| name | query |
| description | Put this project's graph to work toward the deliverable — answer a grounded question, advance a work/ deliverable with citations, run gap analysis ("what's missing for this?"), or read the task and reading-queue views — and file genuinely new conclusions back. Use when the user asks something the brain should know, asks to move a deliverable forward, or asks what's open or unprocessed. |
query
Use the graph toward the work product, and file useful results back. Read CLAUDE.md first.
Steps
- Start from the index. Read
PROJECT.md (goal, work products, task_policy:) and
knowledge/index.md first — don't blind-scan. Pull the relevant notes and follow their links
one hop at a time; the edges are the point. Load knowledge/golden/ as authoritative constraints,
and read knowledge/templates/ (the types) to resolve names against them so the same concept
isn't reinvented.
- Answer, or advance the deliverable. Answer the question, or advance an unlocked deliverable in
work/, citing the notes you used (link to them). Treat golden as hard constraints — flag a
conflict rather than overriding it. Don't rewrite a locked: true file without asking.
- Gap analysis (when asked "what's missing?"). Read the target deliverable in
work/ and the
graph, and surface what the deliverable still needs and what to ingest next to fill it. Propose;
don't fabricate to fill the hole.
- Read the views (don't regenerate them here). For "what's unprocessed?" read the reading-queue
index-view; for "what's open under initiative Y / requested by Z?" traverse the task nodes and
their parent/requested_by edges — or just browse the folder tree, which mirrors the parent
chain (§3a). Secondary index.md nested-bullet outlines are other slices (read-only). (Regenerating
a view is distill's job, §12 / §8 — query only reads.)
- File back. When a genuinely new conclusion or connection emerges, add it as a typed note,
resolve and link it into the graph, and update
knowledge/index.md so it compounds next time.
- Be honest about gaps. If the graph can't answer well, say so plainly and suggest what to ingest
— don't fabricate. When a follow-up needs the user to choose (which deliverable to advance, which
reading of an ambiguous ask), use the structured AskUserQuestion prompt (
CLAUDE.md §17).
Don'ts
Don't answer from memory when the graph should know. Don't regenerate a timeline or index-view here
(that's distill). Don't overwrite golden or locked files. Don't drop a frontmatter key you didn't
write. Don't make a claim the notes don't support without flagging it.
