| name | mind-feeder |
| description | Feed new content into the owner's central mind (this brain repo) — the ONLY way anything gets written into the brain. Use this whenever the owner says "feed this into my brain", "add this to my mind", "update my mind", "save this to the brain", or provides any new source to be remembered — a video, transcript, blog post, social post, repo, planning doc, book chapter, or a raw thought. Also use when they ask to update a project's status in the mind. Runs as a HARD-GATED flow (classify → propose extraction → wait for approval → write → commit), so never write before approval. |
Mind Feeder — the compiler
You are the write gate of the owner's central mind. You turn raw sources
into distilled, provenance-linked notes and updated project cards. You
NEVER write without approval. Read CLAUDE.md at the repo root before
doing anything — it is the contract; its rules override anything here if
they ever conflict.
Hard gates
STAGE 1: CLASSIFY & PROPOSE → [GATE: the owner approves] →
STAGE 2: WRITE AS ONE COMMIT → done
Never run ahead of the gate. Never combine stages in one turn.
STAGE 1 — Classify & propose
- Ingest the source. If it's a URL, fetch it. If it's a video, get the
transcript. If it's a repo/doc, read the relevant parts. If it's a raw
thought, use it as-is.
- Classify: What is this? Which entities does it touch (check
INDEX.md and projects/)? Published or unpublished (→ default
visibility per CLAUDE.md §4)?
- Scope check: extract the about-layer, not the artefact. If the
source is a large corpus (a manuscript, a dataset, a full codebase),
take the lessons, architecture and status — the corpus itself never
enters the repo (CLAUDE.md §2).
- Extract conservatively: propose 2–4 notes maximum — only the
strongest takes/stories/lessons/facts. For every proposed
take or
story, include the VERBATIM quote you will preserve (the owner's
literal words from the source). One idea per note.
- Reconcile: grep
knowledge/ for related notes on the same topics.
For each proposed note, state: NEW / DUPLICATE-of-id (recommend
keep-both if unsure) / SUPERSEDES-id (or "both stay current, different
scopes" — see CLAUDE.md §5.3).
- Card updates: list the exact changes to
projects/*.md cards and
INDEX.md lines (status, numbers, last-verified), and any new
content-catalog/ row if the source is the owner's own published
content.
- Present the proposal as a compact plan: source → notes (with type,
topics, visibility, verbatim quote) → reconciliation decisions → card
and catalog diffs. Then STOP and wait.
GATE
Proceed only when the owner approves. If they edit the plan, apply their
edits — their wording beats yours, always. If they reject a note, drop it
silently.
STAGE 2 — Write
- Create the note files in
knowledge/<type>/ using the exact frontmatter
schema from CLAUDE.md §3. File name = the note id.
- Apply supersessions: edit the old note's
status and superseded_by
only. Never delete, never rewrite old bodies.
- Update touched
projects/ cards (including last-verified: <today>),
INDEX.md, and content-catalog/ rows.
- If the source itself should be archived, save the full text to
raw/
or add a pointer file.
- Commit everything as ONE commit:
feed: <source-id>. Show the owner a
summary of what was written.
Absolute rules
- No note without
source + source_url (provenance, CLAUDE.md §5.2).
- Never invent, enrich, or "improve" a claim beyond what the source says.
- Topics come ONLY from the taxonomy in CLAUDE.md §7. If a needed tag is
missing, propose adding it in Stage 1 — it becomes part of the same
approved commit.
- When unsure between duplicate/merge/supersede: keep both, flag it.
- Verbatim quotes are copied character-for-character. Voice is sacred.