in addition to generating me the lead list, it actually scrapes the leads to begin with. And that sounds kind of crazy, but it's actually fairly straightforward. If you go back over here to Ampify on one of the services that I was showing you guys earlier, um I'll show you guys can actually scrape the leads really easily u manually. And so I'm just going to pipe this in to the flow that we have so that in addition to enriching the leads with all this AI customized stuff. It

[3:58:18] actually goes and finds the freaking leads for me too. We do this thing inside of LeftClick all the time. Uh which is my AI automation agency. I basically build things with cloud code that economize various parts of growth mostly marketing based flows. And as you can see I just pumped in a brief query looking for founders and owners at businesses that were like agencies. So SEO agencies and so on and so forth. You see, we we actually just got a bunch of emails. I mean, I just did this. I just spent 18 cents and then I got a 100 freaking email addresses. What the heck can I do with that? I can do a lot with

[3:58:48] that, right? Anyway, so to make a long story short, um what I want to do is I want to have AI do all this for me. It's not enough for me to like give it a lead list. Um I prefer just to bypass the lead list entirely. And so what I did in addition to that is um you know, I gave it some instructions saying I do all of my lead scraping through a platform called Ampify. What I want you to do is actually steer Apify to scrape the leads for me and then perform this enrichment flow on the leads so that I can just have all of them done in one shot and

[3:59:14] ready for my email campaign. I'll give you all the information you need in order to connect it with Apify like the MCP and the scrapers themselves. And then all I want you to do is just bundle that into the skill and test it on a small subset. Let's do 100 leads to start scrapes because I think that's the minimum. And then you know I just wanted it to enrich 10 of them to tell me um whether or not it it actually ended up working. And so pretty crazy but you can just do this. Um, in my case, the way that you do this with Ampify is, uh, I

[3:59:38] think you have to go via API, uh, or maybe you have to do, if I go back to this actual scraper here, and then I go over here to where it says MCP configurator, it'll give me a URL. So, all you have to do is you actually just copy that URL, go back to cloud, then go

[3:59:54] settings under cloud codes, connectors. You can then create a new custom connector called Appify and paste that link in. And so, this will now just log into Appify for you. And you can see that I actually already ended up doing this earlier. It's the same link right over here. Okay, so that's about how easy it is. It's really not that difficult. Um, fairly straightforward as long as you have the the MCP folder set up. Immediately after it said a

[4:00:16] connected, it can see the MCP tools. Now, let me get the inputs for this scraper and then it actually steers scraper for me. So, I gave it some more instructions as to the sorts of roles that I look for. I usually look for founder, CEO, co-founder, owners, co-owners, partners, co-partners. You can see I kind of went all over the place and I settled on a new idea and my new idea um was as follows. You know, I was thinking about the best way to structure this and I said, you know, how do I typically do it? And it's always a good idea to kind of think about the way

[4:00:48] that you do it. What I do anytime I'm scraping leads for marketing purposes is I'll create a bunch of filters. I'll set them myself, right? Then I'll scrape a list. And then what I'll do is I'll check the list to ensure that the leads in the list um match my ICP which is you know your ideal customer persona and then usually if maybe less than 75% of the leads don't match my ICP you know maybe sorry if less than 75% of leads match basically more than 25% of leads suck, then I'll just redo it and

[4:01:33] I'll do it over and over and over again. And I've realized that this is how I work. So what I did is I explained this to Claude. And then I said, "What I want you to do is scrape a list of 100, then go through 20 out of those 100 leads and determine how many of them are in my ICP. As long as more than 15 out of 20 are in my ICP, we're good. If it's less than 15 out of 20, I want you to go back and I want you to redo it." So what we did is I basically took this SOP and workflow and then I just turned it into

[4:01:58] my skill. Okay? Okay. And you can actually see that right here. It got through. It asked me some questions about this PPC agencies. And then it went it went ahead and did it. And so I ended up actually running this like I think this ran four or five times just redoing it over and over and over again. The first time was 9 out of 20. I think the second time was 9 out of 20. The third time was 15 out of 20. And then finally the fourth time and it just did

[4:02:20] all this entirely on its own. Self-driving was 19 out of 20. So I just sat back. I ran the skill once and then went and actually scraped a giant list of leads for me. In addition to that, I went through and enriched a big list of leads for me with this. And so now we're at the point where it's like, okay, what the heck do we do with all of these leads? Where do we put them? Um, and that's the final step, which I'll show you guys how to do right now. And, uh, hope to god my dang microphone is working this time. Okay. [snorts] So, you know, I was talking about an app

[4:02:46] that's used to deal with leads like this. And, uh, the one that we like using is called Instantly. Instantly is basically a tool that allows you to send emails unass um for various campaigns entirely to a cold audience, people that haven't basically ever heard of you before. And that's different from like the warm audience campaigns which uh are typically managed via things like you know Mailchimp kit and so on and so forth. So, I have a bunch of email campaigns set up here for Clarvo, which is the startup, the namesake of this hat. And uh, you know, we're doing legal intake, staffing and recruiting, automotive, BDC, and so on

[4:03:20] and so forth. Up here, I set up a demo campaign just to show you guys what it would look like if we populated this sequence. And so, what we have here is the same copy that I put together manually is now here in the sequence. Subject line is first name. I have my variables in here. And you're probably wondering, okay, where do we go from

[4:03:37] here? Well, it's actually pretty easy. you can preview um the the forms or rather preview the variables that are filled. And so what I did was I imported that list that I just generated, okay, and enriched. And then I just plugged it in here cuz I wanted to see what it would look like. So Marcus receives an email in his mailbox called Marcus. It says, "Saw you're running a full service Amazon marketing shop. I spent a lot of time around Amazon seller agency circles 2 and wanted to say hi. Looks like our campaign has an extra two. So, I'm [clears throat] just going to remove that extra two. Preview that and then

[4:04:07] see how that how that goes. And that looks pretty good. Maybe I'm going to test somebody else. Um, so you're a Google badge partner running out of downtown Reno. Okay. How about Maya? Performance-based SEO for businesses with customuilt sites. How about rich marketing? Onetrack mind of clients, not just clicks. That one's kind of rough. That one sounds kind of AI, but you're never going to be able to fully escape this. saw you're running an SEO SEM shop out of uh Bissi. I've worked with a few agencies operating out of that region too. Wanted to say hi. So, we now have this actually like in our campaign. And what we could

[4:04:40] do now is we could just send it. And so, that's the skill. Now, obviously, we could turn this into a loop, [clears throat] but it depends on, you know, like would you even need it to be a loop? The only situation which would actually make sense to turn this into a loop is if you knew every single day you were going to make a new campaign. So, every day you needed a bunch of leads available. But this is a good example of like bottleneck thinking. To be honest,

[4:05:00] I don't actually need this all the time. Like I don't need one of my lead lists to be ready for me every morning. Um I could just do it on demand myself. Hopefully you guys see how quick it is. So yeah, you can very quickly and easily turn it into a routine. Cloud routine. You just tell Claude, "Hey man, turn this into a routine. I want it to launch at 5:59 a.m. every morning. Here's the Appify connector. Here's everything else." And I did so um you know in that previous video uh and you know it works well. Routine's right over here. Daily Clervo Legion, it's called. Here's this

[4:05:27] uh big block runs every day at 5:59 a.m. PDT, [gasps and sighs] but I did that more as just like a demonstration that, you know, you can turn anything into a routine. Not necessarily it makes sense for us to turn this one into a routine. Okay, so that that is it in a nutshell. That's how to go basically from a totally manual lead scraping process for marketing services to actually generating these marketing lists entirely autonomously. Um, it's also how to go from like standardized newsletters using kit.com and then turning into like a customized AI fuzzy variable-based newsletter. All very straightforward if you know what you're doing. And

[4:06:02] hopefully I've shown you guys how easy it is to to get out there and do it. And that takes us to, if you think about it, drum roll please, our next uh massive build. So we are now done with our personalized copy, which means we are done with the top of funnel. Nice job everybody. From here we can go down to middle of funnel which is speed to lead for appointment booking via email, SMS and voice. And I just want to outline the problem very quickly because I don't think enough people appreciate just how large of a problem it is. So the problem in marketing is a lot of the time you

[4:06:38] spend tons of money on leads. We're talking like thousands, tens of thousands of dollars. You do so in a variety of ways. I mean, one, you buy lead list, like I just showed you guys earlier, you can do with platforms like Ampify, but [clears throat] probably what you're doing more more likely is you're running ads, okay? And then after a certain number of clicks to that ad, somebody actually ends up booking a meeting with you and and doing some sort of intent, right? So, this is like a typical setting. It's like a booked meeting flow. And uh the issue is you

[4:07:08] spend lots of money on leads, okay? And then you only get a small fraction of that money back. The reason why is because most companies that run advertisements of any kind, hell, even organic brand-based advertisements like I do on my YouTube channel, is they do not treat their leads with the time, energy, and respect that they need in order to, you know, make a ton of money with it. I want you guys to think about leads, okay, this is just some lead as basically being directly convertible to money. This is money, man. And this is me drawing my little, you know, I don't know how to do it, but the dude's

[4:07:54] face, Thomas Jefferson or whatever. This is money. This is just a giant stack of bills. Um, this is a currency that is directly convertible to, you know, freedom dollars, Benjamins, Jefferson's. I'm not American. If it's not clear, I don't actually know what y'all call your money. Um, but you can convert gleeds to money. um if you know how to squeeze out every last drop of it. The issue is most people are just converting at a very low exchange rate right now. So they spend, you know, let's say $10,000 on leads a month and then they will receive maybe $20,000 on their leads a month, which is

[4:08:30] a 2x ros, which is quite poor. Uh means that you're spending 50% of every dollar you earn basically on the marketing to acquire that. It means your cost uh to acquire, you know, a lead is about 50% of what you make. [gasps] The the reality is if you treat your leads better, if you respond to them faster, if you make them think it's more customized and personalized, if you do everything in your power to give them the best experience possible before they pay you, ideally after they pay you, too. But that's definitely not marketing's job. You can turn that single solitary dollar sign into a lot

[4:09:02] of dollar bills, y'all. And so this is more or less what um I actually do all day. And when I mean all day, I mean, you know, like probably like 20 30% of my day. Um, I do it so much that I actually ended up starting a whole business around this with a friend of mine, uh, called Clearville. And the idea behind this is very similar to what I just talked about. You know, you paid for the lead. You just paid, I don't know, like a $30 cost per meeting or something, okay? And if they're in your calendar, you spent $30 on them. You might as well make as much money out of

[4:09:31] that person as humanly possible, right? Just makes sense. So, you know, check in with them immediately. Actually call them ASAP within a few seconds of them filling out your form. send them text messages, follow up and squeeze them like a damn wet rag. I mean, they gave you your information and you bought them, you might as well get the most value out of them. In our case, the way that we do it, and this is just a highle explanation cuz again, I want to make this contextually relevant for people that are watching. um is we will weave

[4:09:57] in flows that call leads immediately. And then we'll also build what are called dispositions, intelligent dispositions, which are outcomes such that if we call somebody and maybe they don't pick up, we route them in the right ways to the right places to maximize the probability of them picking up in the future. We also do things like leave voicemails intelligently. We dial them with good phone numbers that maximize the possibility and probability they'll pick up cuz how many people here get random spam calls a day? Like we all do now. Um we brand them. uh we you know allow call centers uh visible monitoring and stuff like that so we could see sort

[4:10:29] of what people are saying what they're doing and why it's not working. We customized the information on our immediate post uh post ad for form fill calls such that we actually have like AI that can help coach people. We do a lot of stuff and um this is how we deliver absolutely absurd outcomes for a lot of the clients that we work with like a home services company that uh I've mentioned over and over and over again on the channel but they're just probably the best example of this. They were spending god knows how much money on leads before. I don't actually know the figure. We took the revenue from 3 mil a

[4:10:58] month to literally $9 million a month which is 300% revenue growth simply by treating their leads like gold. you know, simply by actually calling these leads ASAP, calling them consistently and repeatedly, and then dialing them with like really good dial practices. [clears throat] Okay, so this is what a speedtole system does really. Okay, and it's um affectionately referred to as S2L in the industry. Those of you guys that just want that spelled out, that's speed to lead. It allows you to squeeze

[4:11:30] significantly more juice out of leads. If before, [clears throat] let's say you get a new lead, somebody that fills out a form or an ad or something, and then I don't know, you have an average wait time of, let's say, 10 minutes before somebody picks up the phone and dials them after a speed to lead system. Okay? They'll fill out the form and then on average, they'll, I don't know, wait maybe like

[4:11:54] 30 seconds. In doing so, the person that filled out the form or the person that signed up to your offer, opted in, they'll still be in the same mindset that they were when they actually did it. The longer that they wait, the lower the likelihood is that they're actually going to want to engage in the offer cuz how many times have you filled out a form for something you don't actually need, but you just kind of felt like you did in the moment, right? Like we all do this sort of thing all day long. So, yeah, that's um that's sort of the idea here. And uh today,

[4:12:20] what I'm going to do specifically with you guys is is I'm going to show you a simple system. Sorry, somebody's blowing up my phone. They do not have branded numbers though, so I'm not picking that puppy up. Is I'm going to show you guys a system that basically you're going to have some sort of, you know, event. So, I'm going to call this intent demonstrated. And this intent can be demonstrated in any way, whether it's a form fill, whether it's a phone call, whether it's an an email to your brand, whether it's an SMS. And instead of it taking, you know, like 10 minutes, um, we're going to cut that down and we're

[4:12:48] actually going to we're actually going to get them talking to you within like 30 seconds. Hell, probably even less. probably like 5 seconds. Uh what we're going to do is immediately after we get the message, okay, we're going to send them a text/ SMS. We're also going to send them a email. Both of these are going to be AI customized just like we customized our outreach a moment ago based off the information in the form fill. The information in the form fill is going to

[4:13:19] be stuff like I don't know their name. It's going to be stuff like why they signed up. It's going to be stuff like their challenges, whatever sort of stuff you collect on your um opt-ins for ads. Then once we have AI do this, then we can even extend this. And I don't know if I'll get to that part today cuz I don't know if that's super important, but we can even extend this to some sort of like merged call. And there are a bunch of services out there that can

[4:13:40] actually assist in the merging of call. That's that's what we do over at Clarabo. But it but it is more kind of intense software that we would have to build where you know there's somebody in a call center or something and they just constantly get merged into calls with people that just filled out a form. Uh and you know because they're getting merged like the second that it happens the person on the other end of the line doesn't have to wait. Like the second they fill out the form 5 seconds past

[4:13:59] they get a phone call. It's like what? You know back in the day I lived with uh this guy Jonno and we were both running kind of like wedding based videography based photography based uh companies. And I remember one day he he made over like $20,000 one month and I was really broke at the time, you know, eating rice and beans for the most part. So I was just like listening through the wall. He was my roommate and I was just like, "What the hell is going on? How's he how's he doing this?" And he had like a thing on his website where it would say

[4:14:23] if you fill out this form, I will call you within 30 seconds or something like that. Maybe it was 60 seconds. And people would fill it out and they would be like, "Well, obviously he's not actually gonna call me in 60 seconds." But he had like a proprietary new, this was six years agoish at the time, version of this. He would text them immediately, be like, "Hey, you know, just saw you filled out the form. I'm

[4:14:41] going to give you a ring in a second. Stay tuned." He'd send them an email. Hey, just following up on the text I sent you a moment ago. Wanted to make sure you had everything. And then he'd immediately call them, too, based using like an automated uh phone merge thing. [gasps and sighs] So, very low hanging fruit, right? But you'd just be surprised at how many companies don't do this. You'd be surprised how many companies make $30,000 a month when they could be making $100,000 a month or 50k instead of like 150 or 200k. And what I want to do is I want to show you guys

[4:15:04] how to do that. um you know using uh speed to lead. Okay, no more perishable leads. Uh we're going to try and reach other people in just a few minutes. We're definitely not going to wait an hour and we're definitely not going to wait the next day. Um the amount of money that your business makes is directly contingent on all of this. So it's very important that we we reach out. So now that I'm done yapping, how the hell are we going to do this? Well, as I'm sure you can imagine, we're going to use Claude. So what I'm going to do is I'm just going to set up like a

[4:15:32] hypothetical form fill. And this hypothetical form fill I'm going to build using cloud just to show you guys that you can also build demos with cloud and little artifacts. And we're just going to pretend for a moment that that uh form fill is like you know an ad based form fill. And what do I mean by ad- based form fill? I mean, like you you literally go on a website like I'm just going to type in like I don't know like HVAC or something and you see that there are uh I don't know people running well actually just interestingly enough there's like only a couple people that

[4:15:57] are actually running ads here which is so so fascinating and they're not even running good ads. Okay, why don't I do this? Let's just go I don't know like air conditioning installation hypothetically. Cool. So you see we have an ad right here, right? Let's say I'm on this page and I'm filling out this form. So, we're basically going to rebuild the flow that would occur immediately after this. And these are exactly the sorts of companies that a speed to lead flow like crushes for, by the way. Like, they tend to be home services. They tend to be blue collar. They tend to be really boring businesses that you never would have

[4:16:26] considered. Uh, but yeah, these guys these guys would crush if they had something like this. And I guarantee you they don't. Maybe I'll fill it out just to see how long it takes for them to reach out to me. But no, I guess I'd have to put my phone number on here, and I don't really want to. Okay, so uh yeah, let me let me run you guys through this. First thing I'm going to do is I'm going to go back to Claude. I'm going to go new and I'll say, "Hey, I'm creating a demo to show people how speed to lead works. Essentially, I want to make a

[4:16:51] simple form fill for a hypothetical error conditioner installation service. Pretend that I'm running an ad for this. When I fill out the form, I want to receive an email and my email address. I'm then going to build a speed to lead system that monitors that email address and then um, you know, emails and then SMSes them immediately after they fill

[4:17:10] out the form. So, help me build this HTML artifact, whatever works, needs to actually send me an email. Yep. Okay. Now, in order to actually receive these emails, we're going to need to um actually fill this puppy out. So, I don't know. I'm just going to do some password. Can I use the term

[4:17:36] password? We'll say, "Hey, baby cakes. Okay. And I'm going to register. Perfect. And uh what I'm going to do is I'm going to use this as like the form fill backend. Basically, you can't actually set up a demo like this unless you have a [sighs and gasps] something set up. Maybe we'll say suggested by AI. What prompted use, dude? Nothing. Let me sign up for the love of God. Okay, we're

[4:18:01] going to add a new form. Okay, and we're back over here. Um I just gave it a little form fill. uh basically something that like wires in an endpoint to allow it to connect to my email. If you don't have something like this, it can't actually send an email directly from this page. So, it's kind of [clears throat] annoying, but u I

[4:18:19] just use a service called Form Scree. It's totally free. It takes like two seconds to sign up. Anyway, you can see we actually have the the form here, which is nice. Um it's nice. Like, it it actually looks pretty good. I'm just going to move my gigantic handsome face out of the way. Um, my girlfriend caught me calling my head really big and fat, so now I call it a handsome head. Uh, Summit error. New AC

[4:18:44] installed in as little as 24 hours. License insured and local. Get a fixed price quote on a new high efficiency system. No sales pressure in a surprise. Sure, whatever. Expect a text and email within 60 seconds. That's pretty cool. So, now what I'm going to do is I'm going to set up a system that actually does this thing. Um, so yeah, let's do it. Uh, it looks like I just need to confirm some form fill service and then it'll actually get sent. And it'll get sent like this, which is about as clear

[4:19:07] a speedtolebased system as possible.

[4:19:14] And and if I go back to my email, you can see it right over here, which is cool. So, we've now verified that this is entirely possible. And we also have information like source, notes, service, home, email, you know, zip code and and all that stuff. [gasps] So, if I refresh this, anything else we should know? Cool. I can put some information in there. Condo, new install, two twotory units about 15 years old. I think what I'll do um make the anything we should know field mandatory and then change it because I want to use AI to take that information and then weave it into a

[4:19:50] cool email reply template. [snorts] So I'm just going to feed that in and then Okay, cool. We now basically have everything we need in order to make this work. So what is the flow actually going to look like? Uh well the flow is going to be fairly straightforward. We're going to start, okay, with the lead filling out the ad. Immediately after, I'm going to build a polling system, which is going to be new to you guys, but it's going to be a system that essentially checks my email constantly to see if emails exist that look like this. And if they do, we are going to send an email to that address

[4:20:25] immediately. We're also going to personalize an SMS. Now, there variety of different ways to do SMS personalization, uh, and stuff like that. I'm going to show you guys a simple way that I do it and then at the end um I don't think we're actually going to do the full call merge, but uh I'll show you guys what you need to do if you did want to do a full call merge like let's say for a company that you're working with that does phone based marketing. And uh yeah, I mean like we can do all this thing. We can also do this in like 5 seconds. We can do this

[4:20:48] not in 5 minutes, in 5 seconds. Um but I think I'll probably spread it out just a little bit so it doesn't seem like too crazy. And that is something worth noting like if you do this too fast people will think it is fake. So, you do have to sort of be in the bounds of reason. Uh, could be like 30 seconds. You could

[4:21:06] insert like a little spelling mistake. That's something that a lot of people will do where they'll say like, "Hey, Pete, just coming back from the office. You know, just driving into the garage right now. I just saw this and I will call you. Can you give me like 3 minutes?" They do stuff like that to make it seem more realistic. It just depends on on sort of how you want to to do it. Um, anyway, cool. So, what's

[4:21:26] going on with your current system? Awesome. Why don't I just take this and load this in a new browser? So, we'll actually go file speak demo. So, this is actually like live. Well, not live. It's just on my computer. Um, and then I'm going to fill this out. So, Nikki wiki, we'll go. There you go. And then we'll go nick at leftclick.aair. That is my business name. And we'll do condo service needed new install. And then I'll say um hm I don't know. Can't make the top floor cold. Please help get my free quote. Cool. You're on the list. Summon air specialist is reviewing your request

[4:22:03] now. I'm just verifying that this 100% works. You can see here in my email that we did just receive it, which is great. So, what am I going to do now? I'm going to go back to Claude and now that we built this demo, I'll say, "Excellent. Now that we have this system, I need a way

[4:22:17] to pull my mailbox for new form fills." Okay. And next up, it is checking my Gmail to see if it can connect to um you know my my address and then grab all the information. I would recommend just doing this through a Gmail connector directly. So you can just go to connector and then go Gmail. I'm going

[4:22:36] to connect this directly. Then I'll go down to continue. I'll select all. And now I can read, compose, and send emails from my Gmail. So then I will go back here. Then I'll just stop this cuz it's on a wild goose chase. Hey, I just added a connector. So, you can uh take a look at that now. Cool. That looks pretty solid to me. And now that we have this information, looks

[4:22:58] like it's skipping the whole OOTH dance. At least that's what it's saying. So, it's going to connect that to nicholas@gmail.com, which just happens to be one of my many emails. And it's even going to go and full pull one full message to see the body format. Hopefully, it's going to pull the nicks one. Let's see. Looks like it's also parsing this out correctly, which is nice. And now it's going to O. Just verifying that the OOTH client is

[4:23:22] usable. Should just try and log me in. Okay, cool. I think it did. Okay. And it looks like we also now have a list of all the leads. And it's even giving me the number of seconds since they got sent, which is awesome. All right. So, from here, we need to build some sort of polling system that actively checks the mailbox uh once every minute to see if there are new entries that aren't already part of, you know, the old list of entries. So, we need two things. We need one, a way to verify that these are new. And then two, we need a way to

[4:23:49] basically take that information and then immediately send an email back to them, you know, asking them questions and then verifying and making it seem real human. I'm going to give you the copy in a sec. Okay, so it's going to go and do that. Um, while it's doing that, I'm just going to go to Google Docs and just write a couple of messages. So, email is going to be a little bit longer, but I think what I'm going to do is I'll just say re first name because that makes sense. And if you get an email, it's like repeater and then I'll say air conditioning. That's going to be the

[4:24:17] subject line. Um, and then I'll say, "Hey, Peter, you know, first name, sorry, just got your request. I hear paraphrased version of their request." Just got your form or just saw you filled out a form. I hear a paraphrased version of your request. Will happily help you with that. I or someone on my team will call you in the next I don't

[4:24:44] know let's just say 5 minutes. Thanks Nick. And maybe just you know for my own sanity I'm not going to use Nick. I'll pretend that the person's name that's calling is Jonah. So hey Nick, just saw you filled out a form. I hear your upstairs is getting pretty hot these days. We'll happily help you with that. I or someone on my team will call you in the next five minutes. Maybe I'll say we'll call you in the next five

[4:25:08] minutes. to book a time to come. Or maybe I won't even mention that and sort this out. Cool. Why don't we do that? Now, um for SMS, typically you want something way shorter. Um there's like 160 character constraints on SMS. So, I'm just going to go here uh and say okay, ooth off me. I'm going to go back here and I'll say SMS limit characters because I just want to see how many. So, it's 160. So, we have to be very very careful about how many characters we're writing in order for this to make any sense whatsoever. Um, just while uh we're doing this, it is going to ask me

[4:25:44] to authenticate. So, I'm just going to verify that it can. And then I'm going to say something like, "Hi, first name. Why don't we start with my email?" Actually, this this is probably already 100 characters to be honest. 176. Yeah. So, it's just a little bit too much. I'll say, "Hi, first name. Just got your request. I know paraphrased version of their request calling you in the next five men

[4:26:17] to sort this out. Jonah, right? So in the worst case, hi, you know, Antonius, just got your request. I know paraphrased version of request and this is quite long as is. So realistically, like this will probably be about 20 or 30 characters anyway. We'll say short paraphrase version of this request calling you in the next five men to sort

[4:26:42] this out and then it'll be Jonah. Maybe we'll actually just make that even shorter. Basically, we want to minimize the probability that this will spill with the AI variable fills. And you know, there is a potential case of them having a really long name. So, if that is the case, that would suck. Okay, so this looks pretty good to me. I'm now just going to fill this in. This is going to be like email copy. This is

[4:27:05] going to be like SMS copy. And then what I'm going to do is I'm going to feed this in to AI and I'll say so this is the copy that I want to send back immediately. Um because we're just working on email right now. Just do the email one. For a paraphrased version of the request right at maximum like 10 or maybe 15 words or so. Keep it really short. The shorter it is the more human it seems. And when you send the email, make sure to send it like really casually formatted just like I provided

[4:27:32] to you. not stray from this template. Um, the 60-cond scheduleuler sounds awesome. Okay. And then we'll fire that off. Okay. And it's looking like we are now sending and receiving emails. This one just came in a moment ago as a test. I'm just going to It looks like for whatever reason, it's like pretty short. I don't know if this is normal. Yeah. Okay. And just while I was doing this, I was experimenting with SMS. So, there's a service out there called Qo. It's formerly called Open Phone. And uh I've had Open Phone for a long time. They've since changed their name and stuff like that, but uh it's just a requirement to

[4:28:04] like have a a b an American business and then set up an American bank account. You need to have like a phone number for legitimacy purposes. And I was just curious to see like, hey, if I had one of these, can I text from it? And I can. Um so, I'm just going to use that as my example because I want to show you guys what an SMS and email combined flow looks like. But if you guys don't have one of these, you can also set them up using Twilio. you do have to wait and actually set up like a phone number via um what's called A2P verification or

[4:28:31] registration. This takes a while. Um and you know I think people on average have waited probably somewhere between like 2 days to maybe like 7 or 14 days. You can also get rejected which is kind of annoying and it's a relatively new um it's a relatively new requirement by the American government to ensure that there isn't just like tremendous amounts of phone spam. So you know it's not a bad thing per se. It's not necessarily horrific, but uh yeah, because of these spam message increases, they've now sort of made it so that you have to wait a little bit. And I had to do A2P2 with

[4:29:04] Kuo, but uh you know, mine's done, so I'm just going to use that as example. Anyway, everything I'm about to show you with Qo is the exact same thing with Twilio. It's just the interface is going to look a little bit different. The same core principles apply. just um you know, whatever platform you use, whether it is Qo or Twilio, which is the main one that most people use, or some other thing, you're just going to go by the API key and just give it to your model. Anyway, just running through the test here, you could see that we just got a new form submission from Dana Whitfield. It

[4:29:31] includes my email address despite the fact that it is Dana cuz I'm I'm sending it back um to me and I obviously want to make sure that email is going to be fine. The upstairs bedrooms never cool down no matter what we set the thermostat to. Unit is maybe 12 years old. So, this is all the information

[4:29:44] that we need. Let's see what happens. Hey, and if we check my sent mailbox, you can see that we just replied saying, "Hey, Dana, just saw you filled out a form. I hear the AC at your house needs a look. We'll happily help you with that. We'll call you in the next 5 minutes and sort this out. Thanks, Jonah." Again, you know, keep in mind, Jonah, Dana, Nicholas, we're all blending together into one formless

[4:30:04] entity. Um, I'm just using my own email. Ideally, this would be something different. One more thing. Now, I want to integrate that phone number flow. So, how am I going to do that? Um, I'm going to go and check to see if they have an API or an MCP that I can plug into. Just looking up MCP. Looks like they do have

[4:30:21] integrations with Claude and Chat GPT. So, let's see. Do we have an MCP here? It looks like we do. Just looking up the service. Um, oh, and it looks like they're actually part of Claude now. So, that's good. I can go back here and then go to settings. Go down to connectors. I'm actually going to type Qo. And it doesn't look like it has my connected one. And so I think that's cuz I need to search through their default ones instead. So I'm going to go back to

[4:30:47] connectors. Then what I want is browse. So I'm looking at theirs. I'm typing qo. I'm seeing this one. I'm connecting now. And this should allow the firing of SMS messages even without me. So there we go. Qo cloud connector is requesting permission to access my business workspace. It's now connected which looks great. Okay. And if I go back to this conversation, I'll also say, could you check to see if you have Kuo access? And if so, can you send a number, a text to my number? Okay. And I don't know if you guys could see this, but I just got a message from this number. U basically

[4:31:32] pretending that I was Dana. Um, hey Dana, just got your request. I know the AC house needs a look. Calling you in the next 5 minutes, Jonah. So, uh, this works, right? We just actually sent the SMS and it only really took me a few seconds. That's how easy it is to integrate different platforms if they are available in the Claude MCP sort of connector list. So, I'm going to go ahead and turn that from a series of prompts into a dedicated skill and then I'm going to show you guys how to loop this. Um, you guys can loop this locally. You can also do it via routine

[4:31:59] since we're cloud connected uh aka we have connector integrations. All of this stuff is capable of occurring entirely autonomously without any of us now. Cool, cool, cool. It's now doing one final test. I just had to reroute it away from building like a script. Uh it just usually wants to do this because it's more computationally efficient. But if you think about it, the whole idea here is we are going to run this like in a loop. We're just going to do it via routines. And speaking of routines, um we are going to set one up here. It's going to be cloud-based. As you can see, we have a Gmail connector. Uh we also

[4:32:28] have a Google Drive connector. We have a Gmail app. I'm going to take that out. A Higsfield. Going to take that out. And I think that's it. We just need Gmail. We need kuo. So with all that in mind, um I'm not going to set this up entirely just yet. I want to see the results of that last test, but I'm just going to have it generate copy for the routine and then run it. We just tested this on a new lead called Priya. Hey Priya, just saw you filled out a form. I hear upstairs is stuck at 80 even with the AC running. We'll happily help you with

[4:32:56] that. We'll call you in the next 5 minutes and sort this out. I also just got a text message saying, "Hi Priya, just got your request right over here." So both of these were basically done immediately, right? And ideally they are done in like a humanized way like this. Some snippet that is short, you know, not super overly long, but still written like a human being would write it. And now what it's also doing is creating me a routine. Um, I asked it to upgrade this and basically turn it into a clawed

[4:33:22] thing. So all of this is good. Everything looked basically perfect. And we are now actually generating the routine.md, which is right over here. I'm then going to paste this in. Just don't want to scroll too far down so I don't leak my damn phone number for the 4 days time. We'll call this speed to lead SMS and email. And then yeah, the

[4:33:40] connectors I want are um Gmail and Quo. And here we're going to want to do a couple different things um in addition. So I mean like we could do it scheduled where like every single time it'll pull, but routines are a little bit different. Like routines um they are capable of getting run via API. What that means is we can basically wait for a trigger and then only respond when we get a request. So if you think about it, what we really could do is we could trigger every time somebody fills

[4:34:09] out that form. Okay, wait a few seconds. That'll land in our inbox and then we'll fire the routine only then. And it's a lot more efficient uh to do it that way. So that's what I'm going to do. I'll go call via API. You could still run the polling approach like I was showing you guys how to do uh locally on your computer. Uh but I think this is probably like a much simpler way. can't really do this API thing if it's local, which is why it's more of a routine based thing. Yeah, this looks pretty solid to me. I'm just going to click create. Now, we have a token that'll

[4:34:35] allow us to fire the routine via post. So, just going to copy it here. I'm going to say here's the token. Please set this up. Cool. And you can actually see these runs working right now. Um, this is actually like literally firing off and being sent. So, initializing the session, starting Cloud Code. Here's a bunch of the the data and the information. And no, I did not leak my phone number. Thank goodness. Um, so it's just going to double check the connectors responsive and actually go ahead and fire them off. Searching for the send and label tools, confirming

[4:35:02] Gmail. Looks like it does have Gmail. And uh, yeah, I mean now we have our little speed delete system set up as a cloud routine and it runs um, via some remote trigger um, sort of API as opposed to something different. So that's that. I mean like we just built a speed delete system. You can obviously take this further if you want to by doing things like automatic call merging. The way that automatic call merging works um just for your convenience is it'll call both of us simultaneously. So there'll be some third phone number that'll call me and it'll call them at the same time and so

[4:35:32] we're both going to get rings on our phones. So ideally you would pick up fast, right? It's trigger based and the first person to pick up is going to be in the call. Um that way there won't be latency like when the other person comes onto the call. But yeah, you could do this with like big setting teams and stuff. And uh the presence of this versus not having the presence of like a call merge system for something that like makes sense to be dealt with over the phone and you know a lot of mid to high ticket offers do makes a real big difference. But think about it. You're

[4:35:59] now slapping them with an email. You're slapping them with an SMS. You're slapping them with a phone call. Really the only missing piece here before you get something that is kind of like Clarvo is just like tracking dispositions and then doing follow-ups. So, pump pumping this into some sort of CRM um you know, scheduling follow-up

[4:36:15] calls by people on your team and so on. That is more sales based. So, I'm not going to talk about that in today's course, probably in a future course. Um but yeah, I just wanted to make sure that we all knew how to squeeze the juice out of the leads to the biggest and best extent possible. And if you think about it, that now takes us to our next task, which you know, now that we've done our middle ofunnel speed to lead for appointment booking and I want to do middle ofunnel data collection and tracking, uh things like UTMs, things like actually building a dashboard and so on. All skills that

[4:36:45] are very important and will serve you quite well in your marketing journey. Uh but first, I am going to go eat a big fat dinner. I'll be right back. Okay, so dashboards. What the hell is the purpose of a dashboard? It's obviously to visualize [clears throat] a data. in our case, marketing data. And you know, a big question is how do we make dashboards that both showcase marketing data in like an effective and simple way while also making use of all of the really cool features of cloud code and uh AI agents in general um to, you know, make them not absolutely ask to look at. So, you know, this is like a

[4:37:17] naive approach that I think most people would probably do. They'd basically just pump in a bunch of their marketing data to cloud code, say visualize it, and then end up with something like this. We're going to make something that is far far better and far higher quality. And we're going to do so in a couple of ways. Um, you know, this is for a hypothetical company called North Beam Digital. If I zoom in here so you guys can maybe see a little bit better. It tracks uh the total ad spend, the sessions, the leads, the blended CPL, closed one deals, one. Um, but the problem is, you know, like when people

[4:37:44] visualize stuff like this, they're not really treating it like what it really is, which is a web design project. Uh, you know, dashboards nowadays are just like HTML. And HTML is hypertext markup language. It's essentially the language of websites and you go a lot further if you treat you know what you are building as uh like literally just like a glorified website um than you know is

[4:38:04] like necessarily a data showcase tool. So yeah, I'm going to show you guys how to make stuff like this way way better and I'm also going to show you guys how to visualize data in ways that allow you to share them with virtually everybody in your organization. If I scroll way down here to my little marketing um section, like you have access to virtually everything that you could possibly need already. You know, meta ads, your spend and clicks, Google Analytics, G4, Stripe, your revenue,

[4:38:29] CRM, email platforms, stuff like that. Your issue is they all tend to be spread between multiple platforms. And so, you know, like all of these realistically will live in different places. The whole idea of like a unified cloud code build dashboard is we get to create a bunch of connectors that connect to these platforms which hopefully you've already

[4:38:46] seen is fairly straightforward, right? You just get an MCP or something and go for it and then build what is technically referred to as an ingestion pipeline. Um, and then your ingestion pipeline just runs on like a daily, weekly or monthly sequence to pump into some sort of dashboard. So that you know that's sort of the idea here. U we're just going to get out of like the super disconnected dashboards and build like a single source of truth. Um what's really cool is you know back in the old days and I say this as somebody that used to run a marketing agency. It was our primary deliverable. Um you know we'd

[4:39:21] have like a report day once a month and uh the whole idea was you know at the end of the month we'd go through all our spreadsheets and stuff like that and put together really nice reports send those reports to the clients and the reports are almost like sales mechanisms. You know we we we try and tune them so that they look as sexy as possible. try and get the client to, you know, review it and be like, "Wow, I'm really getting good return on investment and stuff like that." Um, it was a massive pain in the butt to import data from all these different places. But now the idea is

[4:39:47] we're not necessarily going to have like a report day. It's just going to be a dynamic dashboard that does the ingestion automatically and then, you know, shows the data um pretty constantly. And I think you guys realistically have everything that you need in order to do that, right? Like we've now built loops. We know how to build routines. We know how to build systems that can essentially log into services and grab data without us. So, how far of like a leap is it realistically to like have an online

[4:40:08] asset that is automatically updated? It's really not that big of a leap at all. Um, in addition, I think would be really cool, at least for us, is to take the dashboard that we're building and then provide some sort of like AI questionnaire functionality where not only can we like, you know, talk to our AI about it, but maybe, you know, a client or something can come in on a dashboard and then query an AI maybe locally on the dashboard to tell us stuff about it. So, uh, I'm going to extend our own little marketing dashboard, which is going to look much,

[4:40:34] much better, into something like that. Make it a little bit more dynamic in that way. We'll have basically like an like an like an AI assistant. I'm capable of analyzing stuff. The actual flow itself up here is going to be fairly straightforward. Um, we're just going to build dashboards for a client based off of the information that, uh, we're tracking for them. So, all the KPIs of interest, you know, like meta ads, uh, total number of sessions of their website and stuff like that. It'll then read all the API docs on all the platforms, write the tests and the queries and then generate a dashboard page which is refreshed by a cloud

[4:41:06] routine. And yeah, I mean I don't want to make the whole basis of you know this part just a bunch of AI generated visuals. Hopefully you guys understand we are going to go from super sad frustrated guy over here on the left to super cool chill guy over here drinking his coffee uh with a cute little robot pointing at it. Okay, so that's the idea. How do we actually do so? Well, step one realistically is like assembling your data. Okay. So, um let me just move this a little bit over here otherwise it starts getting kind of

[4:41:32] laggy. Step one is assembling your data. So, I am going to use default sort of templated data here. Um our data is going to include a variety of sources. We're going to have like hypothetical Facebook ads. We're going to have hypothetical like website visits. Um we're going to have hypothetical sales data as well. Um, I know it's not marketing data necessarily, but we're going to have like logs from salespeople. I think we'll even have like email stats and we're going to combine all of this into um, you know, claude as mentioned. Worth noting mine is not going to exist on the cloud. Mine is going to exist in basically like a

[4:42:11] big CSV. CSV just stands for commaepparated value. It's basically like a Google sheet. And the reason why is because I I just don't want to actually pull all of my stats for you guys because a lot of that stuff uh as we saw with the kit marketing emails is sort of proprietary and it's also sort of private. So what I've done is I've just dumped in a bunch of fake data to you know a bunch of files that will just pretend to be our databases. Um now and then once we've assembled the data we're then going to visualize it. Um and what's really cool is we're actually

[4:42:39] going to generate like I don't know five or six different variants. Uh, and so, you know, I think the way most people do this is they'll just try and oneshot it and then they end up with like a really ugly design like this, right? Um, I don't know. I just I don't think this is a very efficient layout. I don't think this is very good for a client. I think you can do way better than that. So, what I'm going to do is I'm going to take that idea and then I'm actually just going to generate a variety of different variants. And then we're going to pick a variant that we

[4:43:03] really like, like B. And then what B is going to do is it's going to make a bunch of other variants. Maybe A, B, C, D. And then maybe we'll pick C. Then C will do a couple variants like A and then B. And then maybe we'll pick A. So now what we've done is we've basically gone through, you know, three different iteration periods to find like the best dashboard ever. And what's cool too is you can just repeat this process anytime if you wanted to visualize different data. Um, so that's going to be sort of my web design process. And then third, we're going to add some sort of like AI

[4:43:29] assistant functionality. [gasps] So why don't we start with the first, which is fairly straightforward, just assembling all of our data. Okay, so this is all of the data in uh my quote unquote database. And as you guys can see, we have a lot. We have DIM channels, clients, sales reps, fact activities, fact ad creatives, agency financials, attribution paths, backlinks, budget pacing, call tracking, uh, number of calls, channel daily. I mean, these are these are massive kind of data dumps, right? U we even have like client summaries and monthly time series. I will show you what a couple of these look like just for your own sanity, but uh you know just for you

[4:44:07] know I think the sake of time and and whatnot. Um I would recommend you just let's find something reasonably big activities. I'd recommend you just assume this is like Google Analytics or something like that. Just going to import this and show you guys what the spreadsheet realistically looks like. So this is fact activities which is going to show a bunch of activities um that include different reps that are doing things. This is total number of calls,

[4:44:29] emails replied and so on and so forth. This is basically like a time log. And so in this way, you know, I'm sure you guys can imagine if you guys have some sort of time logged data within your company. You could figure out, okay, like, you know, how much time has rep four spent, let's say, in the last 24 hours. Well, what I can do is I can sort or filter by activity date. I could clear all the values except for uh well, I guess this is about a year ago, but let's just hypothetically say it's today, which is um August the 6th, right? And then we can filter by, I

[4:44:58] don't know, rep one. And then we can actually add up their total time here. 2,940 seconds to determine how long they've actually spent. 2940 divided by 60 means they've actually only really logged 49 minutes of activities. Maybe then we can make some determination as to their productivity, right? Okay. So, this is the sort of data that we have. Uh a lot of other data as well, like um let me show you just one more that I think is probably a little bit more representative. Um, you know, we have a bunch of sales rep stuff, but I don't know, maybe we'll do G4 data. That probably makes more sense. This is going

[4:45:28] to be a really big one, mind you. 6.1 megabits. So, just uh bear with me here as I actually do the upload. I'm going to append this to the sheet just so that I could visualize both of these simultaneously. We just scroll down a tiny bit. You'll see what that actually looks like. We have over 10,000 rows, right? And so, this activity data was I

[4:45:48] think up to like 30,000 or something. Yeah, approximately. right about here, 50,000, sorry. And then here's where all the Google Analytics data starts. Check this out. I mean, we have like the client ID, the channel, we have the country, the number of sessions on their website, the number of engaged sessions, the bounce rate, the pages per session, and so on. I mean, we we have a lot of stuff here, right? Um, ideally, this would be interpretable to you, but just wanted to make sure you guys see saw like the degree of data we were working with here. It is honestly pretty freaking insane. [gasps] Okay. Anyway,

[4:46:18] so where do we go from here? Um, first things first, if I want to visualize stuff, I'm going to open um up Claude. And then what I'm going to do is in a new chat, okay, I'm going to go and find all of this data, which is in marketing dashboard data. Okay, and then I'm going to hold this button here, and I'm actually just going to feed it a brief description of all of the data, as well as the data folder itself. So now I'm going to say, hey, this is a super detailed um database that contains all of the information on a hypothetical set of companies that I am using to develop

[4:46:49] a dashboard for for a YouTube video that I'm recording. I want you to go through all of the data and for every bit of data, I want you to see if there are ways to combine the databases logically or if the databases are discussing, you know, things that are naturally linked together. Uh then create a highle structure using this. We're going to use this to create the pages of our database. For now, we'll scaffold them really simply. Just provide me a few candidate structures that I can look at and then determine, you know, how to how to lay it out. Okay. And just before I do that, um I want to talk a little bit

[4:47:19] about model intelligence. For most of the course, I've actually been using Sonnet 5, which is like a last generation model. It's not a very intelligent model whatsoever. Um the smarter that you know your needs get, the more intense like your design needs and stuff like that get typically the more intelligent the model you have to use. So I think like last video or maybe the module before I went from Sonnet 5 to Opus 5 just because I was I was wanting to see how much of the errors that the model was making was due to just like its model intelligence and it really wasn't all that much. However,

[4:47:46] for high quality design purposes, I really do recommend using the best models possible and they'll just save you a tremendous amount of time and energy. And so that's what I'm going to do with Fable 5. I'm just going to run it on usage credits. You can see that it's asking me to buy a certain number

[4:47:59] of credits over here. So, I don't know. I'm just going to buy, let's just say, $20 in credits. Uh, make my life a little easy. It'll initiate the purchase and go through. Uh, and you don't need to do this if you have, I think, um, like a max plan. You only need to do this if you're on a pro plan. I actually have like a max pro plan. It's just I've decided not to use it today cuz I want you guys to have the same experience that I do. And I presume most of you guys are probably going to be on, you know, the smaller plan, not like the

[4:48:24] super major plan. So, this is what that UX looks like if you actually buy one. Um, okay, cool. So, we have that. And I don't know, my monthly credit usage limits probably going to be 20 bucks. I'm not going to turn on auto reload. And then what I'll do is I'm just going to um basically rerun this with Fable 5. And then instead of medium, I'm going to go to high. Okay, which tends to be like

[4:48:44] reasonable. I just swapped to Fable 5. I'd like you to revisit these what makes the most logical sense. Assume the analysis done before was by a dumber, less capable model. And so what I've done is I've taken all of the stuff that we had before and I've just fed it to Fable 5 um because you know these models share context. And so this is like the

[4:49:06] Opus 5 recommendations which are okay. Did so off of medium reasoning. Uh I want Fable 5 on high which is just again the absolute most highquality data humanly possible. Okay. So working through it again the structure is actually more constrained than a matter of preference. Three properties of the data set force the answer. The anomaly distribution dictates the primary axis of the seven stories fiber client scoped. One is global, one is rep scoped. A funnel ordered layout scatters each client story across five pages. The deal the death spiral shows up as a traffic blip on one page. They're talking about a specific client. A clientcentric layout

[4:49:39] is the only one each planted story is visible as a story on a single screen. Since the anomalies are the entire demo, this alone settles the primary axis. Two, the coverage gaps punish global pages. E-commerce exists for 4 out of 12 clients. Attribution for 44% of leads. GBP is local business flavored and then the grain mismatches dictate what cannot be client tabs. So here is what it recommends. Overview portfolio grid alert feed performance traffic pipeline search presence commercial and this is client broken down by ID. Um specific

[4:50:06] channels sales teams and then signals. So I like this idea and I'm going to say yes. And you know if there was something here that I didn't like obviously I don't just have to take the model's word for it. This model despite being super smart makes a lot of issue uh a lot of errors as well. Um I just think that this is a reasonable breakdown of the

[4:50:24] data that I was um u working with here. They are like client-based uh and so you can see opus did not do the same thing. Funnel oriented overview acquisition traffic and site client centric agency overview client detail channel analysis candidate C persona question. Are we hitting the number? Where's the money going? Is the site converting? These are all very different approaches and uh Fable basically just jumped right in and then determined the the highest quality approach immediately. So, it's going to scaffold those five pages for me and I'm just going to loop around when it's done. Once it's done the scaffolding, uh then I'm actually going to like turn

[4:50:56] this from just like a scaffold project into like a design project. And I'll show you guys what it looks like when you do cool high quality web design uh with, you know, Fable and these other models. Actually, while I'm at it, I mean, like this is um Fable Design 25. I made this website a while back. Uh it's called Fable 25. And basically, I had Fable just like create a bunch of extremely highquality, well-designed sites. And I say well-designed here in like a relative term. Um these sites are not really the sorts of sites that you'd realistically probably want, you know, to build a dashboard off of, but they

[4:51:27] are very impressive and they they showcase just how high quality Fable really is at stuff like this. So, um I don't know, how about this one here? Right. This is like extraordinarily interactive website where we actually are modeling all of these and we're doing it while also flipping between different font sizes uh and also autocasting them. You know, this over here is like a sexy architecture website with like cool effects, right? You can do pretty crazy stuff these days. This one's like cool typography stuff. So, we don't need to feel constrained by design really to any extent and um Fable 5 can do a really really good job. It's just

[4:52:00] important to approach it in a specific way. Get the data on the page. Once you have the data on the page, then you skin it and select high-quality iterations. You can see the example that it has put together uh which is very in-depth. It's an extremely extremely detailed database. And I mean the design is far better than whatever we uh did before with that much dumber model. As you can see, you also have the ability to click into a specific client, which is kind of

[4:52:23] neat. And you can see the leads per day. You can see the traffic collapse. Uh what they do is they have naive SQL threshold detection over the engagement window. Basically, if something drops off or comes up really really quickly, um you can see it right here. It'll actually show you like it'll say, for instance, hey, there's a weird lead flow collapse happening for CL10, that's client 10 on the week of, I don't know, March uh I don't know, 9th, 2026. On

[4:52:49] average, you guys normally get 57 leads. Now, you guys are getting three leads. So, I think it's like a 5x difference threshold or something like that. And uh that's how you determine whether or not something is weird. But then you also have a breakdown of areas that it kind of makes sense to point out. So, hey,

[4:53:04] you know, here's a conversion rate jump. The baseline percentage is 1.65. Okay, so 1.65% and now it's 2.23. So, what we've done is we basically 1.25 or 1.5xed total results. You can see that sort of law hold across the board here. And so, this is like good information for you to like present to a client. Then down at the very bottom, you can see they're also doing a bunch of stuff on rep rate wins. So, you know, like these are sales representatives and uh you know, this guy has, I don't know, a win rate of 15.6% on deals, so about one in seven calls. Um, team average

[4:53:36] percentage is, oh, sorry, this guy's not a winner. He's a he's a loser. The team average percentage is 24.2%. His win rate is 15.6%. So, it's like, hm, okay, maybe maybe I should do something about that. So, Fable is going through this with me and basically doing the evaluation to try and figure out what's going on. Um, we're clicking through all these different clients. Just making

[4:53:55] sure that like nothing here is is weird. End of all of this. It has gone through and done all of the testing. Um, I just want to verify that I haven't actually ended up using more than my fable usage limit. [sighs and gasps] So, it's looking like the cost of this session was 94% fable $13. I just want to make 100% sure that as something that um just

[4:54:16] worked out. Let me actually go through usage credits. you know, we realistically didn't burn through them. And it looks like for that dashboard, which to be clear is, you know, this controls like 20 clients or something like that and tens of thousands upon tens of thousands upon tens of thousands of rows of data. Like, it's not a it's not a trivial build. That cost me $18 Canadian dollars or about 13 12 or 13 American, which I think you guys are probably more used to. So, that's big, right? I mean, we just spent a lot of money. But the reason why I did this with Fable is

[4:54:48] because now we have basically the scaffold. We have the architecture here. And what we can do is we can take this extraordinarily detailed super in-depth um dashboard and now we can just spin it into like 20 different variants. What do I mean by this? Well, what I'm going to do now is I'm going to go from Fable 5 to Opus 5. Okay. And I'll keep it at um high just for simplicity. And let me just see if I can make this fast. I don't know if I can. Yeah, it doesn't look like I can. That's okay. Fast mode just allows your model to operate a little bit faster. Now, I'm going to go

[4:55:20] back to Opus 5, which um is not being built as part of my usage limits. Uh sorry, it's not being built as like extra credit. It's being built as part of my usage limits. And now I can spin it. Hey, now that we've built this extremely um highquality scaffold, what I want to do is I want to spin this in five different designs. Aka, I want you to generate five highly different variants each. I'd like you to spawn five sub agents all working on this project simultaneously using HTML. Um, go nuts here. Really demonstrate the diversity and range of your visual uh design appeal and give me something that just

[4:55:53] looks really really awesome and interesting on all of them. Key focus should be interpretability obviously uh usability and practical UX. UX that actually does serve uh the user. Okay, sorry it turns out you can't use extra high on Opus 5, I don't think. Um, in case you guys didn't know, you can actually spawn multiple sub agents simultaneously that are all working on things for you at the same time. This consumes vastly more usage limit than if you were to do it in serial or a linear manner like most of us are used to, but it also allows you when you're doing something like creating five different

[4:56:28] designs to design things far faster. The whole principle here is parallelization. And I'm glad we're doing a design project cuz I I did want to talk about this at some point in the program. Um, but basically what we want to do is, you know, if you think about it, the way that we were doing things before is we had one process. Okay, maybe make this a little bit bigger. And why don't I actually draw this on the uh dashboard here? That'll allow you guys to see this um in perpetuity and then we can come back to it. You know what we're doing before is we have one process, okay? And

[4:56:58] then we want to do another process, we want to do another process, and then maybe we want to do a couple more. What we were doing is we were proceeding in like a linear manner. And when you do things in a linear manner with agents, there is a certain time cost. So maybe in order to go from 1 to 2 takes 5 minutes. In order to go from 2 to 3 takes 5 minutes. In order to go from 3 to 4, it takes 5 minutes. In order to go from 4 to 5, it takes 5 minutes. What that means logically speaking is if you add up the total

[4:57:29] amount of time that is 5 + 5 + 5 + 5 20 minutes in total. Well, when you have um parallelizable uh processes and I define a parallelizable process with agents as a process where you can spawn multiple agents working on the same data resource but whose actions are independent and do not impact the actions of other agents.

[4:57:52] Then you can do something really cool. You could start with step one, but then you know what you can do? You can spawn, you know, two, three, four. Well, I guess probably just one more here. And then you can immediately combine those into some next process six such that okay, from 1 to 2 3 4 and 5, it's only 5 minutes. And from 2 3 4 5 and 6 to 2 3 4 5 26, it's only another 5 minutes. So now if you think about it, this whole process here, okay, this is only uh 10 minutes long. And so we're comparing it to a process that's

[4:58:28] previously 20 minutes that has a lot of weight times. Well, now we're just doing that waiting kind of in one shot. We're just immediately generating all of the variants that we want. And this second half might not even be necessary as we will see. But generally speaking, what I'll do is I'll I'll try and combine the best portions of each of these variants and then um stick them in one. So the upside obviously okay if we look at the pro of parallelization it's it's way faster okay it also allows you to search a lot of what's called the solution space which is the space of all possible

[4:58:57] solutions to your problem the cons are you know it's much more expensive so I'm just going to say not cheap [laughter] because you are doing a lot of this and two um in order for you to be able to do this each of these processes need to inherently be like non-moniitored Um, and what that means is, you know, it's like higher I'm going to say higher risk of failure. But it's okay if there's a higher risk of failure because logically speaking um we don't really mind uh you know as long as the models are capable of I don't know going out and doing these things um you know reasonably

[4:59:35] autonomously like it's okay if we do it five times only like one of those actually need to work right or be sufficient to our taste. But anyway yeah that's that's more or less the process. Okay so checking back in on our model you can see that we've now spawned five sub agents. One is doing an editorial style design. Another is doing a terminal style design. Another is doing a Swiss style design. Another is doing a dense operation style design. And the

[5:00:00] last is doing a warm print style design. And two things that built into the brief that matter for the video that I'm recording. Every variant must find the same seven anomalies, which is pretty interesting. I guess it's trying to like improve my ability to make this video because I'd be able to cut between each of these and show you guys what each of the data sources are. [gasps] Okay, so this is the reference scaffold right over here. This is what we were looking at before. Okay. And it's just like bare bones. It's kind of a very neat and kind of sleek design. It's not doing a lot of

[5:00:27] BS here, uh, which I like. But, uh, I wanted I want to kick this up a notch and I want to show us all of the same information just at much higher quality and levels of granularity. Now, what's really cool here on the right hand side, you can see we actually have a monitor that's showing all of the current agents that are in progress. Um if I move my head just a little bit so you guys could see this my my luscious head. Um we can

[5:00:52] see here we have one called Northbeam. Now this is actually the main agent. This is build editorial variant that's agent one. Terminal variant agent two. Swiss variant agent three. Dense ops variant agent four and warm print variant agent 5. So we actually have transcripts of all of these. We can see the tools that and we can actually just jump into this at any point in time to

[5:01:10] see what's going on. And you can see sort of what's happening here, right? It's actually designing like this editorial thing as like a quarterly review style, um, you know, uh, design, which is kind of neat. You can see how deep we're going with some of these designs. Um, I mean, look at that. That is next level. We have basically the same data. It's just being shown to us and sort of portrayed a little bit differently. Dare I say, you know, a little bit better. Um, so I'm looking forward to seeing all these when

[5:01:37] they are done. I will be real with you. This doesn't occur immediately. I mean, I was saying 5 minutes there, not as a uh prescription. Like, it's not going to take exactly 5 minutes. In this case, it's 854, 841, 827, 812, and 756. They're all just like lagging behind each other a little bit. Um, but yeah,

[5:01:53] that's more or less where we're at. Okay. And it's now just doing some final checks. You can see down over here, we are approaching our usage limit, which means we have used a lot of cloud. Um, however, this one is basically finished. and see the same thing happening for I think this terminal variant as well. It's just finishing up now. Uh we can actually see the the website sort of live if I make this bigger. So we can jump to a client very easily just by

[5:02:19] using this little terminal structure. You know, I think I think this is probably one of my better ones. This one here is pretty interesting. It's almost like an auto group. This one here is like the north beam report. This one here is almost like a I don't know. It's like a consulting style channel. Although you could see one of these is a little bit off. And I think this is that original one at the very beginning. So

[5:02:37] I'm going to got 1 2 3 4 5. Cool. Okay. So up here we actually can see the different variants. Um I think I'm going to see if I can go back to number one. But there's now V1's, V2s, V3s, V4s, and then V5s. Um they're unfortunately titled kind of annoyingly. So let me see if I can just copy and paste this. This is dense ops. Take a look at this. I like dense ops. uh looks far better in light mode. I'd say what we see here is overview on the lefth hand side channels underneath. So we can see all the different channels that are contributing

[5:03:09] or not contributing to uh the economics. And this is across all of our clients, not just one of them. Our sales, so our team win rate, total deals won, revenue, closed, average sales cycle, and then signals as well. These signals are the ones that I was mentioning pop up essentially anytime that you um you know have a major increase or decrease in something. So there's wins with conversion rate jumps, but also losses with traffic collapse. Then you have specific client pages. You can actually see their performance, their traffic, you can see the pipeline, you could see um you know rep performance on the account, see who's the best, who's the

[5:03:40] worst, and then search. Like I I hope it's clear this is like a full product in and of itself. It is extraordinarily high quality. And um you know this you can set it up so that it pulls from the database every morning. Um, I've even seen a particular uh company, a big company, you know, they make, I don't know, probably somewhere between 20 to $30 million a year, uh, actually take one of these dashboards and just put it up on their televisions. They, instead of structuring it like this, do so in like a like a slideshow kind of way where the metrics kind of populate as

[5:04:10] the day goes on and then the reps that are responsible for specific things constantly see it in front of them. I mean, you know, maybe that's a little bit more basic and we've been doing stuff like that for a while, but yeah, you can do you can do fantastic work with um Fable nowadays and cloud code. This one over here is that terminalbased style one. So, you can go overview, see everything in front of you here. You can then click into a specific client, which is cool. I like the color scheme of this one a lot. You know, you have LinkedIn ads, Google ads through search, organic

[5:04:38] search. You can drill down like pretty far with this stuff. And obviously, we can add filtering functionality if we want to. [gasps] Um, I like that you can also just go 1 2 3 4 5 6. So that's what I'm doing here with that smooth scroll. So maybe you load up, you deliver a presentation using this dashboard, and you go to two, traffic, you go to three, pipeline, for search. This one over here is much more punchy and engaging, I'd say, but also more like a visual portfolio project. So

[5:05:00] here you can see the client portfolio. You can click into the client itself and actually see the growth over time. I'm a little bit zoomed in, so let me just zoom out, make it look a little bit better. That's what we got here. Um, you could see active, churned, health, MR, all of that. And, uh, yeah, I mean, this one is looking at it, I think, on a 30-day timeline, but yeah, you can obviously change that, too. Pretty clean. Now, the last thing is now that we built a dashboard, we actually need to get this somewhere, and I want to put this on the internet and make this

[5:05:29] accessible to people that are within my organization. So, how do you do so? There are variety of different ways. Um, the way that I'm going to do it is using a platform called Netlify, which essentially allows me just to put static information on the internet for virtually free. Uh, in my case, I think it is free. I think they have like a free plan where it's free forever. You can deploy however many websites you want. Uh, or maybe you can do 300

[5:05:48] websites a month or something like that. But still way more websites than I ever possibly need. So, I already have the site. I'm just going to click login, but you guys can obviously um, you know, sign up to a new one. Once in, we need our Netlefi remote MCP server. So, similar to how we did before, I'm now

[5:06:03] going to go up here, go to settings. I'll then go to connectors, add add custom connector, add this, and then I'm going to call this netifi. And if it starts looking weird like that, you just kind of pull this back and just do um you don't add the npx sections here. All you need is this. Once you're done, click add. It'll then check the

[5:06:22] connection, ask you to connect to it. So, that's what I'll do here. And now I'll authorize this so it has access to my client projects now. And I can go back and so here are all the things that it can do. It can uh I don't know import a cloud design from a URL. It can, you know, set up a new job, create, deploy, and manage my websites. And uh that's virtually what I want. So yeah, as I was doing this, it uh said usage limit was reached. So I did go and I bought I think another $15 of it or something like that. Uh keep in mind that you guys

[5:06:50] don't have to do this. I have just been making like a 5 hour course in basically a day. So 5 hours of a probably much longer course in basically a day. So, the probability that you guys yourselves would breach these loot usage limits as quickly as I did quite low. Um, but what I'll do here is uh just voice discate. My goal is I want to get this on the internet. My favorite was V3. I just connected the Netlefi MCP. I would like you to add this and upload this to

[5:07:16] Netlifi and then password protect it. Make the password 1 2 3 4 5. Okay. So, here is the finished app. Uh, it has both a light mode and a dark mode. this is a great internal thing to be able to send to you or your clients. Uh I'm sure you guys could see how easy it would be to just gate it so that this is just like for one specific client. Maybe you'd use a URL parameter or something like that like you know client equals X and maybe there'd be a special password for them, a different password for somebody else. What we just did is we we

[5:07:42] basically recreated like agency analytics or another one of those sorts of marketing uh visualization softwares and we did so for realistically about like $15 in tokens. Um, but I'll be honest, this is way more data than you would ever ever have on any one of those. You're not just dumping the stuff into a Google sheet here. This is like, as you guys saw, over 50,000, maybe a 100,000 rows of data. So, you also have like a little client health score on the uh side here with like a health thing that has been generated. Um, you could see, you know, their their whole pipeline, their funnel, and where they

[5:08:12] dropped off. So, this would be considered like a warning or a dead zone. And then you can also filter based off of CPL. You can also see there's a big bump in CPL here from an average of about 100 bucks to literally like 1,100. And that whole idea is just so that you guys could see like, hey, you know, what's going on? Oh, there's some sort of alert basis for this client. This client is, uh, you know, suffering for whatever reason. And so, uh, let me see which one was that. Torque Auto Group with CL10. If we go back to signals here, you can see CL10 on the week of

[5:08:39] March the 9th, um, was critical. It got three leads versus its usual 58. uh you know and there were also CPC spikes for 10 as well. Um but I think they were like so great that we're probably not seeing them on this dashboard. Point that I'm making is you guys can design basically your own apps at this point very easily and very um straightforwardly. I would encourage you guys have something like this at least for your own companies. Uh it's very easy to put together along with a data ingestion pipeline. You know you could do this thing in I don't know probably like 30 minutes or so just like I did

[5:09:08] here. Um and keep in mind I was also just using cheap models and I was being quite rate limited as well. um you know, like for my own companies because tokens are not a concern. You know, it's like a business expense and I'm happy to spend a tremendous amount. I think one month I spent like 14 or $15,000 on them. Um I'm I'm fine just like blasting everything with Fable 5, spawning multiple Fable 5 sub agents and then even using fast mode, which is a feature that allows the model, I think, to achieve somewhere around a 2x speed multiple. Um it doesn't really matter to me, but you

[5:09:36] know, in our case, I was trying to do it the way that everybody else does it because I want you to understand and be able to learn. Um, we still got something extraordinarily economically valuable done in just half an hour. And that takes me to that middle of funnel data collection and tracking item now done. All we really have left on the project side is a automation of highquality follow-ups again via email and SMS. We'll use the same um tools that we used back here to do our speed delete. It's just now instead of it being event driven, it's going to be schedule driven. We're basically going

[5:10:05] to come up with the cadence by which we are following up entirely on our own. And I'm going to show you guys that tomorrow. Okay, so today we're going to take everything that we learned in the previous set of builds and then apply it to a really practical marketing activity. And this one also does kind of verge on sales similar to the uh uh not the last one which is the dashboard one before which was speed to lead, but it's still really important and yeah, I think it's worth you guys spending at least a

[5:10:28] little bit of time internalizing that. Um what it's going to be is we're going to automate high quality follow-ups via email and SMS. And so just like last time I used this platform called Kuo, which I'm not affiliated with at all. You guys can do whatever the heck you want. I'm going to use Kuo for this one. Um, but I'm going to walk you guys through how to set it up on Twilio as well. And then over here, I'm just going to use my email connector, which I've shown you guys how to do, too. And this is a really, really high ROI sort of setup. What it looks like for people

[5:10:55] that are unacquainted with this sort of thing is um, most of the time sales and marketing functions within a company will work out of a CRM and that just stands for customer relationship management platform. customer relationship management platforms just have a bunch of leads that are arrayed sort of uh including all of the conversation history between you and the prospect and so on. And so I've actually like built out a really simple customer relationship manager right over here. As you guys could see, we have a bunch of different fields. Let me just make this a little easier to see. Um our hypothetical CRM, and this is actually

[5:11:28] pretty close to my real CRM. Um although it, you know, I made it a little bit different. I'm not just giving everything away. But basically, you'll have like a company name, let's say. You'll have a certain amount of money that the deal is pending for. So, maybe this is a $4,200 deal for Silver Birch Wellness. Uh you have the contact. You have, you know, the date that the invoice was sent. Maybe um I don't know, uh number of days outstanding before people pay you. Uh your due date, you know, when the [snorts] invoice was due and and when it eventually got paid, the various services therein, and so on and

[5:12:00] so forth. And what it really is, if you just think about it, is this is a glorified Google sheet, right? In our case, it's happening on a platform called ClickUp, but obviously you can do this on whatever platform you want. Typically, you'll separate them into different stages. So, these are clients that are actively in progress. These are clients whose invoices are upcoming, who we sent a while ago, and you know, we're like waiting for them to pay for us to start. And, you know, the question behind the CRM is basically, you know, after a while, man, this gets pretty complicated. There are a lot of people

[5:12:24] and a lot of them have different due dates. A lot of them are for different amounts and stuff like that. um how can we stay on top of it and actually make sure to actively follow up with all the people that need to be followed up with that are paying us and or not paying us. And that's where a system like what I'm proposing really comes in handy. So the reality is most people are not following up on their invoices anywhere near as robustly as they should. Also, people aren't checking in with prospects for any uh part of the marketing stack, whether it is beginning the business

[5:12:52] relationship or it's even just like, "Hey, are you making it to the to the meeting that we booked last week?" Um, so if you guys just follow up more, you'll usually make significantly more money. And I don't mean to say that you're going to like, you know, make 100 times the money or anything like that, but you can genuinely add 20 to 30% of the revenue of your organization um,

[5:13:09] simply by solidifying your follow-ups. And you can kind of think about this similarly to how we were thinking about speed to lead before. The whole promise of speed to lead, right? And I cannot draw to save my life today because it's kind of laggy because I'm doing a lot of stuff with agents in the background. The whole promise of speed to lead if you think about it is that we can squeeze out way more juice. So if you know before I would get 10 leads and then only I don't know three of them would come onto the call. Okay, that's a call rate of 30%. After because we're

[5:13:44] contacting them really really quickly and we're following up consistently and stuff like that. Maybe we can make it to six calls. Well this system basically does the same thing. If before, you know, we had 10 potential clients, okay? Uh and now we were only getting, I don't know, two clients, let's say, because I don't know, you know, only 20% of people actually end up signing the paperwork or something in the first place. That's before. Uh after the whole idea is because we're following up so robustly and everything like that, maybe we can turn 10 leads into three clients. This right over here is literally 150% improvement in rev. You know, this over

[5:14:20] here is a 200% improvement in Rev. If you attach these two together, 200* 150 uh or 1.5* 2 rather is 300% improvement in rev, which uh hopefully I'm, you know, not overstating is pretty insane. And if a company was previously making, I don't even know, let's just say $250,000 a month on average, and you apply these systems and take it to $750,000 a month on average, you are producing 500k a month in economic value. And you know, whether you're working for the company's marketing department or it is your company we're talking about here, um, you're going to be compensated a percentage of that. Even 10% of that, 50

[5:15:04] grand a month is pretty huge, right? And so this is this is how I make my money. Hopefully I'm just showing you guys all the sauce because yeah, you can use automations like this to do tremendous tremendous uh things for companies. Okay. Anyway, so yeah, how does this work? We query the the CRM and basically we'll read through all the conversations between me and the prospect historically within my um um email. Then we'll check trigger conditions. So trigger conditions are just a set of things that we're going to put up like, you know,

[5:15:28] does this person actually need a nudge? Have they given us all the information that uh we need in order to determine that they're going to pay or they're not going to pay? What is outstanding? Then we'll just pick and choose a draft from a template library that we've created ahead of time. And then we'll send. And so we're just going to repeat this process over and over and over and over again. Um on my case, I like doing it on a daily schedule. So once a day, we're going to run a system that just loops this over. And as I'm sure you guys can imagine, we are literally going to use a

[5:15:51] loop for this. And in that way, we'll just check to see, you know, hey, who still needs a nudge? Who still needs a nudge? Who still needs a nudge? Because we're using a template library instead of just raw dogging it like most people will do with AI. The quality of the follow-ups tend to be a lot higher and they're closer to your tone of voice, too. So you'll find there are way fewer catastrophic screw-ups because you know the AI is just pulling from like your TVO all the time. And then you also get to send the cadence of how often you want. So for instance, let's say um I

[5:16:17] don't know if they've gone quiet and there's no reply 7 days after the proposal, you can actually check and see, okay, have they replied in the last 7 days? If they haven't, you know, sent any emails or something, then and only then I will send a brief nudge and I'm going to pick from my nudge pool. Um if they ghosted the question, we asked them and they didn't answer, then I'll send a nudge. the nudge will be a little bit different because now it's, you know, to do with them ghosting the question. Hey Pete, let me know if there's anything

[5:16:38] more I can clarify on this. Thank you. If they ask for later, okay, they said follow up next month and it is next month, then we will actually catch that and then we'll follow up then. Then if there's a warm signal, they open the last email a bunch of times, we can send them a nudge, too. These are all just ideas for how detailed you can make the system. Um, and yeah, I think I've just more or less given you guys everything

[5:16:55] that you need in order to make it work. Um, you can make the template libraries as complicated as you want. Um, so yeah, I'm going to do pretty simple ones to start, but hopefully you guys see just how you could evolve a system like this and turn it into somewhere really cool. Okay, so what do we actually do? Well, the first thing we need to do is we actually need to connect this CRM to Claude. And as mentioned, I'm using ClickUp, which is one of many CRM that you could use. Uh, you guys might be doing this in Notion, might be doing this in ASA, you might be doing this in

[5:17:20] Pipe Drive, you know, whatever it is that you're using. I want you to know that's okay. Um, processes are going to be quite similar. either you're going to connect to the MCP or you're going to connect to the API. And I'm going to try connecting to the MCP since obviously that's preferential. So, first thing I'm going to do, and uh this is just what I've gotten in the habit of doing with all my automations, is I'm going to go down to the connector library, and I'm just going to see, do we have a pre-created ClickUp connector somewhere here? I'm going to type ClickUp. If we have a ClickUp connector, we're just

[5:17:47] going to oneshot connect to that. If we don't have a ClickUp connector, then we'll have to do a little bit more complicated work. And as you guys can see here, we actually do have a oneshot ClickUp connector. So I can now just press this button and it'll immediately just ask me to log in. Okay. And so what I'm going to do now is I'm just going to log into Claude. And once I've logged into Claude, it will now try and do the same connection with ClickUp. So this is

[5:18:06] my workspace here called NSM. And I'm now going to connect to that. And now, if you guys think about it, we now have our MCP connector all good to go. It actually just connected to ClickUp, right? Cool. So what I also want to do is I want to find the specific place of this list of mine and uh you know in my CRM where this is is it's inside of the operations space and it's called invoice collection CRM. So what I'm going to do is I'm just going to copy this. I'm going to go back to Claude and then I'm going to say I just

[5:18:37] added and let me voice transcribe this. I keep on forgetting that I can do that. I just added a ClickUp connector. I want you to query a specific list within a space and then get me some of the information therein just to verify that you can actually connect and everything is okay. So as mentioned I always want to start at the end just by verifying can I actually get that information. This is the list and I'll say get me everyone that hasn't been let's see how do I frame this inside of the CRM. Get me anything with a days outstanding of seven days. let's say with a days

[5:19:15] outstanding equal to 7 days. So the whole idea will be this will now um demo my flow and sorry I just reached my usage limit for fable because I don't have any more usage credits if you guys remember from yesterday. So I had to switch that back to 5. Um what it's going to do is it's going to see can we pull tasks that are 7 days late. Uh 7 days late obviously is just like one of the many benchmarks I will use one of the many thresholds. So what it's doing is it's going to figure out how to do this and then we'll sort of have that

[5:19:42] prompted. We'll then be able to use this to build our skill later. Okay. And as we can see, we pulled a bunch of information here on the specific people. So, we have Erica Doyle, amount due on receipts, Dr. Nina Patel, and then Marcus Ben. And these are all different deals. So, what I want to do here is I'm just going to ignore this middle chunk here because there's a bunch of information uh that it looks like it misunderstood surrounding what field I wanted it to pull. I just want to verify, can I also get their email addresses? What are their email

[5:20:08] addresses? So, if I can get their email addresses, then I can actually send a check-in. And it looks like their email address is nickleclick.ai. The reason why is because I just added one of my emails. So, all contact emails route to my own email for demo purposes cuz I want to email them, right? Okay, cool. So, now that I have them, send Erica Doyle this message. Just going to verify this is possible. Hey, Erica, just wanted to check in. How are things going on the invoice? Let me know if you have cues. and I'll say thanks Nick. So, we should have an email to Erica aka

[5:20:44] Nickleftclick.ai with this. Now, it's saying we can't actually do this. We don't have the ability to send from my email. I'm just going to double check why. So, I'll say is there no sending uh tool inside of the Gmail connector. Ask it that question. We'll see how it goes. Let me actually check rather than rely on the tool name list. So, now it's actually going to look through the

[5:21:07] tools. I checked and my claim holds up. The Gmail connector exposes these tools. So, create draft, update draft, no send message, no send draft. Um, so, okay, after finagling it, it's given me three options that we could use to create emails and then actually send them. I was unaware. I thought you could do this in the Gmail connector. Um, doesn't look like you can, but that's okay. Uh, you know, the question is not will I get a bug or will I not be able to do it the way I want to do it, but when I will not be able to do it. So, let me show you

[5:21:35] guys how I deal with that. Um, first I asked, so what are our options? I had to do a bunch of research and now it's giving me Gmail API with your own OOTH client. Let's do Gmail API. What it's going to do now is build a brief pipeline that allows us to do this. It's then going to open up OOTH, which is that like, you know, check-in thing where it like stores my credentials. So,

[5:21:54] it's went it's found some libraries. It's now going to write an authentication script. After it's done authenticating me, I will have to um log in. Basically, still doing just a little bit of testing. Okay. Hey, and I'm now logging in to my email account. So, it was fairly straightforward. The model just asked me a couple questions, gave me a link to click to like give it my uh Google Cloud Console access. And now it's actually going to send a message to my own address to prove that the whole path works without actually touching a client. So, it sent one to me. Let me now take a look. And you can see that it

[5:22:22] has actually done this. Hey Nick, just wanted to check in. How are things going on with the invoice? Let me know if you have cues. Now, we said Erica, but uh you know, it's actually done uh uh you know, Nick. So, a couple things. I'm going to change the subject line now and then I'm going to change the message that we're sending. First, I'm going to say write 10 simple messages like the one you just sent me for checkin

[5:22:40] purposes. Use very light language. Nothing hard, just super soft nudges that fit on one line. Okay. And then um for subjects use straightforward ones like first name or maybe I don't know like invoice first name or you know how are things going etc. So these aren't you know the most pushy of follow-ups but the idea is it's going to run completely automatically on a loop so you might as well um you know be nicer as opposed to be super crazy

[5:23:17] harsh. I've done a lot of following up in my time. There's really smart ways to follow up and there's also kind of hard ways to follow up. So, all of these have m dashes. I'll say no m dashes. We're just going to fix all of that right now. Just checking on invoice. Uh, we don't need the number. Also,

[5:23:35] don't mention the invoice number. So, what else do we got? Cool. And then sign off all of them with thanks dashnick. Cool. That looks good. So, first name, how are things going? Invoice, first name, quick check-in, invoice, following up, hey, first name, anything you need. Check in, invoice. So, now it's just going to write all of these. And the idea is I'm going to I'm going to loop

[5:24:00] all 10 of these into a a follow-up flow. Cool. Looks good. Looks good. We're not seeing any of that invoice information. It's just all pretty simple. So, excellent. This will be our template pool. The idea is we can never send two of these in a row if they're the exact same. They always need to be different templates. So the flow will work like this. To start once per day, you will query ClickUp. You'll determine whether people are 1 day outstanding, 2 days outstanding, 3 days outstanding, 7 days outstanding, 14 days outstanding, 21 days outstanding, 28 days outstanding, uh 56 days outstanding,

[5:24:47] um 84 days outstanding. And what happens is every time that you find a person that is one of those days outstanding, you send an email template. um you fill in the information provided in the merge variables with the customer's actual or rather the prospect's actual uh information. So if it's being sent to a woman named Erica for instance, you should say hey Erica you should not say hey first name and we will slowly and reliably continuously nudge them towards the finish line. One more step just so that we don't end up sending the same email. After you um verify that it is matching one of the days

[5:25:32] outstanding thresholds, go through all of our email conversation history with the prospect to verify and see what sort of follow-ups we've um sent them. That way, we're not going to send them the exact same message twice. Okay, so kind of all over the place there, but I think I got what I wanted to say. Let's see how it performs. Looks like it's going to start by confirming the scopes of my quote unquote token. And uh the history check can't work with the token. So, we're

[5:26:00] just going to have to make a new one. So, I'll say add a read scope to verify. Basically, what we want to do is we just want to like actually check my the real live email to make sure that we're not sending the same thing twice. There are some situations as well in which it might make sense not to send the email at all. Like for instance, if

[5:26:20] a person says, "Please stop emailing me. I've received tons of emails from you. I don't want another one or something like that," then obviously I probably shouldn't email them. So, we'll uh we'll add that functionality in later just uh to verify that, you know, it's actually it's actually working. Cool. Anyway, we did that authentication. Let's double back. So, this is now going through and building uh the flow that I want. It's not actually a skill just yet, but yeah, let's actually run this thing and see what goes on. We're going to send to my email. Um I'm going to go back here first and then open up that. Okay. And

[5:26:48] so now we have, let me see, just running a manual check. Sunhill Property, Pace Orthodontics, Ridgeway Autodetailing. These are the three that should be followed up with. Okay. So I'm going to go back to Claude and I'll say, "Okay, let's run this on a 7day threshold. Again, I want you to go through the whole process end to end and actually tell me everything that you see so that I can verify if the flow is occurring the way I want it to occur." One more change before we do this. There will be some situations in which it doesn't make sense to continue it continue the nudging process. If the prospect

[5:27:23] explicitly tells us not to follow up with them for whatever reason, then flag that on the actual ClickUp task and then don't send the email logically. So now it's actually going to go through the opt- out detection flow and then narrate everything that it is doing. So what ClickUp fields are available for flagging, a notes, and a chase stage with a closed option. Looks like it'll use both of those, which is nice. And

[5:27:45] it'll actually start updating my CRM. Okay, now it's running the full flow. It's found a bunch of my own emails um to myself. So, it's just going to verify kind of what's what. And it looks like clean path earlier is working. So, it correctly remembered templates one and two from earlier selected number three. Now, it's testing the opt out path. Need a real inbound reply containing a stop phrase. So, now they've added that in there basically saying, "Hey man, please don't send me emails. Um, please stop emailing me about this. We are disputing this invoice through our lawyer." Okay, so that's what it came up with. If there

[5:28:14] are situations where that would occur, you should probably not continue emailing the person, right? Kind of makes sense. Um, if that does happen, then you've done something wrong. Okay. Chase stage to close a note and a comment for the audit trail. So now it's actually updating the flow. Let me verify the flag actually landed on the record. So now it's just going to double

[5:28:32] check that its job was done correctly. And now we should actually see what's going on. So it did match the 7day records. It then searched Gmail. It then selected the template high net quick nudge on the invoice. Then had an opt out check just to verify whether or not that worked. And then which recipients got uh went through this flow just cuz I want to verify that was actually those three. Um, why one? There were three with sevenday flags because I never put there were three real contacts. Yes, of course. I want you to run this for real using the info in the CRM. You'll find cloud will do

[5:29:21] stuff like this fairly often because it doesn't want to use tokens, which is generally reasonably positive behavior. So you guys could see here we're 44% of the way through our 5h hour limit which resets in 1 hour and 27 minutes. Um and that's you know just cuz we're on Opus 5 which is a little bit smarter than sonnet 5 which is the model we were using at the beginning of all this. And let's just take a look. How are things going? Hi Nick. How are things going on the invoice? Happy to help if anything's unclear. And then this one here is hey

[5:29:48] Nick. Just checking in on the invoice. Let me know if you have Q. So these were these were sent to our three records basically. Um, it actually went through and it found those three that each have 7-day days outstanding sent to them. Um, it's also then logged a bunch of information about them, I believe. So, this was the uh fake, hey, Nick, please stop emailing me about this. So, we obviously don't run into like crazy stop conditions. And now it's kind of letting us know, hey, you know, Chase stage is set to close, so we're not going to continue following up with them. And we did all this pretty easily just using

[5:30:16] like a demo CRM that I set up in just a few seconds. If you guys run your own real CRM through cloud like the way that I did so today, you can obviously get tremendous alpha. So, what are we going to do now? Well, I basically want to extend this and turn this into a full endto-end flow. So, uh I want to check all of those dates. Hey, this sounds great. I now want you to convert everything that we've talked about so far and all of the fixes therein into a skill. The idea is I will run the skill once every morning. For now, I'll do it

[5:30:40] myself, not as a loop. But we'll run the skill once every morning. And then it will go through uh 1 2 3 7 14 28 56 uh 84 and so on and then match to the templates as appropriately. Um we'll also make sure to do that opt out check and everything else that we talked about. Uh convert it into a simple bulleted SOP so that a future cloud instance that does not have the context of our conversation will still be able

[5:31:08] to execute this reliably. Excellent. So now I'm going to go from that prompt to a skill stage. And then I'm going to retest this. And it's mentioning that the connector is an MCP integration that only works inside a cloud session. That's great. That's basically exactly what we want. The skill is now a standard operating procedure that every time I run it is just going to look for that script. It's going to use its own built-in integration. It's going to run it and then we're basically good to go. Um, if you think about it, what we need to do

[5:31:32] now is we need to turn it into a loop. So this looks great. I'd like now to turn this into a loop that fires every morning at 5:59 a.m. On the local credential side, how could we turn this into a routine where we can run this using a cloud instance as opposed to locally on my computer? You mentioned the MCP integration for ClickUp. I know we can add that as a connector in the cloud environment. How can we do the same with the local uh script? Okay. And after all is said and done, we now have a routine set up for the daily invoice/proposal chase. I'm sure you

[5:32:05] guys could see how you could extend this virtually arbitrarily to whatever the heck you want. Um, I just did mine on that 1 2 3 74 218 56 and 84day schedule. Um, granted, if you are chasing people for more than 84 days and they still haven't got back to you, you probably got bigger issues like, are they even going to sign the deal? Are they even

[5:32:22] going to pay the invoice? Probably not. Anyway, we did all that with clawed code and we did it just a few minutes realistically. Um, so that's that. Some ways to make this better. We could absolutely do some sort of higher quality nudges. We could, you know, give people specific types of value like email sequences in newsletters that drop like case studies and assets. Um, we could we could give them insights. You know, we could go pre or postmeating. I just did proposal here because I thought it was probably the easiest to to verify and see. We could do some sort of customer story. We could also just rip

[5:32:56] and rinse and repeat a bunch of these gentle price objection check-ins that I've been doing. Um, but yeah, I mean like this is fairly straightforward, which officially marks us as having finished the last demo of our course. So hopefully you've learned a lot. Just to give you guys a brief uh revisiting, we started with creative generation. So I showed you guys how to generate ads, organics, uh, images, and videos all using AI. We started off with like very standardized um compositing workflow that used HTML to like generate something that looked really similar to Stripe and uh later I showed you guys how to do AI image generators like GPT

[5:33:32] image 2 and then video generators like you know Gemini Omni and then Cance and stuff like that. I also showed you top of funnel uh for personalized copy. So how to take like email newsletters and then personalize the hell out of them at significantly higher quality outreach to these people than just like hello welcome to my product. I actually then did that for a product that I'm currently running, Maker School, my AI automation community that teaches you how to sell this stuff. And uh as you

[5:33:56] guys saw, it was fairly straightforward. Then I showed you guys how to apply the same idea of automating text and templates through AI, which is something that I had a lot of experience with way back in the day when I was going door to door. Then uh for middle of funnel, I showed you guys how to incorporate speedtolebased email, SMS, and voice uh into a flow. Um very straightforward and easy to do as you guys saw earlier. All we really do is we just hook up um sort of an email polling system that checks our mailbox and then if you want to turn that into like a digitally run or like

[5:34:25] cloud-based routine um we do that via API call. And so we set that up on both email and SMS. Then for middle of funnel we did data collection and tracking and then visualizing all that in this cool sexy dashboard which you could host for your own business or you could sell to people. And then finally bottom of funnel automating high quality follow-ups via email. And I'm realizing now it actually I I said SMS. So, why don't I just round that out for SMS. Um, in order to do something like that, all you would have to do is basically add the Kuo connector, which I will jump on

[5:34:53] right over here, do some editing, and then um actually add that QUO connector. And then [snorts] over here, I'll also check for the phone number. So, I'm going to go here, hold this down. In addition to sending an email, I also want you to um draft and then send a message to their phone numbers using kuo

[5:35:13] quo. Awesome. You'll get the phone numbers from the local ClickUp scrape. Cool. I mean, that is really how easy it is to modify these sorts of flows to be clear. Um I will have to obviously test that if I did want to use that at scale. can't just assume that it's going to work perfectly. As you guys saw, a lot of bugs do crop up throughout the process. Okay, great. So, now that we've made it all the way to the end of the practical builds, where do we go from here? Well, there are two things that I want to talk about. The first is the maintenance and then the

[5:35:47] upgrading of these systems over time. And to be clear, when I talk upgrading of these systems, I mean like wholesale making this better um basically every time you run it. And there are a couple cool things that I'm going to touch on there in a second, but um for now uh I'm just going to give you guys probably like the 8020 and then you guys can

[5:36:03] apply these advanced tips if you want. And then speaking of advanced tips, the second is a bunch of advanced tips. Now the advanced tips are more about how to actually um run this sort of stuff in real organizations, less so academic. So uh yeah, I'm just going to give you guys more or less everything right now. And I've arranged my little worksheet here to essentially work like this. So, my first major uh point of advice on maintenance is, you know, if you're doing this for a client or your own business, it's going to be a little bit different, but I'm assuming right now that you're maintaining these systems,

[5:36:32] let's just say for somebody else. Um, what you're going to want to do is you're going to want to get all of your login and authentication upfront. And the reason why is because u login and authentication in practice actually end up being something like 70% of all of the maintenance requests and and problems that I've ever had. Um typically a pro you know a problem will

[5:36:53] go like this. Uh let me not use one. it'll be like service upgrades and you know when a service let's say some sort of Gmail or something upgrades you lose uh your connector after your connector breaks right then and I'm thinking about this in the context of a client client gets angry or client thinks system doesn't work that's not how you spell system so Then they call you in the middle of the night being like, "Hey, this thing isn't working." You build something that is inherently broken or wrong. And then you have to sort of do a fix last second. And then in doing your fix last

[5:37:35] second, you kind of have to get them on the call. You have to get some sort of communication with them. Um meaning you have to reauthenticate. So number one, it's a it's a bad look for you, but obviously it's also a bad look because this whole maintenance and upgrade loop could have been handled assuming you actually had the core credential um that we're talking about. What I mean by that is despite the fact that it is far more secure not to have credentials for all of your clients or all of the team members within your team's um login and so on and so forth, it is way easier to

[5:38:02] handle stuff like this on the fly if you actually have the core or root login. So much so that you can actually do what's called a debug flow. So, I don't know if you guys remember inside of Claude, just going back here, um, there was a, if we go to new and then go down to our local environment, okay, there is a section where you can actually store keys. And one thing that you could also do inside of this local environment is you could store usernames and passwords. So, what I mean by this is let's say, you know, we have some ClickUp fail that it just continues to

[5:38:37] occur. We could actually store clickup user equals, you know, nick at whatever.com. We could also store clickup_pw equals nick blah blah blah blah blah blah fan whatever you know 45 I don't know cuz nowadays they all require numbers and then what happens is you can actually store a line in your cloud routine or in you know your your your local environment because this isn't a local environment. So probably your loop in this case if you wanted to do in a cloud routine it would be the cloud environment. You could store a line that basically says um you know if o issue log in via whatever the creds are. Okay,

[5:39:23] navigate to you know location where all the stuff is and then ultimately regenerate API key. I don't know if you guys know, but these these models now have browsers, right? And so, as we saw with Claude, when it develops a website, it literally opens up a little browser on the right hand side. Well, it can legitimately use that to log into services for you. And so, what you can do is you can have like sort of a debug flow where if there is

[5:39:48] an issue, it can self-manage that issue. Um, but the only situation in which that is doable is if you already have all the login and authentication. And unfortunately, as mentioned, this is like 70% of all the problems with these sort of recurring loginins. And if you want something to be long-standing business infrastructure, a lot of the time you just sort of have to let go a little bit and say, "Okay, like take the login, log into the service yourself and

[5:40:07] then do everything you need to do." Okay, so that's the first thing. The second thing is um called self-healing. And I'll show you guys an example of self-healing in a second. But if other issue you know my recommendation is you give the model basically total cart blanch autonomy to fix and then rerun. This is called selfhealing and it's a really important part of like robust um skill loop or routine development. To make it much simpler for you guys to visualize, you can think of it as giving your AI systems the ability

[5:40:54] to rewrite their own instructions. If a step keeps failing in the skill, update that skills instructions and try again. So, for instance, you know, let's say I'm doing my usual, uh, creative generation flow or something like that. Um, and for whatever reason, it just can't access, [clears throat] you know, one of the API routes that it used to, uh, access. So, it calls this route over and over and over and over and over again, and then, you know, it says it's just tells me like, "Hey, Nick, so I'm sorry, this can't work." Well, rather than force the model to go through a bunch of the same attempts over and over

[5:41:28] and over again, if you just add a line somewhere at the very bottom of your skill or somewhere at the bottom of your looper routine that just says if you've tried doing this more than three times and it does not work, then odds are there's an issue with the actual core service, figure out what the issue is, see if maybe a route has changed or something like that, and then actually go through and update and add a log of everything that you've done in order to do it so future runs can work successfully. If you do that, um, the model will virtually never cause any problems for you again. So long as it

[5:41:55] has all the resources it needs to actually fix the problem, like maybe some login or something like that, it will heal itself. I've called this multiple things over the uh last year or so called a self annealing. That's a pretty cool term, but I think self-healing probably makes more sense. It's like kind of how Wolverine's body regenerates when somebody slashes him, right? Like that's that's sort of the idea here. Uh, and so typically what'll happen is you should have a phrase, and I'm not going to give you the exact phrase here because I don't think my phrase is perfect, and there are a variety of different things you could

[5:42:22] say, but basically if at any point you run into the same error more than three times, odds are something has materially changed in the systems that you are interacting with. Evaluate these changes, explore, and see if you can solve the problem. You have full agency autonomy. After you've solved the problem, report back and update your own skill or loop or routine, whatever it is that you're running, with a change log of upgrades, including the problem, your solution,

[5:43:01] and then update the actual skill itself. Append to this change log as you continue doing this. So we always have a breakdown of what happened and why. Okay. So this is just something you know this is one of those things you can just copy and paste and stick into literally every single uh window virtually ever and you will have a significantly better and higher performing system as a result. U so this is in practice how you sort of maintain the systems. You get ahead of the brakes by letting the model do more and more and more heavy lifting on its own. And I'm not going to say

[5:43:35] that this is like 100% foolproof. I've had some situations where I give them a little too much self-healing access and because of some spurious issue, maybe like a weird rate limit, it fundamentally changes itself and then it screws up. But those issues are far less common than um issues where, you know, the model just screws up due to some temporary rate limit and then ends up just calling it quits. The whole point of automation really is that it's hands-off, right? It's a system that

[5:43:59] just functions without humans. Obviously, we want to eliminate human labor completely where possible. We don't necessarily just want to minimize it. as such that we can redirect that human labor towards areas of the business that are more profitable and then ultimately probably better and more fulfilling for a human being. Like think about, I don't know, weaving a basket by hand 5,000 years ago. It's like, well, why would you weave a basket by hand uh and you know, try and make a livelihood off of if you could just build a basket weaving machine, you go up one level of abstraction, now you monitor the basket weaving machine, fix it, make

[5:44:28] improvements, and stuff like that. You're you're producing a net good for society, right? That's sort of how automation is. Imagine if that basket weaving machine could now clean itself and like avail itself of little bugs here or there if it had its own hands. That's basically what we're doing. So that's one of the changes. The second major thing I want to talk about on the maintenance end is error logging. You know because we're moving away now in automating you know marketing stuff from the old school platforms like you know make.com nadn these sort of drag and drop tools. Um, we're currently in like an erroring apocalypse where these

[5:45:01] things are failing constantly and we're just not seeing them because we don't have simple like notification routes available to let you know when something is or is not working. So, what do I mean by this? If you guys were using an old school tool, it's funny that this is old school cuz it's kind of where I started on YouTube talking about this platform which I was using to to automate my own marketing. But if you guys are using an old school platform like you know make.com or something like that basically what you get with a tool like

[5:45:24] this is you get automated error logging. And so every time there's an error inside of like one of these flows that's drag and drop based like this. You'll actually have like a full history log. You'll be able to see every change that has ever been made. And then you'll also be able to see all the errors, incomplete executions um physically literally on the diagram. You'll be able to see the errors and uh you'll also receive notifications. And you can set

[5:45:46] those notifications to send you emails. you can set those notifications to, I don't know, do push noteps or whatever. Now, same thing with like N8 even if there are issues on this more kind of modern uh builder, you know, they they'd be able to see these things. You'd be able to see tickets, you'd be able to have error logging flows, you'd be able to get emails. What I'm trying to say is like you would know when something screwed up. And the whole idea is because you know that something screwed up, you'd be able to fix it significantly faster, right? It's much more like observable and transparent um

[5:46:12] than than most things that we're doing now. Unfortunately, because AI has just moved so quickly and because we're all building slightly different things nowadays, we don't have that same standardized observability, we don't actually have the ability to like log errors. And so, what I like to do as a minimum is I basically like to give all of my actual routines, the ones that are running my business, the ability to log errors to a uh de facto error channel inside of my workspace. And so, what do I mean by this? What does this actually look like in practice? Well, um I'll show you guys. I'm actually going to

[5:46:43] wire up one of our flows to to do this. Um, but first of all, let's just call this error handling. And the very first kind of most important thing about error handling is you just have to make it visible. So, you know, you dump this stuff where you work. If your business works in WhatsApp, then dump it in WhatsApp. If your business works on Slack, dump it in Slack. If your business works on Discord, literally just dump it in Discord. Uh but basically every time that an error occurs, I recommend finding some place specifically for

[5:47:14] errors, like an errors channel, okay? And then wiring it up via um some logic, which I'll give you guys in a second, that just says like error with, you know, XYZ. And then what you can do is you actually weave in this self-healing, right? Air with XYZ tried XYZ to fix, you know, succeeded or failed, right? If you know it's a failure, something like this is totally fine and more than enough. Now every person in your organization will have basically full um top to bottom access and visibility into processes that you guys are doing. Okay. So how do you actually do this in something like you know um

[5:47:52] claude? It's actually fairly straightforward. What you do is you go down here and then right over here you see how you can add a connector. What you do is you just add slack and then at the very bottom you say something like if you have any errors send the error notification to the you know # error

[5:48:10] slack. Let's do this um slack chat. Okay. And so once we have the connector it'll be very very easy to do. Let me show you guys a quick example. Um just go back here and then down to the bottom I'll go settings. Then I'll scroll all the way down to connectors. And what do we want? We want Slack, right? So you can see it's one of the most popular ones to connect to because we can then send error messages. I'm not going to connect to it. Okay, now we're back here. Okay, so what I've done is inside of my Slack, I actually have a errors channel. Okay, and I just set this up

[5:48:43] here as a demo. Um, my actual errors channel is private, but um we're just going to use this public one for the sake of uh discussion. And I'm just going to go errors and we'll go public. The reason why we're doing that is because I just want to make sure I have the public version of that and I don't forget it. And then here, what I want to do is I basically want to say, "Hey, I want you to send a sample error notification to and uh I'm going to do

[5:49:06] errors public in Slack." Okay. And then this will now have the ability to basically dump a brief little error message and log as it works. So, it's gone through the errors. Okay. I'm just going to say send it to show you guys what this looks like and then we can actually wire it into a real flow. So boom. Now we have error sample notification service invoice chase worker environment production time this cannot read properties of undefined count one occurrence. This is a test

[5:49:31] message to verify errors are alerting. Okay. So now that we've verified that this does work with the connector. All we have to do is I'm just going to say give me all the info to ensure you don't have to search this channel up again. I will add this to a routines prompt verbatim. Now we can actually just copy this whole message. Go to a routine. And

[5:49:53] why don't we do this first one here? Maker school daily ads batch. I'll just go up here and then um add my Slack notification. Then I'm going to paste this. And then I'll say for the sake of this demo, I want you to pretend there was an error in the flow. then report to Slack. Okay, I'm going to save this and then I'm just going to run this now. I don't know if you guys know, you guys can actually run these live like manually. So, that's what I'm doing here. Um, once you're done running it manually, you can actually go to the bottom here, go today

[5:50:29] at 10:30 a.m. And you'll see the very first thing it's going to do the second that it runs is it's going to clone the repository cuz that was part of our um our action. And then it's going to initialize the session. Uh you can also add a a setup script here if you want to that you know includes all the information in addition to the environment variables. So you can actually have like a script that runs every single time one of these things lo

[5:50:48] loads or launches which is kind of cool. We don't have to do that though. Um the thing that's worth noting is this will work slower than our uh physical cloud code session like the one that's local on our computer. And now it's saying I'm generating the ad batch uploading to drive first. Let me look at the structure. Okay, cool. So that's looking pretty good. It's going to install all the things it needs. Okay. And you can see we've just triggered an error here from our flow. Drive upload blocked for maker school ad batch service maker school ad batch render environment scheduled task. And then we don't have a

[5:51:17] time. So now we're saying the bulk upload of the PGs plus contact sheet HTML stalled. The drive connector only accepts blah blah blah. So now now what I'm trying to say is we have uh we actually have an error flow. And what's cool is because we have this now we have observability and we have logability. Um you can stop a lot of problems before they actually become problems. I mean, you know, if we're running a routine at 5:59 every morning and it just screws up once and the average person gets into the office and we'll need the outputs of that routine at, I don't know, 8 a.m.,

[5:51:44] you know, you have like a solid 2 hours and 1 minute to actually solve that problem. Um, so if you didn't have the logging, you wouldn't have known until somebody does come into the office. then you would have had to bug fix and actually fix and uh eventually implement some sort of production update and then maybe it still would have taken an extra hour and before you know it now you've lagged that core person's um key deliverable the thing that they require in order to move forward their job like five or six hours. So yeah, that's that's my debug flow and um that's how I personally deal with these sorts of

[5:52:11] systems. Um I basically treat them as highly autonomous agents, hence the name that have the ability to rewrite their own code if needed. Now, on the advanced tips side of things, just want to share with you guys some learnings that I've had over the course of the last maybe twoish years of using this sort of AI agent workflows in real production systems for clients ranging from very small to mid-size businesses, maybe a few tens of thousands of dollars a month in revenue, all the way up to companies that legitimately make billions of dollars per year. The first is that effectiveness beats efficiency. So I see a lot of people on the internet trying

[5:52:45] everything in their life to be as efficient as possible. You know, they'll shave off tiny percentage points here there in a process um by I don't know like readjusting how they work. They will try and save tiny little 100 millisecond latencies in some system because they think that you know the more efficient this thing can work the better. They will reduce their margins as much as humanly possible for some core process because they're like well I'm saving more money. I'm spending less money in order to do it. And you know, at the end of the day, they look at their KPI metrics, whatever their logs are for marketing activities, and

[5:53:17] they're like, "Wa, this is way more efficient now than it ever used to be." But then they check out their revenue and they're like, "Wait a second, why is my revenue so low?" The reason why is because what you guys are doing is you are automating processes that probably aren't all that beneficial to your business in the first place. Efficiency

[5:53:33] is just doing things fast and cheap. Effectiveness on the other hand is what you should really be going for and that's about just doing the right things allocating your resources in places that actually you know move the needle of your organization. So I just want to give you guys a brief example. Um you know we do a lot of automation projects for a variety of small to mid mid-sized all the way up to enterprise businesses and I had this one private equity firm that had reached out a while back and for those of you guys that don't know you know in private equity a lot of the

[5:53:59] time what you're looking to do is you're looking to book meetings with qualified business owners that are either interested in investment or want to sell. And so this was a private equity company that works with um kind of bluecollar businesses and that just means like kind of home services, trades businesses, that sort of thing. Think electricians, think junk removal

[5:54:15] businesses, movers, that sort of deal. And so they came to me being like, "Hey, Nick, I want to automate stuff." And I'm like, "All right, sweet, man. What kind of work do you want to automate?" and they said,"Well, right now we're spending about an hour every week on the phones uh with prospective business owners trying to do due diligence on their business, meet them, and then build relationships with them. We want to automate that with a voice agent because we think it would save us a lot of money." And I remember thinking in my head about this and I was like, "Brother, you're going to be automating

[5:54:50] a single hour of your week. It's like uh I don't know that old trade offer meme. I don't know if you guys have like seen the trade alert meme, but basically it's like new trade offer detected. You know, I receive one hour per week of time back and then you receive literal like devastation

[5:55:15] to your business. What I mean by this is if you're automating literally the core thing that makes you money like this private equity firm is, you better ensure you're going to do a goddamn great job. And the harsh reality is if the only way you make money is by jumping on calls with prospective bluecollar businesses to determine their eligibility and to feel them out. If that is the only way that you actually make money in your business, it makes no sense to automate that for a small time savings gain of one hour a week. It's like if you can't spend one hour a week doing the most

[5:55:50] important thing in your business, the only thing that makes you money, you know, uh because you want to save a few bucks, then do you even have a business? No, I don't think so. So, what's really interesting is just in my own business, for instance, I do a lot of stuff manually. And people are always like, "Nick, why are you doing so many things manually?" And it's not that I'm doing everything manually. It's just the only work that I am actually doing is work that requires my full attention. It is

[5:56:12] recording videos like this for instance. It is going on sales calls with mid-market and enterprise businesses where my presence makes a material change to the probability of that sales call working out for us. Uh you know it's like applying my human meat brain, my stuffy sort of fluffy mind to real problems within my business. Uh and then thinking about those not thinking about how to do my bookkeeping or how to scrape the leads. sort of these wrote monotonous tasks but actually applying myself to things that are ultimately effective for my business. So if you're a business that is looking to integrate AI and automation, try to avoid

[5:56:48] efficiency. And I mean that seriously. Don't look for cost. Don't look for speed. Look for um hey, what can I automate that is currently pulling away from my time uh from being more effective. what can I automate away so that I have more time to actually spend on things that drive the the the top line of my business and

[5:57:09] ultimately the bottom line of my life. And on that vein, here are a bunch of things that usually fall under those camps. So things that you should automate, things like data entry, you know, think about it. We just automated a fair amount of my scraping flow, right? We built um a flow that automated the process of enriching personalized ice breakers with things that we have in common and stuff like that. That's pretty routine data entry. I'm not exactly winning any awards there if I do it or if I pass that off to some team member. And it takes a fair amount of my time. So yeah, I'm happy to automate

[5:57:37] processes like that. Why? Because it frees up maybe 40 to 50 minutes of my day and then I can spend that time actually focused on, you know, things like um sales conversations and client relationships. You can automate things like your first draft copy. You know, we automated a lot of ad creative here. I showed you guys how you guys could do this at scale and automate tens if not thousands of images, videos, and pieces of copy uh across the board. That's

[5:58:00] totally okay to have AI help you with. What you don't want to automate is you don't want to actually automate the production of the final asset itself. What you want to do is you want to use AI as an ideation and um sort of concept spec tool. Okay? And then you take those concepts and then ultimately you pick the ones that make the most sense. You know, upgrade them a little bit and then as mentioned push them live. So I would not automate client relationships, but I would automate first draft copy. How about reporting? As you guys have seen, you know, a lot of that reporting stuff

[5:58:26] that you guys typically previously had to do was pretty manual. was like logging into 5 million different platforms and then taking the data and like copying and pasting stuff from one sheet over to the other over and over and over again. Well, you you shouldn't do that. What you should do is you should, you know, have AI or uh some sort of procedural automated system manage the automation of that. Okay? And that's fine. Um so you can automate reporting, but what you should not do is you should not automate, you know, bad news from those reports. A terrible thing to do would be use that signals

[5:58:54] feature that we developed in the dashboard and then have that automatically sent to the client. Hey, here are your signals for the month. It's like no, no, no. That's probably one of the core things that you as a human being make or lose money on. You should be in charge of that and you should actually go sit down and meet with the client. And then ultimately scheduling. You can absolutely automate scheduling. Although I don't actually automate that much of my scheduling. Um, but what you should not do is automate anything where a client, a prospect, somebody that pays you money must feel valued. And a lot of the time that's not

[5:59:20] just like here is my calendar. Go ahead and book, you know, like a good appointment setting flow actually asks, "Hey, yeah, I think we have time on Tuesday or Wednesday. Uh, does 2 p.m. work for you?" Right? Cuz at least that that that makes you feel human. It's not disrespectful. It's not a robot answering uh, you know, and sort of weighing the costs of an interaction against what you stand to gain. The whole idea really is that automation should buy you more time with people. It should not replace the time that you have with people. This is sort of I mean this is an AI generated graphic and I

[5:59:50] think this is a good example. These AI generated graphics that I've been creating throughout this entire course. You know there some of them are kind of funky and cute here. There they save me a tremendous amount of time cuz I don't actually have to go through and make the thing. Do you know what that means? That means that I can actually spend my time talking to you and delivering value in that way which is probably the core way I deliver value. Contrast that with this fell over here who gets a text message from some happy smiley robot which you

[6:00:10] know took two cents in tokens to send. We value you. It's like no you don't really value me. And uh that's inherent in the medium and the way that you send me that message. You know, if you really did value me, you as a salesperson come down to the coffee shop and actually sit down and chat with me. A simple filter that you guys could use is just if you have a process you want to automate, ask yourself, does the customer actually feel this process? Is this like a firstline thing where we sit down and we like we we talk? Is is a relationship um at risk here? If the answer is no,

[6:00:39] you're usually good to automate it. Likewise, if it involves a real relationship like between you and your customer, then you should automate it, but you can QA it by hand. You know, you're happy I'm happy to have you automate processes up until some final filter like with ads and stuff like that, but there should still be a human element. You should have your taste applied to it. You should have your ability to actually uh uh uh, you know, be there for your clients if needed. You know, is a slight little like, "Hey,

[6:01:04] Pete, can you check in on the invoice?" Is that like a relationship based thing? Does the customer feel it? Probably not. That's more logistical. but a hey Pete I'd love to schedule a call with you to talk about X Y and Z I think I can really help you and really make a big deal you know big bottom line impact to your business that over there is something that's worth keeping human okay hopefully everything that I'm saying here makes sense um obviously I'm a big fan of automation that's kind of my whole business model right to be clear but as somebody that has done a lot of automation now I have realized

[6:01:30] that people will overmate tremendously and I've just seen a lot of downsides of that recently thank you guys so much for making it to the end of the course I really do hope that this was valuable and helpful. Uh, you know, I've worked with a lot of companies up to this point in my career. So, this course is basically me just trying to compress everything that I know into just a few hours for you. I I really hope I did a good job. Uh, and I'd love to hear your feedback. So, if you guys did enjoy, please do two things for me. The first is please subscribe to the channel. I

[6:01:57] put a lot of effort into these and for whatever reason, most of the people that watch aren't actually subscribed. I will, of course, continue doing this regardless of how many people are subscribed. So, don't think that I won't, but it obviously helps if I have support. And second, if you guys are considering using these tools to help other people, like other businesses or organizations, uh, please check out Maker School. It's my 90-day AI business program that guarantees you one customer or your money back. It includes a very comprehensive curriculum literally day by day about how to sell this stuff. And I will show up, you know, 7 days a week,

[6:02:26] 365, every single day, and respond to almost every post in the community. I do that literally every day. It's the very first thing I do when I wake up in the morning. I consider Maker School easily the most impactful thing I've ever done. So, if you guys want to be a part of that, then then check out the link in the description. Anywh who, thank you

[6:02:40] guys very much for sticking around. Hopefully, I'll catch you all in another course. I'm going to do a lot more of these. Thank you again and good luck. Get out there and crush it.

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt