AI Company OS Starter

How it works

Most people use Claude like a chat window. Every time, they explain their business again. This setup fixes that with 4 parts that live in one folder.

The brainA few files with your business facts: what you sell, to whom, your prices, how you talk. Every agent reads them first.

The agentsOne per department. Each knows its jobs and its rules, so you just say what you need.

The job mapEvery repeatable job in your business, with a grade that says how much AI can do alone.

The fourth part is the router: a CLAUDE.md file that tells Claude which agent gets which request, and the rules every agent follows.

What mine looks like today: 11 agents, 60 skills and 168 jobs across my business and my school. I started with 4. Start small, use it every day, and add a department when you feel the gap.

Before you start

Claude Code, installed and signed in: claude.com/product/claude-code. It needs a Claude plan that includes Claude Code.
An empty folder for your company OS.
Your answers. Claude interviews you. The more honest and specific you are, the better every agent gets.

Build it

Pick one way. The one-prompt version is faster. Step by step lets you check each part.

1. Open your folder in Claude Code

Terminal

mkdir my-company-os
cd my-company-os
claude

2. Paste this, then answer its questions

Prompt

Build me an AI company OS in this folder. Work in this order and check with me after each part.

1. THE BRAIN. Interview me, one question at a time, and wait for each answer. Cover: what I sell and to whom, my prices, my goals for the next 90 days, the tools I use, and how I write and talk. Then write brain/company.md, brain/offers.md and brain/voice.md. Anything I did not answer, write UNKNOWN. Never invent a fact.

2. THE JOB MAP. List every repeatable job in my business, grouped by department. Give each job a grade: A if AI can do it alone and report after, B if AI drafts and I approve, C if I do it and AI helps. Start most jobs at B or C. Save it as registry/jobs.csv with columns department, job, grade, notes. Show me the 3 departments with the most jobs.

3. THE AGENTS. In .claude/agents/, create one agent for each of those 3 departments, plus a chief-of-staff that runs my daily standup and plans my week. Each agent reads brain/ first, only does its own department's jobs at their grade, never sends, posts or pays for anything, and ends every answer with SOURCES, ASSUMPTIONS, GRADE and NEXT.

4. THE TRACKER. Create trackers/tasks.md: a table with id, task, department, due and status. When I give you a task, add a row.

5. THE ROUTER. Write CLAUDE.md: a table of which agent handles which request, and these rules: draft only, I send; never invent a number, write UNKNOWN; ask one question at a time; log every decision in brain/decisions.md with the date; respect each job's grade.

When it's done, show me the folder and 3 things I can say to test it.

Write the brain

In your folder, run claude and paste:

Interview me to build the brain of my AI company OS. Ask one question at a time and wait for my answer. Cover what I sell and to whom, my prices, my goals for the next 90 days, the tools I use, and how I write and talk. Then write brain/company.md, brain/offers.md and brain/voice.md. Anything I did not answer, write UNKNOWN. Never invent a fact.

You get brain/company.md, brain/offers.md, brain/voice.md

Map every job and grade it

Read brain/. List every repeatable job in my business, grouped by department. Give each job a grade: A if AI can do it alone and report after, B if AI drafts and I approve, C if I do it and AI helps. Start most jobs at B or C. Save it as registry/jobs.csv with columns department, job, grade, notes. Then show me the 3 departments with the most jobs.

You get registry/jobs.csv

Create the agents

Ask Claude to make one agent per department you picked, plus a chief of staff. Each agent is one file. Here is the shape:

.claude/agents/content.md

---
name: content
description: Owns content ideas, hooks, scripts and captions. Use when I say video idea, write a hook, script this, or caption.
tools: Read, Write, Edit, Glob, Grep, WebSearch
---
You run content for my business. Before any task, read brain/company.md and brain/voice.md.

## Your jobs
Only the CONTENT jobs in registry/jobs.csv, at their grade.

## Never
- Never post or schedule anything. Draft only. I post.
- Never invent a number. Write UNKNOWN and ask me.

## Every answer ends with
SOURCES, ASSUMPTIONS, GRADE, NEXT

You get one file per agent in .claude/agents/

Add a tracker

Create trackers/tasks.md: a table with id, task, department, due and status. When I give you a task, add a row. In every standup, read it and tell me what is due.

You get trackers/tasks.md

Write the router

Claude Code reads CLAUDE.md in your folder at the start of every chat. This is what makes it act like your company.

CLAUDE.md

# MY COMPANY OS

@brain/company.md
@brain/voice.md

This folder runs my business. Send each request to the right agent, check the answer, and keep the log.

## Routing
| Request | Agent |
|---|---|
| plan my day, plan my week, what is waiting on me | chief-of-staff |
| video ideas, hooks, scripts, captions | content |
| new lead, follow up, proposal | sales |
| client onboarding, monthly report | delivery |

## Rules
1. Draft only. Never send, post, publish or pay anything. I do that.
2. Never invent a number, price or client detail. Write UNKNOWN and ask me.
3. Ask one question at a time.
4. Log every decision in brain/decisions.md with the date.
5. Respect each job's grade in registry/jobs.csv.

## Grades
- A: runs on its own and reports after.
- B: drafts, I approve.
- C: I do it, the agent helps.

Change the routing table to your own agents.

Test it

Close Claude Code, open it again in the folder, and try these 3:

run my standup
write 5 hooks for my next video
I just got a new lead: [paste their message]. What do I reply?

If an answer misses a fact, the fix is almost always the brain. Add the fact to brain/ and ask again.

Grade every job

The grade is what makes this safe. It tells you, job by job, how much you can hand off right now.

A: autonomousRuns on its own and reports after. For example: pulling last week's numbers.

B: assistedWrites a draft and waits for your yes. For example: a reply to a new lead.

C: human-ledYou do it. The agent preps and helps. For example: a sales call.

When a grade changes

C to B after 5 drafts you were happy with.
B to A after 10 drafts in a row you didn't edit.
Overridden 3 times in a row? Rewrite that agent's instructions, or take the job back.

In my system right now: 61 jobs run on their own, 94 draft and wait for me, and 13 are still mine. It doesn't run itself. It runs the boring parts, and I decide.

Use it daily

Morning: "run my standup". It reads your tracker and tells you what's due today.
During the day: ask for what you need in plain words. The router picks the agent.
When something changes: a new price, a new client, a new rule. Say "save this to the brain".
Sunday: "run my weekly review". What worked, what didn't, what to change next week.

Grow it

Once the starter feels normal, these are the upgrades I added, in the order I'd add them:

More departments as you feel the gap. Mine include brand deals, mentorship, community, back office and school.
Skills for jobs you repeat often, so every run follows the same steps.
Real trackers for leads, clients and content, instead of one task list.
A dashboard to chat with each department and approve drafts in one place.
A morning message on your phone with what's due, through a Telegram bot.

The full tour of my system is coming to my YouTube channel.

Want help building yours?

Inside my community I walk through how I build AI systems, step by step, and you can ask questions on your own setup. Try it free for 7 days.

Join the community