How to build a company brain, part 4: the folders, the summary files and a one-week plan

BRAIN.md is a one-page table of where things live. If an agent opens only one file, it should be this one. Kranich's points to the daily logs and weekly reviews, a folder per client, finance, the 30/60/90-day deadlines page, open loops (things waiting on someone else), the email backlog, people and vendors, tools and subscriptions, and a recommendations page that lists the founder's next actions with the evidence behind each.

An agent won't open BRAIN.md just because it exists. The instruction to read it goes in AGENTS.md at the top of the repository. Cursor reads AGENTS.md on its own. Claude Code looks for CLAUDE.md first, and if it finds one it ignores AGENTS.md. So the day 1 prompt below also creates a CLAUDE.md with a single line that imports AGENTS.md. Both tools then load the same rules.

Every folder has a SUMMARY.md with the same sections: purpose, a contents table (file, date, what it is, where it came from), a timeline, the latest update, gaps and open items, and the date it was last verified. An agent reads the folder's summary before it touches anything in the folder. That one rule prevents a lot of filing mistakes, because the summary already says what's missing and what shouldn't be trusted.

In the product brain, every note also starts with a small header: type, title, a one-line description, a status (draft, stable or deprecated), the date it was created, the sources it rests on, and a date after which it should be reviewed.

Material moves in one direction. Anything new lands in a drop folder. An agent files it, writes a summary and leaves the raw part untouched. A fact that changes how things are now goes onto a knowledge page, but only after a person has reviewed it. Decisions go into the log with their reasons. Dashboards read from the brain and never write to it.

Three routines do most of the work. A daily sync every evening goes through mail, calendar, documents and open loops. A weekly review on Monday morning checks deadlines and the loops that have stopped moving. A monthly close on the 1st reconciles the books. Each one reads before it writes.

Two practical notes before you paste the prompt. First, the agent can only read mail and calendar if you've connected them, which in Claude Code and Cursor means a connector or an MCP server. Until you have that, export or forward what matters into inbox/ and the sync works from there.

Second, scheduling depends on the tool. In the Claude Code desktop app you can save the prompt as a scheduled task with a daily time. A local task runs on your machine, and only while the app is open and the computer is awake. If your tool has no scheduler, paste the prompt yourself each evening. I'd do that for the first weeks anyway, because you should be reading what it writes.

Run the daily sync for this brain. First read BRAIN.md and the SUMMARY.md of every folder you touch. Then go through today's new mail and calendar entries, if you have a connector for them, and every file in inbox/ that inbox/SUMMARY.md does not list as filed. If you have no mail or calendar access, say so and work from inbox/ only. Copy each item to the folder where that folder's summary says it belongs, write a digest next to each receipt (vendor, date, amount, category), update the 30/60/90 deadlines page, and note in inbox/SUMMARY.md what you filed and where. Do not send, pay or delete anything, and do not move originals. Write "unknown" for any fact you cannot confirm and say what would confirm it. Finish by writing today's log in the daily logs folder, with the five actions you recommend and a link to the evidence for each.

This is the order I'd use again:

• Day 1, the entry page and the rules. Create a private GitHub repository, open it in your agent and paste the day 1 prompt below.
• Day 2, one real folder. Pick one client, plus finance. Have the agent fill each SUMMARY.md from what's already there, then read both yourself.
• Day 3, the drop folder. Put a week of unfiled notes, receipts and exported mail into inbox/ and paste the day 3 prompt.
• Day 4, the daily sync. Run the sync prompt above by hand and read the log it writes. Schedule it once a run comes back clean. Add the Monday review and the monthly close after that.
• Day 5, sharing. Set up branches and pull requests so other people can work in it (part 5).
• Day 6, skills. Write your first two skills. Organise and research are the usual pair (part 6).
• Day 7, privacy. Write the list of what never enters the brain and set up the quarantined folder (part 7).

Create the start of a company brain in this repository. Write BRAIN.md with a "where things live" table for: daily logs, weekly reviews, inbox/, clients/, finance/, deadlines (next 30, 60 and 90 days), open loops, email backlog, people and vendors, tools and subscriptions, recommendations. Create each folder with a SUMMARY.md that has these sections: Purpose, Contents (file, date, what it is, source), Timeline, Latest update, Gaps and open items, Last verified. Add AGENTS.md with these rules: read BRAIN.md and the folder's SUMMARY.md before changing anything; never delete; never move an original, copy it; never invent a fact, write "unknown" and say what would confirm it; agents draft, people send, pay and file. Add a CLAUDE.md that contains only the line @AGENTS.md. Do not invent content. Leave a section empty where you have nothing to put in it.

Organise inbox/. Read BRAIN.md and inbox/SUMMARY.md first. For each file: work out what it is, copy it to the folder where that folder's SUMMARY.md says it belongs, add a row to that folder's Contents table (file, date, what it is, source), and write a short summary next to the copy. Leave the original in inbox/ and note in inbox/SUMMARY.md where its copy went. If you cannot tell where something belongs, do not guess: list it under Gaps and open items and ask me. Do not delete anything.

Privacy is day 7 here only because the rules are easier to write once you've seen what actually arrives. It doesn't mean the first six days are a free-for-all. Until the never-enter list and the quarantined folder exist, keep personal data about customers and staff out of inbox/, and keep passwords, keys and tokens out for good. Everything in the repository is something your agent may read and send to its model.