My Simple Claude Cowork System (for normal people)

0:01Hey Co-worker, I am hung over. What are we supposed to be doing today? Finalize outline for workshop, dinner, newsletter write, review expenses. How much did I spend last month? Process last month's credit card statement. My spending is about 1,000% mainly from Bumble Premium. Good job throwing money away while not landing a single That's unnecessarily harsh, though. Did

0:30I really spend that much? Oh. All right. Moving on. Finalize my next newsletter using my tone of voice and just link me to the Notion page. I'm I'm too hung over to do anything today. Okay, final copy under final copy section. All right. Let's open the link. Okay, this is my rough draft up here and final copy. There you probably miss

0:57tip one. All right, not bad. Not bad. Okay. Not bad. That was easy. I'm going back to bed. Thanks. Oh, no, no. Wait. Going back to bed. What the All right, I'm not actually hungover, but what Co-worker did is real and it's insane what it can do with the right setup. And in this video, I'll walk you through how to build that setup without burning through your rate limits. Let's get started. So, how did Co-worker actually do all of that? Let me show you what happened behind the scenes. Here you can see I named my parent folder I Claude OS because get it? It's supposed to be iCloud, but we're using Claude, so

1:34it's I Claude. I'm so funny. Anyways, whenever I start a new session, Co-worker will always first read this claude.md file, which is essentially a set of master rules for Co-worker. Then Co-worker will read the memory.md file to remember what we did last time. So, these these files together, claude.md and memory.md, run my entire CoWork workspace. And here's the single most important takeaway from this video. These are just simple text files, so don't be intimidated. If I click this open, you're like, "Oh my god, this looks like code." It's not. I can select all this text, copy, open up any notes app, paste, and I'm just going

2:16to scroll all the way back up, and all of a sudden, you realize it's not that scary. Header one, Jeff's CoWork Club. Uh this is the central hub for all projects, planning, and work sessions. And underneath, it's simple text. The crazy thing is, you don't even need to write any of this text yourself because CoWork writes all of this for you. And by the end of this video, you'll see just how simple this system is. For example, within my root claw.md file, there is an instruction here that says, "At the start of every session, read memory.md before responding." So, let's open up the memory.md file, and what do you

2:54know? There is a section right at the top here called active projects and work, followed by a list of all the active projects I'm working on. That's how CoWork knows what I'm working on right now. It's the same idea with my tone of voice. Within the claw.md file, there's a section called content and production with a rule saying, "Hey, read this just voice principles markdown file before outputting anything to Jeff." Right? And that's how my newsletter goes from rough talking points to a final draft sounding like me. In plain English, claw.md is the instruction manual that tells CoWork how to behave, and it points to other files

3:35whenever it needs more detail. That's it. That's the entire system. To really help you understand, here's a simple analogy. The United States is governed by a set of countrywide laws called the US Constitution that apply to every person and every state, theoretically. This is level zero, the rooter level, countrywide laws. Moving one level down, we have state laws. For example, California has its own laws that help tech billionaires get even richer, and New York, not to be outdone, has New York state laws helping the finance billionaires get even richer. I'm just kidding with the laws, obviously, not really. But, the point is every state has state-level laws that stack on top

4:11of the US Constitution. So, a business in California is bound by the countrywide Constitution plus California state law. Here's how this maps to our Co-work workspace. The root claw.md file is the Constitution that governs the entire workspace. And one level below, we have individual workstations like email HQ, newsletter HQ, and personal finances that each have workstation-specific instructions that stack on top of the root level instructions. Put simply, the rules in my newsletter HQ workstation only apply when I'm working on my newsletter. But, the root level rules apply no matter what I'm doing. We can extend this one level further to project-specific rules that apply to a single project. But,

4:55we'll keep things simple today and just focus on the top two levels. Now that you have the full picture, let's start building at level zero, the root level. First, create a folder in your documents and name it Co-work OS. And because I'm such an angel, I've prepared three starter templates for you. A claw.md file, a memory.md file, and a voice principles.md file with the exact structure I use myself, all linked below. Pro tip, open up each one of the .md files, go to file, download markdown to make sure you're downloading in the right format. After downloading the files, drop the claw.md and memory.md files into your root folder. Then,

5:34create a 00_resources folder, and within this folder, move the voice principles.md file over. We can now go back to Co-work and select the parent folder we just created right here, Co-work OS. And after we open it, we click always allow. And we can even go ahead and star this so that Co-work defaults to this folder every time we start a new session. Just so we're 100% clear, the claw.md file is the instruction manual that tells Co-work how to behave. The memory.md file is like a notepad that stores things Co-work should remember between sessions. And the root resources folder contains detailed knowledge that Co-work only loads when it needs to. Pro tip,

6:23install the free app Obsidian and under open folder as vault, we can click open and locate the Co-work OS folder we just created, click open, and now we can simply view the claw.md and memory.md files in a more readable format. Don't worry, you don't have to learn how to use Obsidian. I don't know how to use Obsidian. We're just using it to read the markdown files. Clicking into the root claw.md file, there are three things I want to highlight here. First, under memory system, right up top, there's a line that says, "At the start of every session, read memory.md before responding." And there's a rule after that that says,

7:00"When I say remember this, write the information to memory.md." And these two sentences are why Co-work has persistent memory. Second, the routing map section contains a table that tells Co-work which workstation folder to load for which type of task. We haven't built any workstations yet, but when we do later on this video, this is where they'll be all listed. And third, references. These are pointers to files in our root 00 resources folder that Co-work only reads when it needs to. This is one of the ways to keep token usage low. By the way, I had to cut a lot of detail to keep this video within 20 minutes, but I go a lot deeper

7:39and share more templates in my free Co-work tool kit. So, I'll leave a link to that down below as well. Moving on to the memory.md file, we have two sections here: active projects and memory. Starting with the memory section, this is where co-work will automatically add entries when you tell it to remember something. For example, in a new session, I type, "Current events are

7:59distracting everyone from the e-list. Remember that." And co-work will add that to the root memory file as a new entry. First, you can see that under the memory section, co-work has added a new entry right here. And to test this out, we can even start a new session and ask, "What did I say about distractions recently?" And after a few seconds, which I will fast forward, co-work will draw from its memory and answer correctly. Back in our memory.md file, the active project section works the same way. Just tell co-work about projects you're actively working on, and it'll add both the project and the progress of that project

8:36right here. Moving on to the resources folder, we can now start building out our voice principles file. So, assuming you watched my first co-worker video and you connected your co-work to Gmail via connectors right here, what you want to do is go ahead and grab this prompt template that I've linked below. It's the same link. You want to paste this option one Gmail connected prompt template into a new session within co-work, press enter, and what this does is co-work is now going to extract writing patterns from your last 30 sent emails to build your voice profile. If you haven't connected Gmail, find five writing samples that you like. Could be

9:16yours or someone else's, and simply paste the prompt template under option two, no Gmail connected, and it does the same thing. After a few minutes, co-work will tell you that the voice principles file has been updated, and we can click into the voice principles markdown file to see all the patterns it has extracted. For example, I apparently talk in a warm, direct, and professional

9:39tone of voice without being stiff. Co-work will just add more entries as it learns more about your preferences. And just for reference, my actual voice principles file voice principles file is actually over 150 lines long. By the way, the reason I keep saying root cloud MD, root memory MD, and root resources folder is because each workstation below that root level gets their own cloud MD resources folder and memory MD. But before we get to workstations, everything we built today revolves around one important concept, relevant [music] context. And that's something I learned firsthand from today's sponsor, Coursera. Longtime viewers know I took Google's AI central specialization on

10:19Coursera and made an entire video about it. And the one learning that stuck with me all this time is how surfacing implied context is the key to getting better AI outputs. Implied context is the stuff you inherently know, but forget to tell the AI every time. Your writing style, your active projects, how you like your email structured. That's the whole point of cloud MD and memory MD. They remember it for you, so co-work always has the full picture. The AI essentials course covers everything from how AI actually works to practical

10:49frameworks you can apply immediately. And it's taught by actual Google instructors and not random influencers. Not that there's anything wrong with that. Click the link below to get 40% off Coursera Plus. And thank you Coursera for sponsoring this video. All right, so we've covered the root level and going one level down is workstations. And these are basically areas of your life, each with its own cloud.md file, memory.md file, and resources folder. Broadly speaking,

11:13there are two types of workstations. Universal workstations handle things you do across every area of your life. Email is the clearest example. You email your team, your accountant, your sponsors, your friends. So the rules for how you write email apply across multiple workstations. Dedicated workstations handle one specific area of your life. For example, within personal finances, I manage my spending, my savings, and my tax deadlines. Now, let's start by creating a universal workstation, email HQ. I first grab the prompt template from the same link below where you downloaded the starter templates, and go to the email HQ video prompt, and then you're going to copy this, head on back

11:50to your Claude co-work, and paste it in and click enter, and we're just going to let this run for a minute or two, and I'm going to fast forward this next part. All right, now that it's done, let's walk through what the prompt just did. First, the step one, it triggered a rule inside our root Claude MD. So, let's just open that up. Under our root Claude MD, we have a creating new workstations section that say, "Whenever I ask for a new workstation, create a subfolder with that name, and add a Claude MD, memory

12:23MD, and empty resources folder inside." Then, step two, the prompt's own instructions kick in. Similar to the previous example, it basically says to search my Gmail sent folder for the last 4 weeks of emails, then analyze and extract only email-specific patterns that aren't covered by my root voice principles. For example, like greeting and sign-offs, right? And once you have those patterns, uh add them to the editorial rule section of my email HQ Claude MD. Now, in both our co-work OS parent folder and within Obsidian, we can see that co-work has created a email HQ workstation along with a Claude MD and memory MD file, and a empty email HQ

13:05resources folder. And if we open up the email HQ Claude MD, we see that under editorial rules, co-work knows to first follow my voice principles in the global voice principles.md file, and knows that the rules below layer on top with email-specific conventions, like my default greeting, my default sign-off, etc. So, just to be clear, this is where the rule stacking comes [music] in. The next time Co-worker writes an email, it will first read my voice principles markdown document to see that I'm warm, direct, and professional, and I lead with transparency, right? And [snorts] then it will read my email HQ cloud MD file to understand my email specific

13:50preferences. And that's how my emails sound like me and follow my email conventions. Now, if I open my actual email HQ cloud MD, you'll see that I actually taught Co-worker my inbox zero workflow, meaning it knows about the 2-minute rule, what labels I use and why, when to archive, when to snooze emails, and it just works perfectly. If you want my complete system with pre-built workstation templates and a step-by-step walk-through, so you can skip the trial and error of building from scratch, I'm putting together a Co-worker Academy course, and you can

14:21sign up for the waitlist down below. Moving on to dedicated workstations, these workstations handle one specific area of your life. Unlike email, which touches everything, your personal finances live in their own world, right? Their own rules, their own context, their own memory. So, first, with the Co-worker OS folder mounted, I share the past 12 months of my credit card statements, and then I grab the personal finances prompt template from the same link below where you got all the other templates, and you paste it into Co-worker, and then you let it run. While that's working, before I get flooded with questions about data privacy, it's just personal preference,

14:58okay? I'm fine with sharing my credit card statements with Co-worker. If you're not, don't. It's literally that simple. All right, it's done. Let's walk through what happened. First, similar to the email HQ example, the prompt created a personal finances workstation folder, along with a personal finances cloud MD and memory.md, and an empty resources folder inside. Second, Co-worker read every transaction across my statements, broke down the spending by category, and built a master spending tracker spreadsheet inside my personal finances resources folder right here. And [snorts] we can click into preview. We can see that it created four tabs, transactions, yearly summary,

15:38monthly summary, and category taxonomy. And let's say Cohort made a few classification mistakes. For example, this one, Canva, should not be freelancer and contractors. It should be software and subscription, right? That's totally fine because after I correct Cohort once, telling it, "Hey, Canva's actually subscriptions and software," it will remember that for next time. And next month's expenses will be categorized correctly. Pro tip, Obsidian only indexes markdown files, so the Excel file is invisible

16:07in the sidebar here. Even though it obviously exists on disk, right? So, just remember that the Cohort OS folder is a single source of truth when it comes to saved documents. Now, after building almost 30 workstations, I can tell you it's best to start slow. Build two or three workstations out, get used to how the cloud MD and memory MD files all interact and work together, and only build new workstations when the need comes up. One thing to know before I move on, under my dedicated workstations, I also have project subfolders. For example, under my housing workstation, I have a project for refinancing my mortgage. Under

16:45travel, I have one project folder for every planned trip this year. Under speaking engagements, I have a project [snorts] for an upcoming corporate workshop. Each project gets its own cloud MD, memory MD, and resources folder. So, it's the same underlying logic here, all right? And if you want to get into more detail, sign up for the free Cohort toolkit. All right, now let me quickly walk through a few of my use cases so you know what's

17:09possible once everything is set up. Let's say I come across a copywriting framework I know I'll use at some point. I can simply take a screenshot, paste it into Cohort, and say, "I feel like this should be saved somewhere. Figure it out for me." And thanks to the routing map in my root claw.md, Cowerk knows that this belongs within my existing copywriting frameworks file. Another simple example, within my email HQ claw.md, there is a rule that says, "Before drafting any email, search Gmail for existing threads with the recipient on the same topic, right?" Since I hate it when people start new threads for no reason. On to a more intermediate

17:42example, I can tell Cowerk, "Hey, I just finished a meeting, so draft a follow-up email to all recipients." And I won't waste your time here, but basically, Cowerk knows to pull the latest event from my calendar, go to the meeting transcripts folder, read the transcript, and realizes it's related to an upcoming corporate training, load the speaking engagements workstation, and drafts a follow-up email to my client that adheres to the rules from my email HQ workstation. This is the power of Claude Cowerk. Last example, I can tell Cowerk, "Hey, I'm going to Boston from July 17th to 24th for a wedding, set up a project in Notion." And because Cowerk knows all

18:18my Notion conventions found under my command center claw.md, it creates the project page and fills out every property and every section in Notion exactly how I'd want it. I chose a travel project for simplicity, but every project Cowerk creates follows the exact same process. Fill out the properties, fill out the project page, create a Notion thread to track progress, all because I taught Cowerk my rules once. Moving on to two pro tips that will save you time and money. First, whenever I'm done working with a session in Cowerk, I type {forward slash}

18:51session audit. And this triggers Cowerk to scan the entire conversation and look for unsaved principles and preferences it needs to remember for next time. I've included a starter session audit skill in the link below, so all you have to do is download it, and then after that's downloaded, head on over to your Cowerk workspace, customize, skills, plus create a skill, upload a skill, and simply just select the starter session audit skill, click open, and after it gets mounted, assuming that you finish entire conversation, at the end of a session, you can just simply say either forward slash starter session audit, or say before we wrap, audit

19:35this session. Both works. Pro tip number two addresses a tiny, tiny problem you may or may not have noticed, Claude's usage limits. So, here are three ways to keep token usage low without sacrificing quality. First, keep your root Claude.md file within 300 lines. And this is because Co-work loads the root Claude.md by default in every session. You want to keep that file as lean as possible by pointing to other resource files only when the situation calls for it. Two, don't repeat the same rule in multiple files. For example, my root voice principles has rules like no corporate jargon, never bury the lead, right? So, I do not restate those rules

20:14inside my email HQ Claude.md, and instead email HQ only focuses on email specific preferences like thread-first drafting. And three, default to the Sonnet model because 80% of the time Sonnet is more than enough and it's a fifth of the cost of Opus. So, only switch to Opus when the task has three or more steps that all depend on each other. I've been making videos for 6+ years now and I've never feared models, so I don't say this lightly, but you need to start building a system like this today. It doesn't have to be on Claude because Google and OpenAI will definitely come up with Co-work equivalents. You don't even have to use

20:50my templates or take my course, but you do need to start building this personal context system as soon as possible because the rules, the memory, the patterns compound every single session, and you will always be ahead of someone who starts even a day after you. For my part, I'm going to make a lot more co-work tutorials, so I recommend you bookmark this co-work playlist. See you

21:13on the next video in the meantime. Have a great one.