Brain Setup

Three tables and two directories. The brain-setup skill executes this document.

The knowledge itself lives in files, not tables. These tables are machinery: an index so freshness can be queried, a source list so re-ingest is possible, and a question log so gaps are visible.

Tables

Pages

A derived index, rebuilt by Curator from page frontmatter. The file always wins — when a row disagrees with a page, the row is wrong.

• path — single line text
• title — single line text
• area — single line text
• owner — single line text
• status — single select: draft, canonical, superseded, archived
• review_due — date
• last_verified — date
• supersedes — single line text
• source_ref — single line text
• open_questions — number
• indexed_at — date time

Sources

Where material comes from, and when it was last pulled.

• name — single line text
• kind — single select: upload, thread, meeting, url, repo, manual
• location — single line text
• owner — single line text
• cadence — single select: weekly, monthly, quarterly, on_change, never
• last_ingested_at — date time
• status — single select: active, paused, retired
• notes — multiline text

Questions

Every question asked of the brain, answered or not. The demand signal.

• question — single line text
• asked_by — single line text
• asked_at — date time
• area — single line text
• outcome — single select: answered, answered_stale, gap, contradiction
• citations — multiline text
• ask_count — number
• last_asked_at — date time
• page_ref — single line text
• status — single select: open, page_written, wont_document

One table covers answered and unanswered asks on purpose. "What could we not answer" matters, and so does "what do we answer over and over from a thread nobody wrote up" — both are pages waiting to be written, and ask_count carries both.

Directories

• /brain/inbox/ — raw material waiting to be filed. Swept, not triggered.
• /brain/archive/ — superseded pages, kept with a header naming what replaced them. History is not clutter.
• /brain/state/ — Curator's cursor.
• /brain/reports/ — dated sweep and health reports.

Checklist

1. Confirm the schema with the user before creating anything. One line per table; do not paste the whole spec back at them.
2. Ask what the company does, what areas knowledge falls into, and who owns each area. Three questions, not an interview.
3. Create the three tables with create_data_table.
4. Create /brain/inbox/README.md and /brain/archive/README.md with short headers explaining what belongs in each.
5. Rewrite /brain/TAXONOMY.md with the company's real areas, replacing the starter set.
6. Rewrite /brain/INTAKE.md with the real table and column ids so the requests in it actually work.
7. Write the company context into /SUBTERRANEAN.md from what you learned.
8. Tell the user the activation step, verbatim: open a session with Curator and say "activate".

Step 8 is a separate session because an agent may only configure its own automation unless it holds the team permission set — which this template deliberately does not grant to anyone.