TL;DR: By the end of this, you’ll have your own Cowork operating system: a set of folders where Claude keeps your rules, your context, and every decision you make, so you never brief it from scratch again. You’ll learn how the whole thing works, then two prompts build it with you in under an hour — and your existing Claude Projects move in with their memory intact. It lives in Google Drive, so it follows you on every computer and can be shared with your team. A Session Sync skill keeps it learning after every session. No code anywhere.
Everyone wants the same thing from AI now. Not a chatbot you re-explain yourself to every morning. Something that already knows you.
Your work, your projects, your preferences, how you like things done, what you’d never want it to do. Something with a real memory, that remembers your decisions and the context you gave it last week instead of forgetting the moment you close the tab. An AI that opens already caught up, so you skip the briefing and start working.
A personal AI operating system, basically.
Hardly anyone has one yet, for two reasons.
The first is context. Most of what the AI needs to know lives in your head and never gets written down: how you structure a client email versus a team one, the words you avoid, the rule you’d never break. So every session starts close to zero, and the more work you push into AI, the bigger that tax gets.
The second is setup. It looks complicated. What goes where, how to give it a memory, how to lock in decisions so they stick, how to hold every area of your work without it turning into a mess. None of it is hard, but it looks like enough work that most people never start.
Today we fix both with a Cowork operating system that holds your context, your decisions, and every area of your work, so you never brief Claude from scratch again.
Update: I later extended this same system to work across Claude, ChatGPT, and other AI assistants. And if you want to see how it fits into everything else I do, I also shared the full AI system I use to run this newsletter.
In my last piece I gave away a free Claude skill that saves a workflow so you can reuse it instead of rebuilding it every time. A Cowork operating system is the home most of those pieces live in.
That matters more the deeper you’re in. I do a lot of my work in Claude Cowork now, and at that volume the missing structure is what costs you. So I set mine up properly, with an initial nudge from Jeff Su’s setup, and the surprise was how little I had to write myself, because most of it was already sitting in skills I’d made months ago. You’ve probably made the pieces too. What’s missing is how they fit together.
So I took my whole setup and turned it into two workflows that build your Cowork operating system around your actual work. Every file you need, ready to run from the first session. In under 60 minutes, start to finish.
And as of July 2026, Cowork now runs in the cloud, which means your scheduled tasks can keep running while you sleep, even if they work with your local files. Read Part 2 to learn how to set it up.
Here’s what we’ll cover
What a Claude Cowork operating system looks like (my real setup)
How a Cowork operating system works: root, workspaces, projects
The two ideas that make it run: rules vs. facts, and routing
This is the system I (really) wish I’d had months ago. Let me show you what it does.
What a Claude Cowork operating system looks like (my real setup)
Before any explanation, here’s the thing itself:
Just a couple of folders, right? Except these few folders now hold close to my entire working life, organized into one structure Claude reads automatically.
When I start a session, I don’t explain anything. No setup, no paste, no re-briefing. I just ask, and it already has the context, the skills, and the rules it needs.
Here’s a demo with just a few tasks, though it works just as well for anything in my workspace, from client work to end-of-month finance to marketing Amplifiers.
In this demo:
I asked for a carousel from “my last article” and never gave it a link or a title, because it knows where to get it from on its own. It found my latest post, pulled the full text, and built the carousel copy.
I switched to a LinkedIn infographic and post mid-session. It reached for the right skills, matched my brand and voice, and routed the image through the Amplifiers connector, all without being told how.
I asked for internal linking opportunities for SEO. It scanned all 104 of my published articles and handed me the best ones to add, with the exact anchor text and why each fits, in the format I always want. I never told it where to find my articles, what I meant, or how to format the answer.
And changing how it works is just as easy. I say what I want once, and it locks the new rule into the right file, so next time it knows, without me repeating it:
None of that is clever prompting. It’s structure. The system knows where everything lives, so it never has to ask, and I never have to repeat myself.
And since I first wrote this, I've gotten that same structure to work with ChatGPT and Codex too — same folder, same rules, same memory.
Now here’s how the whole thing works, in a way that’ll make immediate sense.
If you run an agency or manage multiple brands, I've built on this same folder structure to create a full AI client-management system — with one isolated project per client, a shared-process workspace, and an orchestrator that scales your skills across every account. 💎 Read: AI for Agencies
How a Cowork operating system works: root, workspaces, projects
Think of your Cowork operating system as a company. Three levels, each one knowing a little more than the level above it.
The company: your root
Every company has things that are true for everyone in it. The mission, the values, how people behave, where to go for what. Set once, and everyone operates under them.
That’s your root, the top-level folder, and it runs on three files:
CLAUDE.md, the org chart and rulebook in one. It holds the rules that apply to everything you do, like your preferences for how Claude should act and talk to you. It also tells Claude where to go for each kind of work, and gets read first at the start of every session.
MEMORY.md, the current state of your world. What you’re working on right now, your recurring tasks, the facts you reference constantly.
ARCHIVE.md, the old records. Finished projects, past decisions, history you don’t need every day.
The difference between those last two matters. MEMORY is the desk you work at, Claude reads it every session, so it stays lean. ARCHIVE is the filing cabinet, Claude only opens it when you ask about something old (”what did we land on for that sponsorship deal back in April?”). So you keep your full history without it slowing down or costing you anything on a normal message.
The departments: your workspaces
A company has departments, each with its own goal, scope, and way of operating, on top of the company-wide rules.
Those are your workspaces, one for each ongoing area of your work. For me that’s the newsletter, content and social, client consulting, Amplifiers, the finances. Each one has its own rulebook and memory. When you work in that area, the rules stack:
Claude reads the company-wide root first.
Then it layers the department’s rules on top.
So when I work with Cowork on an article, it follows my preferences for how Claude should behave with me from the root, plus the newsletter workspace’s own conventions for the writing itself. Each layer adds detail. No layer repeats the one above it.
The projects: your specific work
Inside a department sit projects. Specific pieces of work with a start and an end.
Those are your projects, folders nested inside a workspace, the tightest scope of all. An Amplifiers launch inside the Amplifiers workspace. One client engagement inside consulting. They're temporary, and when one's done you just tell Claude to archive it.
Why it stays fast, and remembers
That’s the whole structure: company, departments, projects → root, workspaces, projects.
And like the map above shows, every level carries its own CLAUDE.md and MEMORY.md. That’s what makes it both smart and fast:
It remembers. Each MEMORY file holds the decisions, context, and current state for its level, written down instead of vanishing when you close the tab. The decisions I make inside an Amplifiers launch live in that project’s memory. Facts that apply everywhere, like my business model, live in the root.
It stays fast. The root never has to hold everything. Ask “what’s happening with the Amplifiers launch”, and Claude checks the root, confirms the project exists, then jumps straight to that project’s memory for the detail. The structure does the navigating.
And you never write any of these files yourself. They’re just plain text, the markdown format, that looks a little like code but reads like normal writing. You tell Cowork what you do and how you work, and it writes and maintains every file for you.
The two ideas that make it run: rules vs. facts, and routing
The two prompts I built to set up your own Cowork operating system already know how this works, so they handle it for you. But understanding the ideas is the only thing that lets you keep growing the system once the foundational setup is in place.
Idea one: rules versus facts.
Every piece of context either tells Claude how to behave, or it’s a fact that could change.
“Always” or “never” (”always sign off with my name”, “never use em dashes”) is a rule. Rules go in CLAUDE.md.
Anything that could change later (a project’s status, a contact, a tool, a decision) is current context. That goes in MEMORY.md.
Keeping the two separate is what keeps the system reliable.
Idea two: routing is just navigation.
This is what makes it feel like the system knows you. Routing is plain instructions in your CLAUDE.md files that tell Claude where to go and what to use, and it works at every level.
So when I say “polish my draft,” the root CLAUDE.md reads it, sees it’s an article, and routes to the newsletter workspace. The workspace’s CLAUDE.md doesn’t spell out every step, it just points to the polishing workflow saved in its resources, which holds the sequence:
Apply my voice DNA to rewrite it.
Use the article-infographic skill to find infographic opportunities.
Use the dense-infographic skill to generate them, then the image-SEO skill to give me a filename and alt text for each.
Use the SEO-audit skill to suggest a title, meta description, and URL slug, and tighten my section headings.
Use the thumbnail-generator skill to suggest thumbnail angles, then generate the one I pick.
I never spell those steps out. They live in the workflow file, and the routing just calls it, so it runs the same way every time and my CLAUDE.md doesn’t get bloated no matter how many of these I build.
That’s the whole concept: a company structure for your context, rules and facts kept apart, and routing that points Claude to the right place.
And if your existing Claude skills have started behaving differently on Fable 5, Sonnet 5, or Opus 5, I wrote a full guide to updating your skills library for the Claude 5 generation.
The prompts build all of it for you, which is what the next part walks through.
How to build your Cowork operating system in 5 steps
You’ve now got the full picture of how it works. Next, you build your own Cowork operating system, in under 60 minutes today.
You could also build all of this manually. Map your work, create every folder, write every file, get the rules-versus-facts split right, wire the routing. It works, but it’s a lot.
So I built two prompts that do it for you:
The Cowork Workspace Planner, which brainstorms your whole setup with you, your areas of work, what’s ongoing, what you already have, then hands you a clean blueprint to build from.
The Cowork Workspace Builder, which takes that blueprint and constructs the entire setup, folders, files, and connectors, in one pass.
They’re both inside Amplifiers, but I’m also giving you the full prompts, so you can paste them straight into Cowork if you prefer it that way.
From here, we build
Here’s what’s behind the paywall, in five steps:
Step 1, plan it. The Cowork Workspace Planner brainstorms your whole setup with you and hands you a blueprint.
Step 2, set up where it lives. How to run Claude Cowork across multiple computers with Google Drive, so your system stays in sync on every device, backed up in full, and shareable with a team from one folder.
Step 3, build it. The Cowork Workspace Builder turns your blueprint into the whole structure in one pass, and migrates your existing Claude Projects, so your workspaces start with their rules and memory already inside them instead of you rebuilding from scratch.
Step 4, make it yours. How to read through what got built and start growing it just by working in it.
Step 5, keep it sharp. A skill I’ll give you to run at the end of a session, it captures the decisions and preferences you’d otherwise lose and writes them into the right file automatically.
If you’ve been meaning to set this up properly, this is the part that does it.
Upgrade to premium.
P.S. We did this in Cowork because that’s where most of you hang out. But if you live in Claude Code instead, you can build the same thing there. Here’s the step-by-step for a Claude Code OS that runs in the cloud 24/7, so your automations keep running even when your laptop’s closed and you’re nowhere near it.
Let’s begin.