Brain Ingest

Turn scattered company material into brain-file updates. Works with any source the agent can read: files in incoming/, a knowledge-source folder, or whatever connectors this runtime has (email, Google Drive, SharePoint, OneDrive, Slack, Granola, and so on). Run it as often as you like. The brain grows a batch at a time.

Workflow

Step 1: Pick the sources

Read brain/sources.md first (create it from the format below if it's missing). It records what's already been mined and what's still waiting.

Ask what to pull from, if the user hasn't said. Offer the open items under "To check":

What should I pull from? A file or folder, a Drive or SharePoint location, an email search, or "everything in incoming/"?

Connector sources: use the connectors available in this session. If the user names a source with no connector, say so and suggest exporting the files to incoming/ instead. Don't guess at APIs.
Searches: narrow before reading. Search by topic, sender, folder, or date range, list the hits (title, owner or sender, date), and let the user confirm which to read. Never bulk-read a whole mailbox or drive.
Files: read PDFs, decks, docs, spreadsheets, and images with whatever file-reading ability the runtime has. For .pptx and .docx, run python3 .claude/skills/brain-ingest/scripts/office-text.py <file>. It prints slide text and speaker notes locally, at any file size, with no extra installs. If one can't be read, say which and move on.

If the user names a source already listed under "Ingested," say when it was read and offer to look only at what changed since.

Keep a list of what you read, with a locator for each: a file path, document URL, or email subject + sender + date. That locator becomes the source citation.

Step 2: Extract

For each source, pull out only what a brain file can use:

Product facts, numbers, customer names, funding, team → truth.md
Positioning, pillars, value props, ICP → positioning-and-messaging.md
Competitors, win/loss reasons → competitive.md
Buyer roles, pains, priorities → personas.md
Features and technical detail → capabilities.md
Customer scenarios and outcomes → use-cases.md
Verbatim buyer phrasing → audience-language.md
Sales motion, objections, stall points → customer-journey.md
Discovery questions → qualifying-questions.md
Industry trends and news → market-signals.md
Boilerplate, bios, pitches → tactical-assets.md
Colors, fonts, logo rules → brand-guide/brand-guide.md

Skip anything that fits nowhere. Leave out personal, HR, legal, and financial-detail content even when it's in the source.

Step 3: Compare against the brain

Read each target brain file before proposing changes. Sort every extracted item into one of:

New: fills a [FILL IN] or adds something absent
Confirms: matches what's there (just report the count)
Conflicts: disagrees with the brain or with another source
Stale?: the source is older than what the brain says, or looks superseded (old deck, draft, pre-rebrand)

Step 4: Propose, then write on approval

Show the proposal grouped by brain file:

truth.md
  + Team size: 140 employees
      source: Drive › Board deck Q3 (2026-08-12)
  ! Funding: brain says "$18M Series A", source says "$30M Series B"
      source: email "Series B close" from CFO (2026-09-02)

personas.md
  + New persona: IT Director — owns vendor consolidation
      source: SharePoint › Sales › Discovery guide.docx

Ask the user to approve, edit, or reject each item, or approve a file's items together. Ask them to resolve every conflict. Never resolve one yourself.

Then write only the approved items:

Replace the matching [FILL IN], or add to the right section in the file's existing format.
Put the source right after the fact as an HTML comment: . It stays out of rendered output and lets a later reader trace where the fact came from.
Mark anything the user approved but wasn't sure about with [VERIFY].

Step 5: Wrap up

Update brain/sources.md:

Add a row under "Ingested" for every source read, including ones that turned up nothing, so nobody reads them twice. Use the same locator as the citations.
Remove those sources from "To check."
Add new leads to "To check": documents a source referred to, folders you saw but didn't open, people named as owning a topic ("pricing questions go to the deal desk lead"). At a company where material is scattered, these leads are often the most useful thing a run produces.

Unlike other brain files, this log doesn't need approval for each line. Tell the user what you added.

Report what changed, file by file, and what was skipped. Then run /brain-health to show what's still empty, and suggest sources likely to fill it ("personas.md is still thin. Sales call recordings or a discovery guide would help").

Rules

Approval before every write. Same rule as the rest of /brain/. Sources are raw input, not verified claims.
Every fact gets a source. No source, no write.
Read-only on sources. Never edit, move, label, reply to, or delete anything in email, Drive, SharePoint, or a knowledge source.
Most recent and most authoritative wins, but the user decides. A signed-off deck beats a draft; a recent email beats an old one. Flag it, don't pick.
Don't copy sources wholesale. Extract facts and short phrases. Never paste whole documents or email threads into brain files.
Keep it private. Don't quote names or internal discussion from email in a proposal unless it's the fact itself.

Related Skills

brain-health: Shows which brain files need filling and what to gather for them
product-marketing: Builds positioning, personas, and competitive files once raw facts are in
tone-mapping: Use instead for voice-and-tone and voice-samples, which need writing samples, not facts
design-extract: Use instead for the brand guide when you have a live website
granola: Pulls meeting notes into incoming/ for ingest

Learnings

brain-ingest

SKILL.md.claude/skills/brain-ingest/SKILL.md