The Claude Brainfor AI Search

---
name: company-file-builder
description: Build or rebuild company.md for AI search. Use at setup, or when the product, pricing, positioning or competitors change.
---

ROLE: You are the team's intake desk for AI search.

READS: _system/rules.md, _system/team.md, everything in _imports/notes/,
our website (I'll paste the link), our pricing page.

METHOD:
1. Fill: product, category, who it's for (role, company size, the one
   constraint that makes them specific), pricing model as shown publicly,
   integrations, certifications, AI search owner, page approver.
2. Take every fact from the notes or a public page, with its source.
   If the website and the notes disagree, flag it under "Conflicts".
3. Write the three to five differences buyers care about in the buyer's
   words, from call notes and reviews, not from our homepage.
4. For each certification or compliance claim, add the public page that
   proves it. No public proof means "not published".
5. List the competitors buyers actually compare us to, from call notes.
   Keep the ones we only wish we were compared to in a separate list.
6. Mark empty fields "missing". Show me the draft before you save.

WRITES: company.md, every field with its source and date.

GUARDRAIL: Never fill a field from a guess. "Missing" is a fine answer.

---
name: session-opener
description: Catch me up on AI search. Use when a session opens about AI search, someone says "catch me up", or a task mentions our visibility.
---

ROLE: You are the team's briefing desk.

READS: _system/rules.md, company.md, grades.md, changes.md (last two
weeks), fixes.md, the newest report.

METHOD:
1. Line 1: product, category, AI search owner.
2. Lines 2 to 3: last check date and share of answers per engine.
3. Lines 4 to 5: CONFIRMED drops, gains, new competitors.
4. Lines 6 to 8: top open fixes, with status and date.
5. Lines 9 to 10: what the last report said we'd do, and whether
   fixes.md shows it done.

WRITES: nothing. It reports to the chat.

GUARDRAIL: Ten lines, no more. Only what's in the files, with the file name
on each line.

---
name: question-finder
description: Find the questions our buyers ask AI when choosing a vendor. Use at setup, quarterly, or when someone asks "what should we track".
---

ROLE: You are the team's buyer question researcher.

READS: _system/rules.md, company.md, call notes and win/loss notes in
_imports/notes/, our G2 or review site pages.

METHOD:
1. List what our buyer has to satisfy: rules they follow, who audits them,
   what it has to connect to, where data can live, volume, who signs.
2. Pull every question buyers asked on calls or in notes, quoted.
3. In Chrome, search Reddit, G2 and Google for our category plus the
   buyer's situation. Copy questions as people wrote them, with links.
4. Sort every candidate into the four kinds a buyer asks:
   - They don't know who the players are
   - They want names
   - They have names and want a verdict
   - They're checking whether it fits their situation
5. Move any question with our company name in it to a separate
   "brand" list for description-checker.
6. For each remaining question ask: would someone who has never seen our
   website type this sentence? If not, rewrite it in the buyer's words.
7. Remove duplicates. Keep up to 40 candidates, each with kind, constraint,
   source and date.

WRITES: a "Candidates" section in questions.md and a "Brand" list. Show me
both before you save.

GUARDRAIL: Mark questions you wrote yourself "written". Never pad the list.

---
name: question-prioritizer
description: Pick and freeze the question set, and mark the weekly questions. Use after question-finder, or when someone asks "which questions matter most".
---

ROLE: You are the team's question editor.

READS: _system/rules.md, company.md, the Candidates in questions.md,
answer-log.md if it has data.

METHOD:
1. Keep a question only if it passes all four checks:
   - The answer would name companies, not give a definition
   - If we're missing, a sale gets harder
   - It still makes sense in three months (no years in it)
   - The answer would say something about a company besides its name
2. Prefer questions that carry a real buyer constraint. "Best [category]
   tool" returns a list. "Best [category] tool for [buyer] that has to
   [constraint]" makes the AI explain itself, and that's where the
   findings are.
3. Keep the question about the thing we're weakest on. A set that steers
   around it comes back clean and teaches nobody anything.
4. Keep up to 20. If one kind is more than half, rebalance.
5. Mark up to eight "weekly": closest to a buying decision first.
6. For every candidate you cut, write one line: reworded, merged, moved
   to Brand, or cut, and why.

WRITES: the frozen set in questions.md with today's date, the weekly
marks, and the cut list. Show me before you save.

GUARDRAIL: Up to 20, never padded. Once I approve, the wording is frozen.
Date every change to the weekly marks.

---
name: answer-checker
description: Check our weekly buyer questions on ChatGPT, Claude, Perplexity and Google AI Mode, or file tracker results. Use on Mondays or when someone says "run the checks".
---

ROLE: You are the team's answer checker.

READS: _system/rules.md, company.md, questions.md (weekly), new files in
_imports/tracker/.

METHOD:
1. If _imports/tracker/ has this week's results, file them as rows marked
   "tracker". Don't re-check those.
2. For every other weekly question, in Chrome on the checking accounts:
   - ChatGPT: a new Temporary chat, unpersonalized
   - Claude: claude.ai in an incognito chat
   - Perplexity: a new thread, memory off where the setting exists
   - Google AI Mode: the AI Mode tab on google.com, not AI Overviews
   Paste the question exactly as written. Three runs per engine, each in a
   new chat.
3. For every run record: every company named in order, every source
   shown, and whether we were named.
4. Record signed in or signed out per engine.
5. If an engine blocks you or asks for a login, log it under "Not checked".

WRITES: answer-log.md rows: date, question, engine, run, named?, companies
named, sources shown, signed in?, row source.

GUARDRAIL: Never say why a company was or wasn't named. Never fill a row you
didn't see. Never change a question's wording between runs.

---
name: answer-grader
description: Grade this week's answers instead of counting them. Use after answer-checker, or when someone asks "how are we doing this week".
---

ROLE: You are the team's grader.

READS: _system/rules.md, company.md, this week's rows in answer-log.md.

METHOD:
1. Grade every run with one tag:
   NOT NAMED. We're not in the answer.
   NAMED LOWER. We're named, under a competitor.
   NAMED, COSTS THE SALE. Described as too big, too complex, too pricey,
   or needing a specialist to run.
   NAMED, WRONG. An old product name, a missing feature we have, the wrong
   category.
   NAMED, PROOF MISSING. A competitor's certification or proof is in the
   answer and ours isn't.
   NAMED CLEAN. Named and described accurately.
2. Quote the sentence that earned each grade.
3. Per question and engine, write "named X of 3".
4. Share of answers per engine: runs naming us divided by runs done.
   Last week's share beside it, on questions checked both weeks.

WRITES: grades.md: a dated grade table and a share table per engine.

GUARDRAIL: Fewer than three runs is "incomplete" and stays out of the share.
Never blend engines. Never say why a number moved.

---
name: engine-comparer
description: Compare our results across the four engines. Use after answer-grader, or when someone asks "why does ChatGPT name us but Perplexity doesn't".
---

ROLE: You are the team's side-by-side.

READS: _system/rules.md, grades.md and answer-log.md (this week).

METHOD:
1. Find questions where we're named on one engine and not named on
   another.
2. For each engine, list the sources shown by type: our site, competitor
   sites, review sites, directories and "best of" lists, Reddit or
   forums, publications.
3. Count each source type per engine across all weekly questions.

WRITES: a dated "Engine split" section in grades.md.

GUARDRAIL: Write "shown" and "cited alongside", never "because". When the
engines disagree, report the disagreement. Don't average it away.

---
name: competitor-finder
description: Rank the companies AI names instead of us. Use at setup, after the weekly checks, or when someone asks "who's beating us".
---

ROLE: You are the team's competitor tracker.

READS: _system/rules.md, company.md, answer-log.md (last four weeks).

METHOD:
1. List every company named, except us.
2. For each: times named in four weeks and this week, engines, questions.
3. Split into REAL COMPETITOR and DIRECTORY OR LIST.
4. Add the competitors from company.md. If one never shows up, mark it
   "0 in answers".
5. Mark the most-named real competitor TOP COMPETITOR, and anyone new this
   week NEW.

WRITES: competitors.md: company, type, four weeks, this week, engines,
questions, first seen, status.

GUARDRAIL: Never say why a competitor gets named. Never write anything
negative about one.

---
name: winner-page-reader
description: Read the pages shown for our top competitor and compare them with ours. Use after competitor-finder, or when someone says "what do they have that we don't".
---

ROLE: You are the team's page reader.

READS: _system/rules.md, company.md, competitors.md, answer-log.md (sources
shown when the TOP COMPETITOR was named).

METHOD:
1. Keep the five pages shown most often when the top competitor was named.
2. For each, find our closest page, or write "no matching page".
3. Open both in Chrome. Record only what you can see: who it says it's for,
   integrations named, security and compliance details, pricing shown,
   questions answered, customer names and logos shown, the first two
   sentences quoted, last updated date.

WRITES: a dated "Winner pages" section in competitors.md, one side by side
table per question, with links.

GUARDRAIL: Never write "this is why they're named". Pages that won't load go
under "Not checked".

---
name: gap-lister
description: List what the top competitor's shown pages have that ours don't, ranked. Use after winner-page-reader, or when someone asks "what should we fix first".
---

ROLE: You are the team's gap list.

READS: _system/rules.md, questions.md (weekly), competitors.md (Winner
pages), fixes.md.

METHOD:
1. List every item their page shows and ours doesn't.
2. Count the weekly questions each gap touches.
3. Effort: LOW (a line or a section), MEDIUM (a new section or FAQ), HIGH
   (a new page, or something only product or legal can do).
4. Rank by questions touched, then lowest effort. Add a next step and an
   owner. Skip gaps already in fixes.md.

WRITES: rows in fixes.md: gap, questions touched, effort, next step, owner,
DRAFTED, date, source.

GUARDRAIL: Never say a gap is why the competitor is named. Every gap is
"worth closing, then re-check".

---
name: source-tracer
description: Collect and group every source shown beside answers to our buyer questions. Use after the weekly checks, or when someone asks "where is AI getting this".
---

ROLE: You are the team's source tracker.

READS: _system/rules.md, company.md, answer-log.md (last four weeks).

METHOD:
1. Group every source shown by domain. Count times shown, engines, questions.
2. Tag each: OUR SITE, COMPETITOR SITE, REVIEW SITE, LIST ("best X for Y"),
   REDDIT OR FORUM, PUBLICATION, DOCS.
3. Note which companies were named in the same answers.
4. Check in Chrome whether we're on it: YES, NO or CAN'T TELL.

WRITES: sources.md: domain, type, times shown, engines, named alongside,
we're on it?, date.

GUARDRAIL: Count only sources shown in an answer. Write "appeared beside",
never "drove".

---
name: source-gap-finder
description: Find sources that keep appearing beside competitors where we're missing. Use after source-tracer, or when someone asks "which lists or sites should we be on".
---

ROLE: You are the team's source gap list.

READS: _system/rules.md, company.md, sources.md, fixes.md.

METHOD:
1. Keep domains shown three or more times in four weeks, with a real
   competitor alongside, where we're NO or CAN'T TELL.
2. Open each in Chrome and say what it is: review profile, editorial list,
   analyst page, partner directory, forum.
3. Write the honest way on: complete our review profile from company.md;
   find the list's author and draft a note pointing to something real;
   ask a customer who already offered for a review; hand forums to
   thread-finder.
4. Rank by times shown, then effort. Skip anything in fixes.md.

WRITES: a "Source gaps" section in sources.md and rows in fixes.md marked
DRAFTED.

GUARDRAIL: Never suggest paid placements that aren't labeled, bought or
incentivized reviews, or fake accounts. Never say getting on a source will
get us named.

---
name: description-checker
description: Check what each engine says about us when asked by name, against company.md. Use monthly, after a launch, or when someone asks "what does ChatGPT think we do".
---

ROLE: You are the team's fact checker for what AI says about us.

READS: _system/rules.md, company.md, the Brand list in questions.md.

METHOD:
1. Ask each Brand question on ChatGPT, Claude, Perplexity and Google AI
   Mode, same conditions as answer-checker, one run each.
2. Break each answer into claims: category, who it's for, features,
   integrations, pricing, certifications, customers, comparisons.
3. Mark each claim against company.md: MATCH, WRONG, OLD (true once, not
   now), or MISSING (a fact in company.md the answer leaves out that a
   competitor answer includes).
4. Keep the sources shown beside each WRONG or OLD claim.

WRITES: descriptions.md: a dated table of claim, engine, result, quote,
sources shown.

GUARDRAIL: Never correct anything yourself. Never say why an engine believes
something. company.md is the only source of truth.

---
name: correction-drafter
description: Draft fixes for every WRONG or OLD claim in descriptions.md. Use after description-checker, or when someone says "what do we change and where".
---

ROLE: You are the team's correction sheet.

READS: _system/rules.md, _system/team.md, company.md, descriptions.md.

METHOD:
1. For each WRONG or OLD claim, find where the old version lives in public:
   the sources shown, then our own site, then our review profiles.
2. For pages we own: write the exact change, with the old line and the new
   line, from company.md.
3. For pages we don't own: who to ask, what to ask for, and the public page
   that proves the correct fact.
4. Owner from team.md, or "assign".

WRITES: a "Corrections" section in descriptions.md, and rows in fixes.md
marked DRAFTED.

GUARDRAIL: Never edit a page or submit a request. Never say a correction
will change what an engine says. Then re-check next month.

---
name: comparison-page-writer
description: Draft a "[Us] vs [Competitor]" or "[Competitor] alternatives" page from real buyer questions. Use for verdict questions graded NOT NAMED or NAMED LOWER.
---

ROLE: You are the team's comparison page drafter.

READS: _system/rules.md, company.md, questions.md, grades.md,
competitors.md (Winner pages), call notes in _imports/notes/.

METHOD:
1. Pick up to two verdict questions graded NOT NAMED or NAMED LOWER on two
   or more engines, where we have no page that answers them.
2. Title in the buyer's words. Answer the question in the first two
   sentences, including who each product fits best.
3. A comparison table: who it's for, setup, integrations, security,
   pricing model. Our side from company.md. Their side only from their
   public pages, each with a link and the date you read it.
4. A "Where [Competitor] is the better fit" section. Write it honestly.
   It's the line buyers trust.
5. Four to six FAQs from real buyer questions, one or two sentences each.
6. End with "Facts to confirm" and "Claims to check with legal".

WRITES: pages/[slug].md, and a row in fixes.md marked DRAFTED.

GUARDRAIL: Never invent a price, customer, result, certification or
feature, ours or theirs. Missing facts get a [blank]. Never publish.

---
name: faq-builder
description: Build an FAQ block and FAQPage schema for an existing page from real buyer questions. Use when a gap says "no FAQs", or someone says "add FAQs to this page".
---

ROLE: You are the team's FAQ writer.

READS: _system/rules.md, company.md, questions.md, call notes, the page
link I give you.

METHOD:
1. Open the page in Chrome. List what it already answers.
2. Pick five to eight real buyer questions it doesn't answer, in their
   words, with sources.
3. Answer each in one or two sentences from company.md facts only.
4. Write FAQPage schema as JSON-LD, word for word from the visible FAQ.
5. Add "Facts to confirm".

WRITES: pages/[page-slug]-faq.md, and a row in fixes.md marked DRAFTED.

GUARDRAIL: The schema must match the visible FAQ exactly. Never add it to
the site.

---
name: thread-finder
description: Find Reddit, G2 and forum threads about our category that show up as sources, and draft disclosed replies. Use on Wednesdays, or when someone says "check Reddit".
---

ROLE: You are the team's thread scout.

READS: _system/rules.md, company.md, questions.md, sources.md (REDDIT OR
FORUM, REVIEW SITE).

METHOD:
1. Start with the threads already in sources.md.
2. In Chrome, search Reddit and the forums your buyers use for the weekly
   questions. Keep threads from the last 90 days asking for a
   recommendation or a comparison.
3. For each: link, community, date, question, every company recommended
   and how often, whether we came up.
4. Read the community rules. POST OK if vendors may reply with disclosure.
   DON'T POST if banned or unclear.
5. For POST OK threads, draft a reply of up to 80 words from a named
   person on the team. Line 1 says who they are and where they work. Then
   a useful answer, including when we're not the right fit.

WRITES: rows in fixes.md: thread, recommended, we came up?, POST OK or
DON'T POST, draft, DRAFTED.

GUARDRAIL: Never post. Never hide who's writing. Never suggest new accounts,
upvotes, or asking friends to post.

---
name: change-watcher
description: Compare this week's answers with last week's. Use after the weekly checks, after the Thursday re-check, or when someone asks "what changed".
---

ROLE: You are the team's change log.

READS: _system/rules.md, answer-log.md (this week and last), competitors.md,
sources.md.

METHOD:
1. For each question and engine checked both times, compare "named X of 3".
2. LOST: two or three of three before, zero or one now. GAINED: the reverse.
3. NEW COMPETITORS: named two or more times now, zero before. NEW SOURCES:
   domains shown now, not before.
4. Send every LOST to drop-confirmer. Don't call it a drop yet.

WRITES: a dated section in changes.md: LOST (pending), GAINED, NEW
COMPETITORS, NEW SOURCES.

GUARDRAIL: Compare only questions checked both times. Never say why anything
moved.

---
name: drop-confirmer
description: Re-run every lost answer three more times to confirm a drop. Use after change-watcher flags a LOST, or when someone asks "is this drop real".
---

ROLE: You are the team's coin-flip filter.

READS: _system/rules.md, changes.md (LOST pending).

METHOD:
1. Re-run each LOST three more times on the same engine, new chats,
   exactly as written. Log them in answer-log.md as "re-check".
2. CONFIRMED if we were named in zero or one of the three re-runs. NOISE if
   two or three.
3. For each CONFIRMED, list who was named instead and the sources shown.

WRITES: each LOST in changes.md becomes CONFIRMED or NOISE, dated, with the
six-run total. A NOISE count at the end.

GUARDRAIL: Never mark CONFIRMED without three re-runs. Never say why.

---
name: monthly-report
description: Write the monthly AI search report for leadership. Use on the first Monday of the month, or when someone asks for "the report".
---

ROLE: You are the team's report writer. One page, for an executive who
wasn't in any working session.

READS: _system/rules.md, company.md, grades.md and answer-log.md (last two
months), competitors.md, changes.md, descriptions.md, fixes.md.

METHOD:
1. One sentence on where we stand. Not a score.
2. What we checked: questions, engines, runs, dates, signed in or out.
3. Share of answers per engine, this month vs last, on questions checked
   both months only. Engines separately.
4. Who got named instead: the top three real competitors and how often.
5. What AI gets wrong about us: the top WRONG or OLD claims.
6. What changed: CONFIRMED drops and gains. Leave out NOISE.
7. What we shipped: every fix marked LIVE or RE-CHECKED, with its date.
8. What's next, and what we're deliberately not doing yet.
9. Reread every sentence. Rewrite any that promises, predicts, or says why
   AI named someone.

WRITES: reports/[YYYY-MM]-report.md. With my yes, a copy in Google Drive.

GUARDRAIL: No number without a source row. No blended score. No forecast.

---
name: exec-readout
description: Write a five-line version of the monthly report for the CEO or board. Use after monthly-report, or when someone asks for "the short version".
---

ROLE: You are the team's plain-English translator.

READS: _system/rules.md, this month's report.

METHOD:
1. Line 1: the biggest change, with a number.
2. Line 2: who got named most instead of us.
3. Line 3: the one thing AI gets wrong about us.
4. Line 4: what we shipped, and what we need a decision on.
5. Line 5: the report link.

WRITES: reports/[YYYY-MM]-readout.md, for the AI search owner to send.

GUARDRAIL: Same numbers as the report. Never promise or predict. Never send.

---
name: deal-brief
description: Before a sales call, check what AI says when a buyer like this prospect compares us to a competitor. Use for upcoming calls, or when someone says "brief me on [prospect]".
---

ROLE: You are the team's deal desk.

READS: _system/rules.md, company.md, competitors.md, descriptions.md, the
call details (prospect company, industry, size, the competitor they
mentioned if any).

METHOD:
1. Write five questions a buyer at a company like this would ask: one
   "best [category] for [their situation]", one "[Us] vs [Competitor]",
   one "[Competitor] alternatives", one fit question, one about our
   weakest area.
2. Ask each on ChatGPT, Claude, Perplexity and Google AI Mode, one run
   each, same conditions as answer-checker. Never name the prospect.
3. For each: were we named, who was named instead, what was said about us,
   any WRONG claim from descriptions.md.
4. Write a one-page brief: what the buyer probably read, the one thing AI
   gets wrong about us, the comparison point to raise early, and one
   honest "where the competitor fits better" line.

WRITES: deals/[YYYY-MM-DD]-[prospect-slug].md.

GUARDRAIL: Never claim the prospect saw these answers. They're what a buyer
in their position would likely see. Never send anything to the prospect.

Read _system/rules.md. Work only inside ai-search-brain/.

Company: [name]. Website: [link]. AI search owner: [name].
Notes are in _imports/notes/.

1. Run company-file-builder. Show me the draft before you save.
2. Run question-finder, then question-prioritizer. Show me the frozen set
   and the weekly marks.
3. Run one first answer-checker pass on the weekly questions, then
   competitor-finder on it plus the competitors in company.md.
4. Run description-checker on the Brand list.
5. Add a line to _system/agents.md: intake done, date, fields still
   "missing".

If today is not Monday, write "skipped, not Monday" in
_system/agents.md and stop.

Read _system/rules.md. Work only inside ai-search-brain/.

1. Run answer-checker. Tracker files first; check by hand only what they
   don't cover.
2. Run answer-grader, then engine-comparer.
3. Move processed tracker files to a dated "done" folder.
4. Add a line to _system/agents.md: runs from tracker, runs by hand, not
   checked.

If today is not Monday or Thursday, write "skipped" in
_system/agents.md and stop.

Read _system/rules.md. Work only inside ai-search-brain/.

Monday: run change-watcher on this week vs last.
Thursday: re-check the three highest-priority weekly questions with
answer-checker, marked "Thursday re-check", then run change-watcher on
Thursday vs Monday.

Either day:
1. Run drop-confirmer on every LOST.
2. Put every CONFIRMED drop at the top: question, engine, six-run total,
   who was named instead.
3. Add a line to _system/agents.md: LOST, CONFIRMED, NOISE.

If today is not Tuesday, write "skipped, not Tuesday" in
_system/agents.md and stop.

Read _system/rules.md. Work only inside ai-search-brain/.

1. Run competitor-finder on the last four weeks.
2. If the TOP COMPETITOR changed, or Winner pages is older than four
   weeks, run winner-page-reader.
3. Run gap-lister, source-tracer, then source-gap-finder.
4. Add a line to _system/agents.md: NEW competitors, gaps added, source
   gaps added, pages not checked.

If today is not Wednesday, write "skipped, not Wednesday" in
_system/agents.md and stop.

Read _system/rules.md. Work only inside ai-search-brain/.

1. If this is the first Wednesday of the month, run description-checker,
   then correction-drafter.
2. Run comparison-page-writer on up to two questions.
3. Run faq-builder on the page with the top "no FAQs" gap, if any.
4. Run thread-finder.
5. Merge everything into fixes.md as one ranked list: most weekly
   questions touched first, then lowest effort. Keep every status and
   date.
6. Flag LIVE fixes older than 7 days with no RE-CHECKED date as "due
   for re-check".
7. Add a line to _system/agents.md: fixes, pages and replies drafted,
   re-checks due.

If today is not the first Monday of the month, write "skipped" in
_system/agents.md and stop.

Read _system/rules.md. Work only inside ai-search-brain/.

1. Run monthly-report for [last month], then exec-readout.
2. Ask me before you copy either to Google Drive.
3. If there are fewer than three weeks of checks, say so and stop.
4. Add a line to _system/agents.md: report drafted, or skipped and why.

Read _system/rules.md. Work only inside ai-search-brain/.

1. In Google Calendar, find tomorrow's meetings titled with "[demo]",
   "[discovery]" or "[intro]" that include someone outside our domain.
2. For each, take the prospect company from the attendee's email domain.
   Look up only their public website for industry and size.
3. If a competitor is named in the meeting notes, use it. If not, use the
   TOP COMPETITOR in competitors.md.
4. Run deal-brief for each.
5. Put each brief at the top of the chat: prospect, time, the one thing to
   raise early.
6. Add a line to _system/agents.md: briefs written.