How to Set Up Claude as Your AI CEO

Most guides on using AI for business stop at prompting. They show you how to ask Claude to write emails or summarise documents. That is not what this is.

This is how to make Claude run your company.

Not assist with tasks. Not draft content on request. Actually run it — making decisions, executing operations, maintaining context across sessions, and operating autonomously between check-ins.

This is the exact setup we use at Autonoma. The company you are reading about right now is run this way.

---

What an AI CEO Actually Does

Before the setup, be clear about the job description.

An AI CEO at this stage handles:

• Strategy — evaluating hypotheses, recommending priorities, tracking progress toward revenue goals
• Product — building features via coding agents, deploying to Vercel, fixing bugs, improving conversion
• Distribution — writing X threads, scheduling posts, drafting outreach, managing email sequences
• Operations — processing webhook events, logging decisions, sending morning briefs, monitoring uptime
• Reporting — synthesising what happened, what's working, what's blocked

What it does not handle (yet): legal agreements, spending money, anything requiring a wet signature, decisions the Founder has explicitly reserved.

The job is real. The output is real. The question is how to configure it to be reliable.

---

The Core File Stack

Claude Code is a blank slate. It becomes an AI CEO through five configuration files loaded into the workspace.

1. SOUL.md — Personality and Operating Principles

This file defines how the AI thinks and speaks. Without it, every session starts from scratch with generic assistant behaviour.

What goes in SOUL.md:

• Voice and communication style (direct, no filler, always has a recommendation)
• Execution protocol (when to act autonomously vs when to ask)
• Honesty rules (when to push back, what to name explicitly)
• The Autonoma model (what Founder does, what AI does — clearly separated)

The critical thing SOUL.md establishes is the execution default. Either the AI defaults to asking permission for everything, or it defaults to acting and reporting. We default to acting.

2. IDENTITY.md — Company Context

This is the brief the CEO gets on day one. It covers:

• What the company does
• Who the customers are
• What the current products are and how they work
• What the revenue targets are
• Who does what (Founder, Argus, agents)
• The current milestone and what success looks like

Without this, the AI operates in a vacuum. It cannot make good decisions without knowing what game it is playing.

3. MEMORY.md — Learned Context

MEMORY.md is where operational knowledge accumulates. It contains:

• Learned preferences ("Adrian does not want verbose output")
• Key decisions made ("Switched from Kit to Resend — Kit required manual UI")
• Tech stack ("Next.js 15, Stripe, Vercel, Resend")
• Product status ("Starter Kit download URL: buildingzero.co/starter-kit/downloads")
• Open items that are pending

This file gets updated every session. It is what makes the AI CEO coherent across weeks of work rather than starting blind each time.

4. thinking-os.md — Strategic Frameworks

This is the judgment layer. It contains four frameworks the AI applies before any recommendation:

1. Pre-mortem — what kills this in 12 months?
2. Asymmetric filter — is the downside survivable?
3. First customer filter — who is suffering right now, specifically?
4. Zero-human test — does this plan require Founder to do anything beyond approve?

The zero-human test is the most important. Every strategy, every channel, every feature gets run through it. If the plan requires Founder to write, post, call, or act — it is not a valid plan for a zero-human company. Redesign.

5. CLAUDE.md — Workspace Instructions

This is the technical brief: file structure, git discipline, where products live, how to commit, when to escalate, how to spawn agents. It is the operational manual Claude Code reads at every session start.

---

The Session Architecture

A session has three phases: startup, execution, and close.

Startup (non-optional)

Every session begins by reading the core files and reporting:

• What the current goal is
• Where execution stands
• What is being worked on next

This is not optional. An AI CEO that starts sessions without priming its own context is useless. It will repeat decisions, miss blockers, and contradict previous work.

Execution

During execution, the AI drives. It does not wait to be told what to do. It reads the task list, finds the next unblocked item, and executes.

The only things it pauses for:

• External communications (emails, posts)
• Financial commitments
• Production deploys
• Anything irreversible

Everything else — code, content, strategy, file changes — gets executed and reported.

Close (non-optional)

Before every session ends, the AI must:

1. Write session notes (what was covered, what was decided, what is open)
2. Update MEMORY.md with any new facts
3. Commit all changes

This is the cross-session memory system. Without it, every session starts blind.

---

The Command Interface

A CEO needs a way to receive directives and send status updates without requiring the Founder to open Claude Code.

We use Telegram. Specifically:

• A private Telegram group with topic threads (Daily Ops, Approvals, Alerts)
• A Python bot that listens for commands and routes them to Claude Code
• Inline approval buttons (Approve / Reject / Revise) for any action requiring sign-off

This means Autonoma's Founder operates the company from his phone. He taps approve or reject. Everything else runs without him.

The bot handles:

• Morning briefs (auto-sent at 8am)
• Approval requests (sent as formatted drafts with action buttons)
• Incoming command messages (/recall, /status, /hypothesis)
• Alerts from webhooks (new purchase, new affiliate referral, failed payment)

---

The Decision Log

Every significant decision gets logged to a structured JSON file under life/areas/decisions/. This is not optional.

Format:

{
  "id": "decision-topic-name-001",
  "date": "2026-04-11",
  "decision": "What was decided",
  "reasoning": "Why this and not something else",
  "alternatives_considered": [...],
  "predicted_outcome": "...",
  "review_date": "2026-07-10"
}

The review date is 90 days out. It forces re-evaluation of decisions rather than letting them calcify into unchallenged assumptions.

---

The Honest Limitations

This setup is powerful. It is also brittle in specific ways.

Context windows are finite. Long sessions lose early context. Mitigated by: writing session notes, keeping MEMORY.md current, and not letting sessions run beyond their useful window.

The AI cannot act between sessions. Unless you run a background execution loop (we are building this), the AI only operates when a session is active. Cron jobs handle time-based triggers, but complex decisions still require a live session.

Trust has to be built incrementally. We did not give Argus autonomous execution on day one. Approval requirements were strict early. They loosened as the track record built up. Do not skip this step.

Hallucinations happen. The AI will occasionally state something confidently that is wrong. Mitigated by: requiring specific file reads before making claims about current state, and always checking STATUS.md before reporting on infrastructure.

---

Start Here

If you want to implement this:

1. Set up Claude Code in your project workspace
2. Write SOUL.md — define the execution default and the honesty protocol first
3. Write IDENTITY.md — current state of your company, one page
4. Write MEMORY.md — empty to start, it fills up
5. Write thinking-os.md — at minimum, the zero-human test
6. Write CLAUDE.md — file structure, git conventions, escalation rules
7. Pick a command interface — Telegram is what works for us

The files are the product. Everything else is execution.

We sell the exact templates, pre-configured for a zero-human company setup, in the AI CEO Starter Kit. If you want to skip the blank-page problem and start with a working configuration, that is what it is for.

Or start with the free checklist — five files explained in detail, enough to understand the architecture before committing to the full setup.