The Company Brain Starter Kit, Explained

By Darby Rollins · Published August 6, 2026

Key takeaways

The kit is 46 files: four at the root that launch it, and 42 that become the live brain.
One prompt runs the whole setup. Everything else in the kit is either read by the AI or filled in during that conversation.
Point Claude at the Company-Brain-Atlas folder, not the zip root, or CLAUDE.md will not load.
Six files ship as deliberate blanks. soul.md is the important one and nothing else gets built until you confirm it.
Seven folders ship containing only a README. That is intentional, not a broken download.
The kit says three steps. The workshop said hours. Both are true and they describe different things.
The kit's soul.md is your company's identity, not your assistant's personality. Those are different files.

At the end of the live workshop we handed out a zip: a complete company brain folder, pre-structured, with one prompt that conducts the entire setup. It is free and it does not require an account.

This page documents what is in it, file by file, and what happens when you run the setup.

Get it from os.agenticsociety.com/company-brain. Read this before you run anything.

What you actually downloaded

The zip has two layers.

Four files at the root launch the setup and then get left behind. They are not part of the live brain.

START-HERE.md carries the one prompt and the two prerequisites. This is the ignition.
QUICK-START-GUIDE.md is the map: what you downloaded, three install steps, standing rules, troubleshooting.
VERSION.md is a six-line version stamp.
WHATS-NEXT.md describes what the paid tier adds.

Company-Brain-Atlas/ is the brain itself, 42 files. This is the folder you point Claude at, and this is the part that lives on after setup.

Point your AI at Company-Brain-Atlas/, not at the zip root. CLAUDE.md sits inside Company-Brain-Atlas/, and Claude Code auto-loads it only if that is the folder you selected. Point at the root and it will not load, and nothing will behave the way the demo did.

The six pieces that matter

Everything else in the kit supports these.

File	Its job	You fill it in?
`soul.md`	Who you are. Read first, every session	Yes. Ships as a labelled blank
`CLAUDE.md`	How to read the brain. Boot order, filing rule, the hard gates	No. Leave it alone
`company-brain.md`	The front door. The map, plus your Chief of Staff’s charter	Partly. Three charter blanks
`Company/roles/`	What each seat’s AI is for. One file per seat	Yes. Only `owner.md` ships
Drive access	Who sees what. Google Groups and folder restrictions	Yes. This is not a file
`Daily-Notes/`	What happened. The ledger every session appends to	No. The AI writes these

soul.md gates everything else

It ships empty on purpose. Every line is a prompt in parentheses rather than content, and it ends with a status marker saying that until you fill it in and confirm it, every line is a hypothesis.

The filing rule in CLAUDE.md says any new context contradicting soul.md triggers a stop. Until soul.md has real content, that check has nothing to check against, which is why the starter prompt builds nothing else first.

One confusion to head off. In this kit, soul.md is your company’s identity: what you sell, how you decide, your voice, your red lines. In the workshop demo, the soul file I was describing is my Chief of Staff’s persona, including personal reminders like the Friday date-night rule. Both are called soul.md and they are not the same thing. Your assistant’s personality belongs in its own skill file, which the setup creates in Module 2. Put charter material into the company’s soul.md and you will trip the Module 1 gate.

CLAUDE.md is the contract

You never open it. It is the only file the harness loads without being asked, and it carries five things: the boot order, an eight-step filing rule, write guardrails, the hard gates, and the access note.

The hard gates: nothing sends, publishes, spends or creates accounts without your explicit go. Never delete, archive instead. Write only inside the brain folder.

company-brain.md has an invisible line in it

The file is two documents separated by an HTML comment: above the line is your approved charter, below it is the map the AI maintains freely.

That comment does not render. In a Drive preview or any markdown viewer you see one continuous document with no visible boundary. Keep it in mind before you edit the file.

The starter prompt

This is the whole setup. Copy it, paste it into Claude Code with the Company-Brain-Atlas/ folder selected, and answer the questions.

You are my AI Chief of Staff and onboarding conductor. My Company Brain lives in this Company-Brain-Atlas/ folder — read, in order: CLAUDE.md, soul.md, company-brain.md, then your SOP: Company/onboarding-runbook.md. Then run the modules with me, in order, one gate at a time.

First question to me — which path fits?
- HARVEST (I already run a business with folders full of context): I'll attach my existing folders. You index everything into Company/sources-map.md — asking me one question at a time when a folder's purpose isn't obvious — then DRAFT my soul.md and ownership picture from my documents (holdco, entities, locations — whatever is actually real), tag every claim CONFIRMED or INFERRED, and walk me through the INFERRED list to correct. Then we run the migration sweep, one approved branch at a time.
- INTERVIEW (starting fresh): ask me the five soul questions, one at a time, and capture my answers live.

Standing rules, every session:
- ONE BRAIN ONLY. If a Company Brain already exists anywhere I can see (a soul.md alongside a company-brain.md — this folder or another), we build on that one: adopt it, fill what's missing in place, and never scaffold a second brain, a duplicate folder tree, or copies of files that already exist. If you find two candidate brains, stop and ask me which is canonical before touching either.
- Identity first — nothing else is built until I confirm soul.md.
- Help me name you, write your charter from my goal, org chart, and bottleneck, and save my core skill in this workspace (start from Company/skills/my-chief-of-staff.TEMPLATE.md — fill the blanks, keep the boot protocol verbatim) so I open it and talk to you from now on.
- Confirm before writing. Never delete — superseded goes to Archive/. Nothing sends or shares without my explicit go.
- Every folder you create gets a one-paragraph README.
- Anything nobody can answer from the brain goes to Company/faq.md under Unanswered.
- Set up the rhythms when we reach Module 4: the daily journal close and the call-filing sweep, both in Company/Playbooks/ — schedule the daily close as a recurring task (Claude Code, end of business day).
- End with the re-architecture + postmortem, then the proof tests. Do not skip them.
- Any time we ingest a large new batch of context later, offer another re-architecture pass — the structure follows the knowledge.

Start now: ask me which path.

Three of those standing rules exist because of things that went wrong in practice. ONE BRAIN ONLY stops the AI scaffolding a duplicate folder tree next to a brain you already have. Identity first stops it building on assumptions you never confirmed. Confirm before writing, never delete, nothing sends is the gate that keeps an eager assistant from doing something on your behalf that you cannot undo.

If you already have a brain from our earlier training, you do not need to start over. Paste this and then ask what it would improve about what you already built.

What happens when you paste it

The prompt hands the session to the AI. It reads CLAUDE.md, soul.md and company-brain.md, then opens Company/onboarding-runbook.md, which is a six-module SOP written for the agent that you never have to read.

Module 0 checks it can actually read the folder. If it cannot, it stops and helps you fix the setup rather than proceeding on a broken install.

Module 1 confirms soul.md. In harvest mode it drafts this from your documents and tags every claim CONFIRMED or INFERRED, then walks you through the inferred ones. Nothing proceeds until you say yes.

Module 2 names your Chief of Staff, writes its charter from your goal, org chart and bottleneck, and saves your core skill. From here on you open that skill and talk, instead of pasting the prompt again.

Module 3 is fill or migrate, and it is the long one. In harvest mode it builds a sources map first, then proposes a migration table and moves one approved branch at a time, copying and never cutting. The milestone is that you can ask anything about the business and it finds the answer.

Module 4 installs the two rhythms: the daily journal close and the call-filing sweep.

Module 5 re-architects. It proposes the best structure for your specific business as a before and after tree, writes a postmortem, then runs three proof tests including the negative test: ask a walled question from an unauthorised login and confirm it comes back empty.

Six files to fill in

Everything else either runs itself or fills up through use.

soul.md with the owner, confirmed before anything else
Three charter blanks in company-brain.md: your number one goal, your org chart today, your constraints
Company/call-filing.md, the routing table for recordings. Its header says to customise it before your first sweep
Company/connectors.md, registering each live system
Company/sources-map.md, built during the harvest
Company/roles/_TEMPLATE-role.md, copied once per real seat

About those empty folders

Seven folders ship containing nothing but a README: Company/SOPs/, Company/standing-rules/, Company/Vendors/, Company/Transcript-Vault/, Archive/, Private/ and Daily-Notes/.

That is intentional. Your download is not broken. They fill up through operation, which is the point: the structure follows the knowledge rather than the other way round.

The one worth filling in by hand early is Private/. Payroll, tax records, bank statements. Owner only.

Before you start

These came up during the live build.

Google Workspace, if you want the team version. Shared Drives and Google Groups need it. On a personal Gmail you can still build a version one in a local folder, and everything on this page works. The permissions model is the part that needs Workspace. The prep page covers this in full.

The five soul questions, so you can think about them in advance rather than on the spot: who you are and your ownership model, what you sell in your own words, your voice with three to six real lines of your own writing, what the business is not, and your red lines.

It will drift, and you should read what it produces. This came up twice on the call. Claude sometimes goes in a direction that is subtly off, in a way that causes problems later if you do not catch it early. Watch the first few sessions the way you would watch a new hire’s first few days. It takes time to train this.

Your Company and Atlas are placeholders. They appear in the file headings as they ship. Setup replaces them with your real names during the conversation.

How long this takes

The kit says three steps. The workshop said hours. Both are true and they are describing different things.

Install is three steps and about five minutes: unzip it, open your AI inside the folder, paste the prompt.

Setup is a multi-session project. Gathering your context is an hour minimum. The build itself runs several hours, better as an hour a day over a few days than one long push. The migration was estimated at up to four hours. And the harvest run itself, in our live demo against roughly 600 files, was still going forty minutes later and did not finish inside the session.

None of that is a reason to avoid it. It is a reason not to start it thirty minutes before a meeting.

The prep list covers what to have ready before you unzip. The playbooks are what to run once it exists. The full workshop is the reasoning behind all of it.