Learn 95% of ChatGPT Work in Under 20 Minutes

0:03You're in the right place if you use ChatGPT every day, know that ChatGPT work is a major upgrade, but you're still using it like a normal chatbot because no one, not even OpenAI, can explain what you're supposed to do differently. Now, ChatGPT goes from just answering questions to doing work for you by connecting to all of your tools.

0:23Okay, Dominic, but what does that mean? I guess I'll take a stab at this. Here in ChatGPT work, I've asked it to create a Google Calendar reminder for my brother's birthday. A simple enough task that regular ChatGPT can do as well. But, watch what happens next. In my next prompt, I tell ChatGPT work to add a cake emoji to the event title, put reminder to buy gift in the calendar description, and I also say to carry these rules forward for every birthday event. And after fast-forwarding, we can see the AI has saved both defaults for every future birthday event. But, what exactly does that mean? To answer that,

1:04let's open a brand new session within ChatGPT work and say, create a birthday reminder for Sam Altman on April 22nd. And let's fast-forward to see what happens. And we see the cake emoji in the title and the gift reminder in the description, even though I never asked for either one in this session, which illustrates ChatGPT work's biggest advantage over regular ChatGPT. ChatGPT work remembers the rules you want it to follow and carries them into future tasks. So, the more you teach it about how you work, the more useful it becomes. And that's just the tip of the iceberg. So, let's get started. Before diving in, let's first take a quick look

1:42at all the products you have access to. On one end of the spectrum is good old ChatGPT chat. I'm not going to waste your time with this today. On the other end, we have Codex, which is designed for coders and programmers. So, we're not touching this today, either. The middle is where it gets confusing because there are two versions of ChatGPT Work. ChatGPT Work Web lives inside the web version of ChatGPT. It has a few unique advantages, but it's a lot less capable than the desktop version. ChatGPT Work Desktop is the one I recommend using for everyone who's not technical. So, that's what we'll be

2:16focusing on for the rest of this video. Good job, OpenAI, for making this so simple. As an ex-Googler, I can tell you we never made our product lineup so confusing for our consumers. And something like this never, ever bit us in the ass. Once you've downloaded the desktop version of ChatGPT Work, you only need one other thing to unlock its

2:36full capabilities, a Google Doc. Think of this document as the master rulebook for ChatGPT Work, written entirely in plain English. The purpose section explains how this document works. The global rule section contains the instructions ChatGPT Work follows for all tasks. Voice principles controls how the AI writes, and maintenance explains how ChatGPT Work should keep the file organized as it grows. Taking a closer look at the Google Doc, we see that ChatGPT Work added the Google Calendar birthday rule from earlier as a subsection under global rules. The voice principal section is thin on purpose because you'll learn how to teach her to sound like you later on, and the same

3:20goes for the routing map. I've linked the starter template for you down below, and when you download this, make sure you go to files, download, and download as a markdown file, okay? And you're going to need to rename this exactly, all right? You're not going to rename this to like agent, you're not going to rename this to agent one, you're going to rename this to agents .md, exactly like that, okay? Or ChatGPT Work is not going to recognize it. Next, create a folder anywhere on your desktop, call it whatever you want. For me, I'll just say

3:51Jeff Oracle OS. And then, we're going to move the agents.md file into that folder, all right? And then back in the ChatGPT desktop app, make sure the work toggle is selected. And then under choose project, we're going to pick that folder we just created, all right? And once your folder has been added or mounted, ChatGPT Work will now automatically read your agents.md file before it performs any task. And the best part is, you don't even need to manually maintain this file. You can just tell ChatGPT Work, "Remember this rule. Whenever I ask you to rename image files, use a short but descriptive name so I know what each image is without

4:34opening it." And let's just fast forward here. Clicking into the agents.md file opens it up in the preview pane. And we can see that ChatGPT Work added a subsection under global rules with our file name conventions. Now, I can drop a folder of photos into my Work OS folder, and in a new chat, I can say, "Rename the photos in the images folder." And after about a minute, we're just going to skip ahead here, we can see the images have been renamed according to our new rule. So, just to be clear, ChatGPT Work could already analyze and rename images, but without rules in place, it'll rename the photos

5:13differently every time, giving you inconsistent results. Which leads us to the most important takeaway from this video. If you use ChatGPT Work without mounting a folder, you're basically using regular ChatGPT. Once you mount a folder with an agents.md file inside, ChatGPT Work reads that file before every task and follows your preferences, so you get consistent results. One thing I'm choosing not to cover today is the three ChatGPT models, because that's a whole separate rabbit hole. But, luckily for us, today's sponsor, HubSpot, has a free guide that breaks it down. It's one of the clearest explanations of which model is good for what, and honestly, I've been making the mistake of

5:57defaulting to the most powerful model, Soul, for everything, wasting like a bunch of tokens. But, now I know better. For a recent presentation, I started off using Soul to build the narrative and to analyze a bunch of earnings call transcripts and financial statements. Then, I actually jumped to Luna to pull thousands of numbers from the financials into a spreadsheet, and it was surprisingly fast and accurate. Then, finally, I used Tara to actually build

6:23out the slide titles and speaker notes. Inside, you also get nine ready-to-run workflows, from just getting started all the way to connecting Gmail, Slack, and Drive for recurring automations. So, check out the free guide by clicking the link down below. Thank you, HubSpot, for sponsoring this video. Next up, let's see how we can supercharge plugins with our agents.md file. As a quick reminder, plugins connect ChatGPT work to tools you already use, like Notion, Gmail, and Slack. In Claude, this is called connectors, and in Gemini, this is

6:53called connected apps. And connecting to an app within ChatGPT work is very easy. You just go to plugins and then search for a tool like Google Calendar, and you simply follow the instructions. But, just giving access by itself isn't that useful. Teaching ChatGPT work how you like to use that app is another thing entirely, and that's where the agents.md file comes in. For example, with your folder mounted, start a new chat and say, "Remember this. Meetings default to 45 minutes, and format Google Calendar

7:26descriptions so they're easy to scan. Start with bottom line up front, followed by three short bullets with relevant context." And we're just going to skip ahead here. Opening up the agents.md file, we see ChatGPT work added the new rule under Google Calendar rules directly, which is a great example of us, the user, setting the rules and the AI doing the grunt work. Now, in a fresh ChatGPT session, I can upload my apartment lease and say, "Create a Google Calendar event for a call I have with my real estate agent tomorrow at 10:00 a.m. I've attached my lease for

8:00your reference." And let's fast-forward. And again, we see ChatGPT work has created the event following the instructions saved inside our agents.md file. 45-minute event, bottom line up front with three bullets, even though we didn't repeat those instructions in the prompt. But, there's a problem. Google Calendar is only one app, and as you connect more, Notion, Gmail, Google Calendar, Teams, Granola, Spotify, so on and so forth, adding all those rules for all those apps in one agents.md file will get very messy. So, here's a pro tip. We can tell ChatGPT work, "Create a tool-conventions.md file that will house all rules governing tools moving forward. Move every tool or

8:47plugin-specific rule from agents.md into it. Replace the moved content in agents.md with a link." And we're just going to send this off. While that's working, let me quickly explain why we just did that. ChatGPT work loads agents.md before every single task. So, a bloated file means worse outputs and wasted tokens. And since we're not using tools like Calendar in every task, moving those app-specific rules into a separate file keeps agents.md lean and organized. Opening up the newly created tool-conventions file, we see that both the image file and Google Calendar rules

9:25got moved over. [music] And over in our agents.md file, that entire section got replaced by a pointer to the newly created tool-conventions file. Now that you understand how plugins work together with agents.md, here are three useful plugin rules that you can add today. For Gmail, never include a sign-off in my Gmail replies because I already have a default signature set up. For Notion, always use the existing templates in my Notion databases instead of creating pages from scratch. For Slack, when summarizing Slack threads, always separate the output into decisions, action items, and open questions. As a quick recap, plugins let you connect ChatGPT work to your apps. And although by itself, this

10:08is nothing new, you can use agents.md to teach the AI how you use those apps. By the way, my free Cohort Toolkit includes all the templates you need to organize your rules, folders, and workflows in 1 week. Yes, it is built for Cloud Cohort, but the exact same logic and frameworks apply to ChatGPT work. Link down below. Moving on to my favorite feature, the side panel. So basically, the side panel in ChatGPT work has three completely different functions. First, it lets you preview and edit files, as you've seen already. Second, it's a fully functional browser, and we're going to get into the benefits of that in a bit. And third, I

10:46can use the side chat function to hide stuff from the main chat. And this is not to be confused with a side chick. Okay, that was really bad. I'm I'm kind of struggling today. Hang on. Hang on. Hang on. Starting with the preview pane, let's go through an example. If [snorts] we open up our tool conventions.md file in the side panel, we can highlight the entire Google Calendar section, click edit, and we can say, "Add a section for Gmail, Notion, and Slack using the same structure, H2 headers, with clearly labeled placeholder rules under each one." And let's fast forward. Just a few seconds later, the same file has matching

11:27sections for all four tools. And if you made a mistake, you can just click the undo button here. I'm just going to click redo. Simple enough, right? Pro tip, whenever you see a hyperlinked file or folder, you can right-click and choose show in finder or the Windows equivalent to quickly locate it within your workspace, letting you easily drag the file or folder into another location. Next up, the in-app browser can be opened by pressing command and control T, right? Just like opening up a

11:55new tab and typing in the URL. And you should know that we could technically ask ChatGPT to open up the in-app browser in a prompt, but obviously that's super slow. Why would we ever do this though? Let's go through two real use cases. Here, I'm on my partnerships page, and instead of taking screenshots or copy pasting the URL, I can click the annotate icon, select this specific section of the page, and say, "Hey, create a preview of this section

12:25in dark mode." And let's see what happens. Think of annotation mode like leaving a comment directly on a website. Instead of describing where something is on the page, I can point to the exact element and explain the change I want in plain English. First, we see that ChatGPT work gives us a mock-up of how the dark mode would look, right? Light mode, dark

12:44mode, which is already pretty cool. Better yet, because ChatGPT work is connected to my website hosted on Ghost, I can actually ask ChatGPT work to implement this change, which I did off camera. Meaning, if you go to my website right now, you will see there is now a dark toggle, and it looks pretty incredible. To be

13:06clear, those were two separate steps. Anyone can open a public webpage and ask for a mock-up, but actually pushing the change live requires ChatGPT work to be connected to that platform. A lot of you are going to be horrified by this next example, but technically, you can log into your accounts within the in-app browser. So, if you trust OpenAI and Sam Altman, I mean, Sam Altman,

13:32you can hand off a lot of tedious work. For example, I recently launched a course and I had ChatGPT work test every single link within the lesson descriptions to check for dead links and to confirm they all open up in a new tab, so the learning experience isn't interrupted, which would have been super annoying to do manually. Lastly, we can type {forward slash} side to open up a temporary side chat that can see everything in the main conversation without interrupting anything. So, as an example, here I have an active task where ChatGPT work is reviewing a dense PDF. And I have a question, but I don't

14:11want to interrupt the task. [music] And so, I open side chat and ask, "Based on the uploaded PDF, what does personal assessment mean in plain English versus salaried tax?" And fast forward a little bit here, we see it answers my question without interfering with the main task. A few things to know when using side chat, it already has the context from the main task, so you don't need to re-explain or re-upload anything. And second, nothing inside the side chat flows back into the main chat. And third, closing it deletes the side chat, so you need to copy anything you want to keep. As a quick recap of the side

14:47panel, here's the mental model to carry forward. Use the preview pane to quickly review and or edit a file ChatGPT work surfaced in chat. Use the in-app browser when the context you want to share with ChatGPT lives on a website. Sign in only if the convenience is worth the privacy trade-off to you. Use side chat when the question or tangent needs the same context as the main task, but shouldn't change the deliverable. Now, let's move on to skills in the context of ChatGPT work. As a quick reminder, a skill basically takes some sort of input like messy notes, processes that against a checklist of rules you create and gives

15:23you an output like a structured briefing. Diving right into an example, here I'm in the middle of a working session where ChatGPT Work gave me the first draft of my next newsletter. [music] And after making some edits, I've pasted the final draft I'm actually going to send out. What I'm about to do next will teach ChatGPT Work to sound like me for

15:43[music] every newsletter moving forward. In the same chat, I say, "Compare this final newsletter against your original draft earlier in this chat. Identify the repeatable differences in my writing style, then create an improved skill I can run whenever I give you a draft and my final version. Show me the proposed learnings before updating the voice principal section of agents.md." Just to be super clear, we're doing two things here. One, we're creating a skill called improve. And second, the input we're giving it contains both the initial V1 draft and my final draft. Then the skill compares the two to extract patterns and the output are the writing rules to

16:25carry forward. Here, we can see that ChatGPT Work identified seven repeatable learnings, ignore changes not related to our tone of voice, and proposed updates to the voice principal section, and created an improved skill that I can use in the future. [music] And after approving those updates, fast forward a little bit, we can then open the agents.md file to see all the changes. And that's the key takeaway when it comes to using skills in ChatGPT Work. You are going to make those edits and corrections anyway, so why not create skills that teach ChatGPT Work your preferences so you re-explain yourself less and less over time. All right, that covers the basics, so let me

17:08share the next logical step on your way to building your own AI system. Here in my Work OS folder, it looks a little empty, but what if I dragged another folder with my travel preferences in here? [music] And And then, in a new ChatGPT work session, I say, "Review the new travel folder I just added to this workspace. Add a travel entry to the routing map, so future travel-related

17:29requests are routed to this folder. Finish by showing me the exact entry you added and explain why." And we're just going to fast-forward here. I'm actually going to ignore the output on purpose, because it's not going to make any sense to you right now. Instead, I'm going to start a new chat and say, "I want to plan a trip to Iceland. How will this workspace handle that?" And after just a few seconds, we see that whoa, ChatGPT Work knows that this ask is related to like traveling and stuff, and therefore should be managed within the travel folder we just added. So, what's going on? How did it know? Well, if I

18:07open up the agents.md file, we see that a new entry has been added to the routing map section, telling ChatGPT Work that all travel and trip-related tasks should be routed to the travel folder, since the travel folder holds all context related to planning trips, like my Marriott membership, my seat preferences on flights, [music] and what happened on previous trips. In plain English, instead of maintaining one giant messy folder, the routing map lets us break our work into smaller subfolders, one for each recurring part of our job or life. [music] And because I know a lot of you will ask, I use both ChatGPT Work and Claude CoWork, since I can just mount

18:53the same folder to both, right? And you'll see in my actual like Jeff OS workspace, I have more than like 30 areas or like what I call workstations. And I have agents.md that ChatGPT recognizes and a Claude.md that Claude recognizes. And if you want something like this, but you don't want to build from scratch, my Co-work Academy starts you off with a whole structure already built. The folders, the routing map, and the rule files. It's taught in Cloud Co-work, yes, but as you just saw, the same folder works in both. Link down below. If you found this useful, check out this tutorial next on how to give

19:27your AI system perfect memory. See you on the next video in the meantime. Have a great one.