Hands-on · Built to run with Claude

The Company Brain Guide

The educational version of this idea lives one page back: why a pile of documents is not a brain, why the challenge loop has real science behind it, and why rulings must persist. This page is the build. Ten phases, in order, each with the exact prompt to hand to Claude. You need no other tooling.

Foundation sources build the base Living signals keep it honest

Everything accumulates in one artifact: your Brain file, a single document called brain.md that you keep somewhere safe. Phase by phase it grows from a source inventory into an agreed, evidenced, versioned understanding of your company: who you serve, what you claim, how you talk, and what you have decided when the evidence disagreed. That file is the deliverable. Everything else is process.

One setup step saves you friction all the way through: create a Claude Project, put brain.md in its project knowledge, and update it there after each phase. Every prompt in this guide then works without pasting the file. No Projects on your plan? Paste brain.md into each session instead; everything still works, it is just more copying.

A warning before you start, from someone who has watched companies do this badly for 30 years: the value is not in the writing, it is in the rulings. When the brain shows you a contradiction, the temptation is to explain it away. Do not. Confirm it or rebut it, record the decision, and move on. An agreed understanding you actually agreed to beats an eloquent one you never tested.

The build, at a glance

Phase	Title	What you get
01	Gather the corpus	A dated inventory of every identity source
02	Extract who you say you are	A baseline identity with every claim sourced
03	Audit what you publish	The say side of your say-do gap, measured
04	Audit who you win	The do side: who actually buys, from deal data
05	Adjudicate: the first challenge session	Agreed Understanding v1.0, with rulings recorded
06	Wire in the calls	A verbatim language bank and the first living challenges
07	Read the search data	Demand evidence, in the market’s own phrasing
08	Install the gate	A reusable pre-publish check for everything you ship
09	Brief the company	A one-page briefing everyone who speaks for you can use
10	Run the monthly loop	A versioned brain, a dated log, and a habit

Phases 1 to 5 build the brain and take a focused day, spread over a week if you prefer. Phases 6 to 10 wire in the living signals and settle into a monthly hour. The thinking behind every phase is on the educational page.

Gather the corpus

What you get: A dated inventory of every identity source

2 to 4 hours of gathering, 20 minutes with Claude

Why

No model can tell which of two conflicting documents is the current one, because authority is not in the text. It lives in dates, ownership, and what supersedes what. So the build starts unglamorously: list every place your company describes itself, and put governance metadata on each one. This inventory becomes the first section of your Brain file, and every claim you record later will cite back to it.

Do this

1 Collect what exists: brand guide, website copy, the last 20 blog posts or articles, sales decks, one-pagers, the ICP (ideal customer profile: the specific kind of customer you have decided to pursue) or positioning slide, onboarding docs.
2 Attach files or paste text; both always work. Links only work if your Claude can browse the web, and pages behind a login (Google Docs, Notion, Canva) usually fail even then, so export those to PDF first.
3 For each source, note what you know: when it was written, who owns it, and whether anyone still treats it as current. Do not clean anything up; stale and contradictory sources are the raw material.
4 Run the prompt below with your list. It will propose an authority ranking and wait for you to confirm it before writing the final inventory.
5 Save the output as the start of a document called brain.md, and put that file in your Claude Project knowledge if you made one. This file is the asset the whole guide builds.

Hand it to Claude

You are helping me build a company brain: a single reconciled, evidenced understanding of my company that I will maintain in one document called my Brain file. This first session creates the Source Inventory.

Rules, which you must follow exactly:
1. Include ONLY the sources I list below. Never add a source I did not name.
2. For each source, record: name, type (brand guide, website page, article, deck, other), date or effective period, owner, and status (current / stale / unknown). Where I have not given you a date or owner, write [UNKNOWN] rather than guessing.
3. Propose an authority ranking: which source should win when two sources disagree. Base it on governance (most recent, most official, closest to the money), not on which is best written.
4. This is a two-step session. First present the proposed ranking and stop, asking me to confirm or reorder it. Only after I answer, output the final Source Inventory section with the ranking marked as confirmed.
5. Note the gaps: which identity sources a company my size would normally have that I did not list (for example win/loss notes, call transcripts, a documented ICP). List them as [MISSING] with one line on why each matters.
6. Output everything as a clean markdown section titled "Source Inventory" with the confirmed authority ranking at the top, ready to paste into a file called brain.md.

My sources:
- [SOURCE 1: name, what it is, date if known, owner if known]
- [SOURCE 2: ...]
- [SOURCE 3: ...]

My company, one line of context: [COMPANY NAME], [WHAT IT DOES, ROUGHLY].

Check it worked. brain.md exists and starts with a Source Inventory in which every source has a date, an owner, and a status, or an honest [UNKNOWN]. If everything is marked current and nothing is marked missing, you were not honest enough. Run it again.

Extract who you say you are

What you get: A baseline identity with every claim sourced

about an hour once materials are in hand

Why

This layer is deliberately your version of the story, because you cannot correct a story you never wrote down. The output is a ranked identity: category, ICP, value claims in priority order with proof points, and the voice you write in. Ranking is the point. When two claims compete for one headline, the hierarchy has already decided. And every conflict between your own sources gets recorded, not resolved. Resolving comes later, and it involves you.

Do this

1 Open a fresh Claude session. Attach the brand guide, your 3 to 5 most important website pages, 5 recent articles, and your main sales deck. More does not improve the extraction; it dilutes it.
2 Make sure Claude has your brain.md (project knowledge, or paste it) for the Source Inventory and authority ranking.
3 Run the prompt below.
4 Read the Conflict Register it produces. If it found no conflicts at all, be suspicious; a company whose sources fully agree is vanishingly rare.
5 Append the output to brain.md. Do not resolve anything yet.

Hand it to Claude

You are building the Baseline Identity section of my company Brain file. I have attached my source materials, and my brain.md contains the Source Inventory and its confirmed authority ranking.

Rules, which you must follow exactly:
1. Extract, do not invent. Every statement you write must cite the source it came from, in the form [source: NAME]. If the materials do not answer something important, write [TODO: not stated in any source] rather than filling the gap plausibly.
2. Produce these subsections:
- Market category: what kind of thing my company says it is.
- Stated ICP: who we say we serve, as specific checkable attributes (industry, company size, buyer title, trigger) rather than adjectives.
- Value claims, ranked: at most four, in priority order, each with its proof point. Where a claim appears with no evidence anywhere in my materials, keep the claim but mark it [UNPROVEN].
- Voice and vocabulary: how we actually write, words and phrases we use constantly, and any words the brand guide bans.
3. Where sources disagree with each other, do NOT pick a side, average them, or smooth the difference over. Record each disagreement in a subsection called "Conflict Register" as a numbered entry: what source A says, what source B says, quoted verbatim, with the source names. Number the entries sequentially starting at 1. Challenge numbers are permanent for the life of the Brain file: future sessions continue the numbering and never reuse a number. Silent conflict resolution is the failure mode this whole system exists to prevent.
4. Rank nothing as more authoritative than the confirmed authority ranking in my Source Inventory allows.
5. Output one clean markdown section titled "Baseline Identity v0.1" containing the subsections above, ready to append to brain.md.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

Check it worked. Pick any three claims in the Baseline Identity and check the citation on each: open the named source and find the claim. Then count the Conflict Register. One or more recorded conflicts means it is working.

Audit what you publish

What you get: The say side of your say-do gap, measured

about an hour

Why

Your published content is your theory in use: it shows who you actually write for, whatever the strategy slide says. This audit is mechanical on purpose. A fixed window, no cherry-picking, every piece tagged, a distribution at the end. The gaps it finds become challenges, written as questions with evidence on both sides, because that is what the adjudication session in Phase 05 consumes.

Do this

1 Gather your last 20 published pieces: articles, posts, talks, newsletters. Fewer than 20? Use everything you have and tell Claude the count; the distribution works at any size, it just gets noisier.
2 For each piece, paste at least the full opening section, and the full text for short ones. Titles alone are not enough for an honest audience judgment.
3 Fixed window, most recent first, no skipping the embarrassing ones.
4 Run the prompt below with your brain.md available.
5 Read the distribution before you read the challenges. The number usually stings. That is the point of measuring. Then append the new Conflict Register entries to brain.md.

Hand it to Claude

You are running the content audit for my company Brain file: comparing what we publish against who we say we are. My brain.md, including the Baseline Identity with its stated ICP and value claims, is available, along with my recent published pieces.

Rules, which you must follow exactly:
1. Tag every piece with: the audience it is actually written for (be specific: role and company size, judged from the content itself, not the title), the problem it addresses, and which of my ranked value claims it supports. A piece supports a claim only if the support is explicit; do not stretch.
2. Where I gave you only a title and a summary for a piece, tag it anyway but mark the tag [LOW CONFIDENCE] and exclude it from the distribution unless I confirm the tag.
3. Compute the distribution: what share of the pieces addresses the stated ICP, what share supports each ranked claim, and what share supports none of them.
4. Where the distribution diverges materially from the Baseline Identity, write each divergence as a numbered challenge in the Conflict Register, continuing the existing numbering (check the register for the highest number; never reuse one). Phrase each as a question with the evidence for both sides. Example of the form: "The stated ICP is mid-market operations leaders [source: positioning deck]. 14 of 20 recent pieces address enterprise IT buyers [pieces 2, 4, 5...]. Which is right?"
5. Do not soften the findings and do not editorialise about what we should do. Your job in this session is measurement and honest questions, not advice. A challenge I later confirm and a challenge I later rebut are equally successful outcomes for you.
6. Output: a tagging table, the distribution summary, and the new Conflict Register entries, in markdown, ready to append to brain.md.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

My recent pieces (full text or opening sections):
[PASTE THEM]

Check it worked. You can state, as a number, what share of your recent content addresses the ICP you claim. If someone asked you that question yesterday you would have guessed. Now you know, and the difference between the guess and the number is the size of your say-do gap.

Audit who you win

What you get: The do side: who actually buys, from deal data

about an hour once the deal list exists

Why

Positioning worth having is derived from the customers you actually win and delight, not the ones on the aspiration slide. Deal data is where reality keeps its records. Two honesty rules, and the prompt enforces both: call nothing a pattern under three deals, and treat rep-entered loss reasons as weak evidence, because reps and buyers routinely disagree about why deals close.

Do this

1 Export your last 20 to 30 closed-won deals (deals you actually signed) from the CRM. Strip customer names; keep industry, company size, buyer title, deal size, sales cycle length, and retention or expansion if you have it.
2 No CRM? Build the list from memory and invoices: for each customer, industry, rough size, who signed off, roughly how long from first conversation to signature, and whether they are still a customer. Ten rows honestly filled beats thirty guessed.
3 Add closed-lost deals if you can. If the CRM has recorded loss reasons, include them, clearly labelled as CRM-recorded.
4 Run the prompt below with your brain.md available, and look hardest at the segment that closes fastest and stays longest. That is the market voting.
5 Append the new challenges to brain.md.

Hand it to Claude

You are running the deal audit for my company Brain file: comparing who we say we serve against who actually buys. My brain.md is available, along with an anonymised export of recent closed deals.

Rules, which you must follow exactly:
1. Score every closed-won deal against the stated ICP attributes from my Baseline Identity, attribute by attribute. Report the overall match share and the match share per attribute.
2. Find the pattern in the best deals: which segment closes fastest, which stays longest or expands, and what the champion (the buyer-side person who drove the deal) has in common. A pattern requires at least three deals; below that, report it as an observation, not a finding.
3. Treat CRM-recorded loss reasons as weak evidence and label them so; reps and buyers routinely disagree about why deals close. If loss reasons matter to a conclusion, recommend which two or three lost buyers I should actually interview, and give me the three questions to ask, starting with "walk me through the decision".
4. Where the won-deal profile diverges from the stated ICP, write each divergence as a numbered challenge in the Conflict Register, continuing the existing numbering (never reuse a number): evidence on both sides, phrased as a question, no recommendation attached.
5. Do not flatter the data. If the sample is too small or too skewed to conclude anything, say exactly that.
6. Output: the scoring table, the best-deal pattern (or the honest statement that there is not one yet), and the new Conflict Register entries, in markdown, ready to append to brain.md.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

My deal export (anonymised):
[PASTE ROWS: industry, size, buyer title, deal size, cycle length, won/lost, retention if known]

Check it worked. You know which segment closes fastest and retains best, and whether it matches the slide. If the answer surprised you, the audit worked. If the sample was too thin to say, you learned something just as valuable: your next 90 days of evidence-collection have a job.

Adjudicate: the first challenge session

What you get: Agreed Understanding v1.0, with rulings recorded

one quiet hour

Why

This is the phase that makes it a brain instead of a report. Every contradiction gathered so far gets presented to you, one at a time, with the evidence for both sides, and you rule on it: confirm the stated identity, rebut it in favour of what the evidence shows, or amend it. The ruling is recorded with the date, the evidence, and your reasoning. Machine learning calls the underlying discipline active learning; data management calls it stewardship. Either way, the rule is the same: the ruling must persist, or the same question comes back forever.

Do this

1 Set aside a genuinely quiet hour. This is judgment work, the part only you can do.
2 Run the prompt below with your full brain.md available, now carrying the Conflict Register entries from Phases 02 to 04.
3 Rule on every challenge. Confirm, rebut, or amend. If you genuinely cannot rule, park it with a named piece of evidence that would settle it.
4 Update brain.md by replacing the changed sections with what the session outputs. Before saving, scroll the output and confirm no section says anything like "rest unchanged"; if it does, ask for the missing section by name.

Hand it to Claude

You are running the adjudication session for my company Brain file. My brain.md is available, with a Conflict Register of numbered challenges built up from my own sources, my content audit, and my deal audit. In this session I rule on each one, and you record the rulings.

Rules, which you must follow exactly:
1. Present ONE challenge at a time: the question, the strongest honest case for each side from the recorded evidence, and nothing else. No recommendation, no lean. Then stop and wait for my ruling.
2. I will answer with one of: CONFIRM (the stated identity stands), REBUT (the evidence wins; change the identity), or AMEND (a third formulation, which I will give you). If I try to explain a contradiction away without ruling, ask me plainly: confirm, rebut, or amend?
3. Immediately after each ruling, print the completed Rulings Log entry for it: date, challenge number, the ruling, my stated reasoning, and the evidence considered. Printing as we go means nothing is lost if this session ends early. Rulings persist: in any future session, a ruled question may only be reopened by NEW evidence, never re-asked because the old evidence still exists.
4. If my ruling contradicts strong evidence, record the ruling faithfully AND add one flat sentence noting the tension, for example: "Ruled CONFIRM; note that 14 of 20 recent pieces still address a different audience." You are not here to argue with me, and you are also not here to pretend.
5. If I cannot rule, record the challenge as PARKED with the specific evidence that would settle it and where it would come from. Parked challenges stay open.
6. After all challenges are ruled, output ONLY the changed sections, each complete and clearly titled, for me to paste over the matching sections of my master file: (a) "Agreed Understanding v1.0", the Baseline Identity rewritten to reflect every ruling, with each changed line marked [ruled: date, challenge #]; (b) the complete "Rulings Log"; (c) the updated "Conflict Register", with resolved entries marked RESOLVED and their numbers retained (numbering is permanent; future challenges continue from the highest number ever used), and parked entries marked PARKED. Do not regenerate the Source Inventory or any unchanged section.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

Check it worked. brain.md now says Agreed Understanding v1.0 at the top of the identity section, the Rulings Log has a dated entry for every challenge, and every Conflict Register entry is marked RESOLVED or PARKED. Read the identity out loud. For the first time, every sentence in it is something you actually decided.

Wire in the calls

What you get: A verbatim language bank and the first living challenges

45 minutes, then monthly

Why

A brand guide is a photograph; your sales calls are the moving picture. Conversation-intelligence practice extracts a small stable set of signals from every call: competitor mentions, objection phrases, and the exact words customers use. Verbatim matters. Voice-of-customer copywriting practice mines transcripts for the customer’s own sentences because paraphrasing strips the phrasing that makes messages land. This phase starts your language bank and tests the agreed understanding against live conversation.

Do this

1 Get transcripts of two or three recent sales or customer calls. Any recording tool’s raw transcript is fine; so are detailed call notes if that is what you have.
2 Neither recordings nor notes? Start recording your next three calls (ask permission on the call) and come back in two weeks. This phase is worth the wait; do not skip it permanently.
3 Strip names if you prefer; the content is what matters.
4 Run the prompt below with your brain.md available, then append the output.
5 Repeat with a couple of fresh calls every month. One call is an anecdote; the value compounds when themes recur.

Hand it to Claude

You are mining call transcripts for my company Brain file. My brain.md (with the Agreed Understanding) is available, and recent call transcripts are pasted below.

Rules, which you must follow exactly:
1. Extract into a "Language Bank" section, everything verbatim, never paraphrased:
- Competitor and alternative mentions: who the customer named, including "we would just use a spreadsheet" style non-vendor alternatives, and whether it came up early or late in the call.
- Objection phrases: the customer's exact words for every hesitation or pushback.
- Value phrases: the customer's exact words for what they want, love, or fear, sorted into three lists: love, hate, worry.
- Vocabulary gaps: terms the customer used for their problem that appear nowhere in my Agreed Understanding's vocabulary, and terms we use that no customer said.
2. Compare what you extracted against the Agreed Understanding. Where the calls support it, say so in one line each. Where they contradict it, apply the pattern rule: a theme that appears in at least two separate calls becomes a numbered challenge in the Conflict Register, continuing the existing numbering (never reuse a number); a theme from a single call goes on a "Watch List" instead. Every Watch List entry carries two dates: date added and date last seen.
3. Respect the Rulings Log: do not re-raise anything already ruled unless these transcripts contain genuinely new evidence, and if they do, say explicitly what is new.
4. No advice, no rewriting of my messaging in this session. Extraction, comparison, challenges, nothing else.
5. Output: the Language Bank, the comparison notes, any new challenges or watch-list items, in markdown, ready to append to brain.md.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

Transcripts:
[PASTE THEM]

Check it worked. The Language Bank contains sentences you would never have written but a customer actually said. Read the love list out loud and compare it with your homepage headline. If the homepage loses, you have found next month’s most valuable edit, and it costs nothing.

Read the search data

What you get: Demand evidence, in the market’s own phrasing

30 minutes, then quarterly

Why

Search data is identity evidence you already own: the queries your site appears for show how the market phrases your problem space, rising impressions show demand forming, and queries ranking positions 8 to 20 show topics where you are almost credible. Standard practice, and the prompt enforces it: exclude your brand name first, because brand traffic drowns the signal of what strangers actually want.

Do this

1 Open Google Search Console, Performance report, last three months, and export the Queries table. Never set it up? Do it now (your web person or host can verify ownership in 15 minutes), skip this phase, and come back in a quarter when data exists. Nothing downstream depends on it.
2 Delete rows containing your brand or product names before pasting; you want the strangers, not the fans.
3 Run the prompt below with your brain.md available.
4 Append the output to brain.md, and rerun this quarterly; search moves slower than calls.

Hand it to Claude

You are reading search demand for my company Brain file. My brain.md is available, and a Google Search Console query export (brand terms removed, last three months) is pasted below.

Rules, which you must follow exactly:
1. First, judge the data honestly. If the export is too thin to cluster meaningfully, say exactly that and stop; do not manufacture clusters from noise.
2. Cluster the queries by intent, using modifiers as the guide: informational (how to, what is), comparative (vs, alternative, best), and transactional (pricing, consultant, services, hire). Show the biggest clusters by impressions.
3. Build a "striking distance" list: queries with average position roughly 8 to 20 and meaningful impressions. For each, suggest which of my known pages (from the Source Inventory and site sources) most likely almost ranks, and tell me to confirm the exact page by clicking the query in Search Console. Updating a near-ranking page usually beats writing something new.
4. Identify rising themes: queries or clusters whose impressions are clearly growing across the period. Flag them as demand forming.
5. Compare the market's phrasing against my Agreed Understanding's vocabulary. Where the market consistently uses words for our problem space that our identity does not, record it: a strong repeated mismatch becomes a numbered challenge in the Conflict Register, continuing the existing numbering; a weak one goes on the Watch List with date added and date last seen.
6. Respect the Rulings Log, as always: ruled questions stay ruled unless this data is genuinely new evidence.
7. Output: the intent clusters, the striking-distance list with suggested pages, rising themes, and any challenges or watch-list items, in markdown, ready to append to brain.md as a "Demand Evidence" section.

My brain.md so far:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

My query export:
[PASTE ROWS: query, clicks, impressions, position]

Check it worked. Either you can name three queries where you are in striking distance and the page to update for each, which is the cheapest visibility gain available to you this quarter. Or you learned the honest alternative: your site is too young for search data to speak yet, and setting up Search Console today was still the right move.

Install the gate

What you get: A reusable pre-publish check for everything you ship

ten minutes per draft

Why

Serious publications separate the writer from the checker, and the checker verifies against source material, not against the writer’s confidence. Your Brain file makes the same discipline available to a company of any size: before anything ships, it gets checked against the agreed audience, the agreed voice, and the agreed claims. The gate is a checker, not a ghostwriter. It flags; you decide.

Do this

1 Save the prompt below somewhere you can reach in ten seconds. If you use Claude Projects, save it as the project instruction so the gate is always one paste away from running.
2 Before publishing anything that speaks for the company (article, page, deck, important email), run it through the gate with your current brain.md.
3 Fix what deserves fixing, overrule what does not. The gate advises; the human ships.
4 When the gate keeps flagging the same thing and you keep overruling it, that is not friction, that is data: either the identity needs amending or the habit does. Take it to the next monthly session as a challenge.

Hand it to Claude

You are the pre-publish gate for my company. My Brain file (brain.md) is available, and pasted below is a draft that is about to ship. Check the draft against the Agreed Understanding. You are a checker, not a rewriter.

Rules, which you must follow exactly:
1. Check three things, in order:
- Audience: who is this draft actually written for, judged from its content, vocabulary, and assumed context? Does that match the agreed ICP? Quote the phrases that reveal the audience.
- Voice: does it match the agreed voice and vocabulary? Flag banned words, off-voice constructions, and places where the Language Bank has a customer phrasing that is stronger than ours.
- Claims: list every factual claim the draft makes about us, our results, or our market. For each, mark it SUPPORTED (it matches the Agreed Understanding or cites its own evidence), UNPROVEN (asserted with no support anywhere), or CONTRADICTS (it conflicts with an agreed fact or a ruling; name which).
2. Give one verdict: PASS, PASS WITH NOTES, or HOLD, with HOLD reserved for a CONTRADICTS finding or an audience mismatch, and say in one sentence why.
3. Suggest specific line-level fixes only for what you flagged. Do not rewrite the piece, do not polish style beyond the flags, do not add content.
4. Stay independent of my feelings about the draft. If it is off-identity and well written, it is still off-identity, and pleasing me is not your success metric.

My brain.md:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

The draft:
[PASTE IT]

Check it worked. Run the gate on something you published last month. If it comes back PASS, your identity and your output are aligned, which is rarer than you think. If it comes back HOLD, you have just seen the gate catch a real one, retroactively, and you will not ship the next one blind.

Brief the company

What you get: A one-page briefing everyone who speaks for you can use

30 minutes, plus a walkthrough with the team

Why

What keeps a company consistent is a short ranked message hierarchy: the claims in priority order, proof under every claim, and governance so outdated versions stop circulating. Your Brain file now contains everything that page needs, already agreed and evidenced. The multiplier is distribution: the brain only changes what the company says if the people saying it carry the same page.

Do this

1 Run the prompt below with your brain.md available.
2 Send the briefing to everyone who writes, sells, or speaks for the company. Walk it through once; do not just attach it.
3 Use the kill list: the old decks and pages still carrying superseded claims. A stale version that looks official is worse than none.
4 Reissue the briefing whenever the Agreed Understanding changes version, and never let two versions circulate.

Hand it to Claude

You are producing the company briefing from my Brain file, which is available. This is the one-page artifact that everyone who writes, sells, or speaks for the company will carry.

Rules, which you must follow exactly:
1. Produce "Company Briefing vX.Y", where X.Y is the current Agreed Understanding version in my Brain file. It contains, in this order:
   - One umbrella statement: the single sentence version of who we are and who we serve, drawn from the Agreed Understanding, not newly invented.
   - The ranked value claims, each with its proof point in one line. Claims marked [UNPROVEN] in the brain do not appear; list them separately at the bottom as "not yet claimable" so nobody uses them by accident.
   - The agreed ICP in plain checkable terms.
   - Voice in ten seconds: five words we use, five we never use, drawn from the agreed vocabulary and the Language Bank.
2. Everything must trace to the Brain file. If the briefing needs something the brain does not contain, mark it [TODO] and tell me; do not fill the gap.
3. Then build the kill list: from the Source Inventory, identify every artifact that now contains claims superseded by rulings, and list them with what specifically is stale in each. These are the documents to retire or update.
4. Keep the briefing to one page of markdown. If it does not fit, cut the least-ranked material, never the proof points.
5. Output: the briefing, then the kill list.

My brain.md:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

Check it worked. Ask two colleagues, separately, who the company serves and what its top claim is; if their answers match the briefing, distribution worked. Solo founder? Read the umbrella statement against your own homepage and your last three sales emails. If they tell a different story than the briefing, the kill list has work to do.

Run the monthly loop

What you get: A versioned brain, a dated log, and a habit

one hour, monthly

Why

Product changes, competitors change, markets change; all three fire continuously, which is why the understanding needs a cadence, not an anniversary. But the counterweight matters just as much: positioning takes far longer than a quarter to establish, so the default outcome of a monthly session should be no change. The brain earns trust by mostly agreeing that you are who you said, and by having the receipts on the months it does not.

Do this

1 Book a recurring hour, monthly. This is the whole maintenance cost.
2 Bring whatever signals the month produced: a couple of call transcripts, the quarterly search export when fresh, notable wins and losses, anything a customer said that stuck.
3 Run the prompt below with your current brain.md available, rule on anything it raises, and update the changed sections in your file. Same habit as always: scroll the output and confirm nothing says "rest unchanged" before saving.
4 Expect most months to end with "no change". That is the system working, not failing.

Hand it to Claude

You are running the monthly review of my company Brain file. My current brain.md is available, and this month's new signals are pasted below.

Rules, which you must follow exactly:
1. Process the new signals exactly as the earlier phases did: verbatim extraction into the Language Bank where relevant, then comparison against the Agreed Understanding. Thresholds by evidence type: call and language themes need two instances before becoming a challenge; anything based on wins and losses follows the deal audit's three-deal rule. Below threshold, items go to the Watch List with date added and date last seen.
2. Check the three structural triggers and report plainly, one line each: has the product materially changed, have the competitors appearing in real deals changed, has the market context changed? Only evidence from the signals counts; no speculation.
3. Review the Watch List and the parked challenges. Watch List: promote items that gathered their second instance into challenges (continuing the permanent numbering), update date last seen on items that recurred, retire items whose date last seen is three months or more before today, and say which you did. Parked: if this month's signals contain the evidence a parked challenge named as settling it, present that challenge for ruling; otherwise report parked items still waiting, with what they wait for.
4. Present any challenges one at a time for my ruling, exactly per the adjudication rules: evidence both sides, no lean, wait for my answer, print the Rulings Log entry immediately after each ruling. Allowed outcomes: CONFIRM, REBUT, AMEND, or PARKED with the settling evidence named.
5. The default outcome is no change. Do not manufacture a challenge to seem useful. A month where you report "the understanding held; nothing to rule" is a successful month, and I want you to say exactly that sentence when it is true.
6. Housekeeping, once a year or when the file gets hard to navigate: move RESOLVED Conflict Register entries and Language Bank quotes older than 12 months into a separate archive section I will keep as brain-archive.md, leaving their Rulings Log summary lines in place. The Agreed Understanding, current Language Bank, and logs stay in brain.md.
7. Finish by outputting ONLY the changed sections, each complete and clearly titled, for me to paste over the matching sections of my file: the updated Agreed Understanding if any ruling changed it (increment the version: v1.1, v1.2...), new Rulings Log entries, the updated Watch List and Conflict Register, and one appended line for the "Review Log": date, signals reviewed, challenges raised, rulings made, version. Do not regenerate unchanged sections.

My current brain.md:
[PASTE IT, OR SKIP IF IT IS IN PROJECT KNOWLEDGE]

This month's signals:
[PASTE TRANSCRIPTS / EXPORTS / NOTES, OR "light month: nothing new" IF SO]

Check it worked. After three months you have a Review Log with three dated lines, a version number that moved only when something real changed, and at least one month that ended "the understanding held; nothing to rule". That log is the difference between a brain and a binder: it proves the understanding is alive and that it does not thrash.

What you have when you finish

One document. A Source Inventory with authority settled. An Agreed Understanding where every sentence was ruled on, not asserted. A Rulings Log that remembers why. A Language Bank in your customers’ own words. Demand evidence from your own search data. A gate that checks everything you ship, and a monthly habit that keeps the whole thing honest without letting it thrash. No platform, no subscription, no integration project. A file, a model, and your judgment, which was always the scarce part anyway.

The research this guide stands on is on The Company Brain page, and the wider operating model is in What Is a Cyborg CEO?