company-brain

/company-brain — Team-shared AI-ready knowledge base

Company Brain (n.): Your team's shared, AI-ready knowledge base — people, companies, meetings, SOPs, and decisions structured so Claude can answer questions on your team's behalf.

Team-scope sibling to second-brain (personal-scope). Same core compile → wiki → outputs pattern; different raw schema optimized for multi-author, sales-heavy, ops-heavy team use.

Also the operational backbone for the Company Brain Setup productized service (was previously called "Second Brain as a Service"; renamed to match the skill).

Mental model

Three layers, same as second-brain — but the raw/ layer is structured, not flat:

raw/          →  wiki/         →  outputs/
(structured      (compiled        (generated
 by category)     interlinked)     artifacts)

Structured raw/ dirs (each is its own top-level folder in the vault):

Dir	What lives here
people/	Contacts with context — CRM-lite. One markdown file per person.
companies/	Org profiles — last touchpoint, opportunity size, status. One file per company.
meetings/	Call/meeting transcripts + notes. Naming: YYYY-MM-DD-<company-or-topic>-<slug>.md. Auto-sync source.
sops/	Standard operating procedures. Named: <team>-<process>.md (e.g., sales-outbound-cadence.md).
decisions/	Decision records (narrative form; decide skill's structured form is different).
customer-language/	Verbatim phrases from prospects/customers/users. Fuels copy, headlines, objections.
recurring-questions/	Questions asked 3+ times across calls. Each becomes a pre-answered SOP/FAQ/script.
sales-objections/	Library of objections + best responses. Assembled into sales scripts.
raw/	Legacy / uncategorized captures (fallback bucket, minimize use).

wiki/, outputs/, and INDEX.md work the same as second-brain.

Reserved dirs (never modified by company-brain): Projects/, Team/, Templates/, Drafts/.

Multi-author discipline

Every capture stamps:

source: <URL / call / email / manual entry>
author: <who added this — email or handle>
captured: YYYY-MM-DD
trust: unreviewed

Wiki pages track cumulative contributions in the ## Sources section (per source file, per author). No overwriting — always append + attribute.

Sensitivity tagging (optional but recommended):

sensitivity: internal        # any team member can read
sensitivity: leadership      # exec team only
sensitivity: confidential    # named list only (list access in the file)

Default: internal. Query mode respects sensitivity — refuses to include confidential content unless the invoker is on the access list.

Trust levels

The other half of multi-author discipline: not everything captured deserves equal weight as context. Every structured-raw file carries a trust: field.

(The field is named trust, not status, because companies/ and decisions/ already use status: for lifecycle — prospect/customer, decided/reversed — and the two must not collide.)

Trust	Meaning	Query treatment
unreviewed	Captured but no human has confirmed it (default for every new capture)	Usable, but flagged — answers leaning on it note lower confidence
verified	A human reviewed it and confirmed it's right	Full weight
deprecated	Wrong or obsolete — kept for history only	Never used as context
superseded	Replaced by something newer — add superseded_by: [[target]]	Never used as context; queries point to the replacement

Deliberately an enum, not a numeric weight — teams keep a four-value field current; nobody maintains a 0–1 float.

Deprecation replaces deletion. The "never delete raw files" rule stays intact: when info turns out wrong or stale, mark it deprecated (or superseded with a pointer) instead of removing it. History is preserved; context is protected.

Trust is orthogonal to sensitivity — a file can be verified + confidential, or unreviewed + internal.

Existing vaults: files predating trust levels simply lack the trust: field — treat them as unreviewed. If the vault's schema doc (CLAUDE.md / AGENTS.md) predates trust levels, offer to add the trust spec to it on the first /cb review run (the vault's CLAUDE.md stays authoritative — extend it, don't override it).

Step 1 — Load vault config + schema

1. Read references/vault-config.md for the vault path (default: ${COMPANY_BRAIN_VAULT:-$HOME/Documents/CompanyBrain}/)
2. Read the vault's schema doc — <vault>/CLAUDE.md or <vault>/AGENTS.md, whichever exists (if both, read both) — for the authoritative team schema. If present, trust it over references/schema.md — the team's vault is the source of truth.
3. If neither exists, fall back to references/schema.md — the team schema starter kit.

Step 2 — Parse mode

Invocation	Mode
/cb capture / /company-brain capture / "capture this into the team brain"	capture
/cb compile / "compile the company wiki"	compile
/cb query <q> / "what does the team know about X"	query
/cb review / "review the company brain" / "cull the team brain"	review
/cb lint / "lint the company brain"	lint
/cb connect / "find cross-team connections"	connect
/cb search <term> / "search the company brain"	search

Step 3 — Run the mode

capture

Same intake mechanics as second-brain, but the routing is different — pick the structured dir based on content type.

1. Detect content type + route to the right dir:

   • Call/meeting transcript → meetings/YYYY-MM-DD-<company-or-topic>-<slug>.md
   • Person's LinkedIn / bio / contact context → people/<name-slug>.md
   • Company profile / prospect / client → companies/<company-slug>.md
   • Documented process / how-we-do-X → sops/<team>-<process>.md
   • Decision made by leadership / team → decisions/YYYY-MM-DD-<decision-slug>.md
   • Verbatim customer quote → customer-language/<theme-slug>.md (append to existing themed file if one exists)
   • Question asked in a call → recurring-questions/<question-slug>.md (append counter if repeat)
   • Sales objection heard → sales-objections/<objection-slug>.md (append variant if repeat)
   • If ambiguous, ask.
   • If the target dir doesn't exist yet, create it with this capture. New vaults start with only one pilot workflow's dirs (see references/vault-config.md → "Start narrow").

2. Add multi-author metadata (top of file):

   source: <URL / call with X on YYYY-MM-DD / email from Y / etc.>
   author: <who captured this>
   captured: YYYY-MM-DD
   trust: unreviewed      # every capture starts unreviewed — review mode promotes it
   sensitivity: internal  # or leadership / confidential
   

3. Save + report file path + one-line summary.

Don't compile into the wiki here — capture is fast intake.

compile

Same core pattern as second-brain's compile mode — process unprocessed structured-raw files into wiki pages, update INDEX.md, add Sources sections.

Differences from second-brain:

• Multi-author attribution: Sources section includes author, not just filename

  ## Sources
  - `people/jane-doe.md` (added by @alex, 2026-06-30) — CTO of Acme, evaluated us Q2
  

• Cross-category compilation: a wiki page on "Acme Corp deal" might pull from companies/acme.md, meetings/2026-06-15-acme-discovery.md, sales-objections/acme-pricing.md, and people/jane-doe.md — all into one wiki page.
• Sensitivity inheritance: wiki pages inherit the highest sensitivity of any source. If any source is confidential, the wiki page is confidential.
• Trust filtering: deprecated and superseded sources are excluded from wiki pages. If a source that already fed a wiki page later gets deprecated, recompile flags the affected pages for re-review and drops the source, noting it in Sources using the file's reviewed + reviewed_by stamps: - meetings/2026-06-15-x.md (deprecated 2026-07-01 by @alex). Pages built mostly from unreviewed sources get a > ⚠ Mostly unreviewed sources callout at the top.
• INDEX.md categories for teams: Sales, Customers, Ops, Product, Team & People, Decisions, Playbooks. Extend as needed.

Everything else (one-page-per-concept, [[wikilinks]], Connections mandatory, quality > quantity) is identical.

query

Same as second-brain query, plus:

• Sensitivity check first: identify the invoker; refuse to include content above their sensitivity level.
• Trust rules: prefer verified over unreviewed, and recent over old. Never use deprecated or superseded content as context — at most cite it as a pointer: "(deprecated — see [[replacement]])". When two sources conflict, prefer the newer + higher-status one AND surface the disagreement in the answer.
• Confidence flag: if the answer leans mostly on unreviewed sources, say so up front: "Low confidence — 3 of 4 sources are unreviewed. Run /cb review to firm these up."
• Author-aware answers: when citing, include who contributed the info: "Per [[Acme Deal]] (source: meetings/2026-06-15-acme-discovery.md by @alex)..."
• Route external gaps to deep-research, same as second-brain.

Save to outputs/<YYYY-MM-DD>-<question-slug>.md with the answer + wiki pages consulted + sensitivity level of the output.

review

The human culling pass. This is how a team keeps garbage-in from becoming garbage-context: everything gets captured freely (nothing is lost), but only reviewed info earns full weight.

0. Sensitivity check first — same rule as query mode: identify the invoker and exclude files above their sensitivity level from the queue. Report the exclusion count: "3 items above your sensitivity level were skipped — someone on the leadership list needs to review those."

1. Build the triage queue:

   • All trust: unreviewed files across the structured-raw dirs (including files with no trust: field at all), newest first
   • Everything lint flags (checks 1–12; check 13 is about review itself)
   • Files whose review dates have lapsed, where those fields exist: decisions/ files past review_by, sops/ files past last_reviewed + review_cadence

2. Walk the queue one item at a time. For each file show: one-line summary, source, author, captured date, and which wiki pages cite it. Offer four dispositions — every disposition except skip stamps reviewed: YYYY-MM-DD + reviewed_by: <handle>:

   • verify → trust: verified
   • deprecate → trust: deprecated (wrong or obsolete; kept for history)
   • supersede → trust: superseded + superseded_by: [[target]] (ask for the replacement)
   • skip → leave as-is, resurfaces next review

3. Batch-apply the frontmatter updates — don't rewrite file bodies, only the metadata block.

4. Flag downstream effects: if a deprecated/superseded file feeds existing wiki pages, list those pages and offer to recompile them now.

5. Close with a summary: "12 reviewed: 8 verified, 3 deprecated, 1 superseded. 2 wiki pages recompiled. Next review suggested: ." Save the summary to outputs/<YYYY-MM-DD>-review.md so the cull itself has an audit trail.

Cadence: weekly for active vaults; pair with loopify to schedule it so the cull actually happens instead of depending on someone remembering. A vault where reviews lapse >1 month shows up in lint (check 13).

lint

Same seven checks as second-brain PLUS:

8. Stale people/companies — people/ or companies/ file with no update in >6 months for active accounts
9. Recurring-questions above threshold — questions asked 5+ times without a wiki page or SOP
10. Objections without responses — sales-objections/ files with no linked response in sops/ or wiki/
11. SOP freshness — SOPs not touched in >12 months (may be stale as the business evolves)
12. Author load imbalance — one contributor doing >80% of captures (usually signals the vault is one-person-dependent — bad for team continuity)
13. Review backlog — >20 files sitting at trust: unreviewed, or no review pass (no outputs/*-review.md) in >1 month. Points at /cb review.

connect

Same as second-brain plus cross-category link suggestions — e.g., sales-objections/pricing-too-high.md should link to customer-language/willingness-to-pay.md and sops/discovery-call-cadence.md if they exist.

search

Same. Grep across all structured-raw dirs + wiki/.

Optional: auto-sync sources

Team vaults benefit from automated capture. See references/auto-sync-sources.md for the setup patterns:

Source	What it captures	Setup
Fathom / Gong / Granola	Call/meeting transcripts	Webhook → append to meetings/
Slack export	Team discussions worth preserving	Manual or scheduled export → raw/slack-<channel>-<date>.md
Email (Front / Missive / Superhuman)	Customer-facing threads worth preserving	Forward-to-address → append to people/ or companies/
CRM (HubSpot / Attio / Pipedrive)	Deal state, contact info	Periodic sync → companies/ + people/

Auto-sync is optional — most teams start with manual capture and add automation as the vault matures. Pair with loopify to schedule periodic sync jobs.

Multi-writer git sync (team members + remote agents)

A team vault is multi-writer by definition, and git is the coordination layer. Back the vault with a hosted remote (GitHub/GitLab); the remote then doubles as a capture API for agents without filesystem access — cloud agents, scheduled sync jobs, teammates' machines. Anything that can reach the git host's API (directly, or through an MCP integration layer like Executor) can read the wiki and commit captures into the structured raw dirs.

The discipline that keeps writers from diverging:

1. Every local session pulls before writing: git pull --rebase --autostash before vault work, push after committing. With multiple humans and agents committing, local copies go stale fast.
2. Obsidian users: the community Git plugin with auto-pull on an interval (~10 min) + pull-on-startup, auto-commit off — commits should stay semantic (one per capture/compile), not "vault backup" noise. Every team member's machine needs this, not just one.
3. Remote agents and auto-sync jobs commit append-mostly: new files in the structured dirs, descriptive commit messages, author stamped in the capture frontmatter (the multi-author trust model depends on it). Distinct-file appends make conflicts rare; rebase absorbs the rest.

Verify the loop once per machine when onboarding: remote commit via API → local pull → file appears.

Composes with

• second-brain — sibling. Use second-brain for your personal wiki; company-brain for the team's. A person can maintain both simultaneously with separate vault paths.
• skillify — use to author new skills that read from the company brain (e.g., a weekly-team-brief skill that queries company-brain every Monday).
• loopify — schedule auto-sync jobs (Fathom pull daily, Slack export weekly, review pass weekly, INDEX lint monthly).
• toolify — wire up integrations that feed the company brain (Fathom webhook receiver, Attio API, etc.).
• deep-research — when query finds gaps, route external. Save deep-research results into raw/ for future compilation.
• decide — decisions/ folder complements decide's structured archive. decide records the evaluation; decisions/ records the narrative + outcome + review notes.
• pm — team task management sits in Projects/ (reserved from company-brain). pm owns Projects/; company-brain reads it for context but doesn't modify.
• jab-hook — customer-language/ fuels social copy that resonates with actual prospect language.
• A blog-drafting skill (yours or a companion plugin) — pulls from customer-language/, recurring-questions/, and sops/ for authoritative blog drafts.

Sibling implementations (reference)

Same lineage as second-brain:

• Gbrain — Garry Tan's team-scale brain (146K pages, 24K people entities). Postgres/PGLite backed with graph traversal + scheduled maintenance. When a team's company-brain outgrows markdown-only, Gbrain is the upgrade path.
• Hermes' llm-wiki — reference for the 3-folder pattern.
• Notion AI / Glean / Mem — commercial "Company OS" tools. Company-brain is the Claude-native, markdown-first alternative — cheaper, more portable, better for teams that already live in Obsidian / Git-backed docs.

Notes on quality

• Structured raw > flat raw at team scale. Second-brain's type-prefix works for one person; teams need dedicated dirs for people/companies/meetings/etc. so multi-author search stays fast.
• Multi-author attribution is non-negotiable. Every file stamps author: and captured:. Wiki pages cite by source + author.
• Sensitivity is respected end-to-end. Query mode refuses to include content above the invoker's level. Wiki pages inherit the highest sensitivity of any source.
• Never delete raw files. Same rule as second-brain — the structured dirs are the source of truth. When info is wrong or stale, deprecate, don't delete — trust: deprecated removes it from context while preserving history.
• Capture freely, weight deliberately. The trust enum means dumping information in is safe — nothing unreviewed poisons answers at full weight, and /cb review is the regular cull that promotes or retires it.
• Start narrow, earn the right to expand. Seed one revenue-adjacent workflow's dirs, pilot with one or two people, and widen only once queries are getting answered and review is holding. See references/rollout.md.
• Never modify Projects/, Team/, Templates/, Drafts/ during company-brain operations.
• Auto-sync is optional. Start manual; automate as the vault matures. Don't burn cycles on Fathom webhooks before the team is capturing meetings regularly by hand.
• One person shouldn't be the whole vault. If lint flags author-load imbalance >80%, the team is one bus-factor away from losing the brain. Broaden contribution.