Hey,

It's been a while! I got slammed running my software business last year, and then started up a new agency on top of that. Writing the newsletter just stopped fitting in.

But now I'm back, and want to start writing to you again, weekly.

Here's the first one.

Since the beginning of the year I've been running most of my work with OpenClaw, the AI agent system I built to handle my content, leads, ops, the whole thing. (I was also recently on the Greg Isenberg Podcast to talk about my set up).

BUT AI moves so fast and in the last 2 weeks I've actually rebuilt most of my OpenClaw setup directly inside Claude Code, and I've been calling it ClaudeOS. The honest reason I did this is OpenClaw was breaking every other day unfortunately and I was spending more time debugging the agent than using it (I'm hopeful it will get better over time though). Claude Code as the base has been way more stable for me. Also, I am able to use it with my Claude Max plan.

So what's ClaudeOS?

It's a Claude Code setup that learns about me and my work over time, and gets sharper the more I use it. Regular Claude or ChatGPT chats typically start blank every session and you need to feed it new context everytime. ClaudeOS remembers things about me and my business and what tools I use.

A few examples of what it does day to day:

• I tell it "pull my warm leads from Attio and draft follow-ups" and it does.

• I drop a video link and ask for shortform scripts in my voice. It pulls the transcript, checks my older scripts for tone, and hands me back drafts.

• It remembers what I'm working on across sessions, knows my offers, clients, and the way I write, and automates things for me.

The real time saver is when you stop re-prompting the same context every day. The system wakes up already knowing you, runs the boring stuff without supervision, and the longer you run it, the better it gets.

Today's email is the full walkthrough so you can build your own.

Prefer watching?

→I Built My Own AI Operating System in Claude Code (Full Tutorial)

Read time: 5 minutes

🧱 The 4 Layers

ClaudeOS is built in 4 layers:

1. Folder structure and instruction files

2. Tools (CLIs, MCPs, APIs)

3. Skills

4. Routines

Let's go through each.

📁 Layer 1: Folder structure and instruction files

Create one project folder for ClaudeOS. This is where everything lives.

Inside, four core sections:

• Instruction files: who the agent is and how it behaves

• Context files: what it knows about you and your business

• Memory files: short-term and long-term memory

• Secrets: your .env (API keys and tokens, always gitignored)

Instruction files are the three that define your agent's behavior:

• Claude.md: the main operating manual. References every other instruction file. The most important file in the whole system.

• Identity.md: gives the agent a name and personality. Mine is called Clawdia.

• SOUL.md: values, mission, principles. The deeper "why" the agent operates by.

Without these, the agent has no character and forgets what it's actually for.

Context files I use:

• User.md: facts about me (role, preferences, working style)

• Business/: folder with offers, ICP, pricing, processes

• TOOLS.md: references every tool the agent can call

• api-docs/: docs for any APIs it touches

Memory files:

• Memory.md: durable long-term memory index

• memory/: topic-specific memory files

• Dream: overnight routine that consolidates the day

The skeleton everything else hangs on.

🔌 Layer 2: Tools (CLIs, MCPs, APIs)

ClaudeOS is only as useful as what it can touch. Three ways to plug things in, in my order of preference.

CLIs are my default. They're just programs installed on your machine, the agent calls them like any command, and they don't take up any context window space. If a tool has a CLI, I use that:

• gws: Google Workspace (docs, sheets, calendar)

• gemini cli (nano banana): image generation

• wacli: read and write WhatsApp messages

• op: 1Password

• apify: scrape the web

• postiz: post to social media

• ffmpeg: edit audio and video

• remotion: programmatically create videos

• yt-dlp: download videos

MCPs for when no CLI exists. They're easy to add through the Claude desktop app and Claude Code calls them directly. The trade-off is every active MCP tool definition lives in your context window, so I only add the ones I actually use:

Attio (CRM) • Gmail • Google Calendar • Google Drive • Notion • GitHub • Slack • Context7 (live docs) • Excalidraw • Pipeboard • Tokscript • Typefully

APIs for tools without an MCP or CLI yet. Drop the key in .env and reference docs in api-docs/:

Instantly • MillionVerifier • HeyGen • Apollo

Reference all of them in TOOLS.md so the agent knows what's available and when to reach for which.

🛠️ Layer 3: Skills

Skills are reusable workflows the agent already knows how to run. Set them up once, then just say "run X" and the whole thing executes.

Project-level skills I use:

content-weekly-plan • manychat-browser • multi-platform-video-post • notion-video-resource • rewe-grocery • skool-browser • upwork-thumbnail • video-upload-workflow • youtube-comment-pinner

User-level skills (available everywhere):

canvas-design • docx • excalidraw • frontend-design • notion • pdf • pptx • skill-creator • x-thumbnail • xlsx

My rule: if I run the same multi-step workflow more than twice, it becomes a skill. After that I never re-prompt, I just call it by name.

🤖 Layer 4: Routines

Routines are automations that run without you prompting. This is what turns ClaudeOS from an assistant into an actual operator.

Local routines (run on your machine):

• CEO Brief: daily business summary

• Weekly grocery delivery

• Dream: overnight memory reflection

• Todo list / planner

• Skool Posts

Remote routines (run in the cloud):

• Weekly content plan

• YouTube monitor

Where to start: a daily CEO brief is the highest-leverage local routine to get running first. Once that's stable, layer in remote routines for things that need to run when your machine is off.

🎯 Recommended Build Order

Don't try to ship all 4 layers in one weekend. Stack them:

1. Folder structure and instruction files (Claude.md, Identity.md, SOUL.md)

2. Add User.md and your Business/ context

3. Connect 2 or 3 MCPs you'll actually use today

4. Run for a week with just MCPs before adding CLIs

5. Build your first skill from a workflow you already repeat

6. Add one local routine

7. Layer in remote routines once locals are stable

🔑 TLDR

Set up right, Claude Code can run as a full AI operating system.

Four layers (folder structure and instructions, tools, skills, routines), built in order, and you're running your work through one place instead of bouncing between 10 apps and losing half your context every day.

Start small. Get one layer stable before stacking the next.

Hit reply and let me know what you'd like me to cover next, or just say hi. I read everything.

See you next week.

Moritz

P.S. I'm spending most of my time these days helping founders and agency owners install systems like this in their business, under a new offer called AgentOS. If that sounds useful, here's a 10-minute walkthrough.

If you'd rather DIY it, the Prompt Warrior community has the templates, courses, and weekly Q&As to get you there.