Search Corpus

Table of Contents

• Workflow
• Query forms
• Output format
• Worked example
• Guardrails

Related skills: Called by intuition-builder (before generating framings), editor (before reviewing a draft to surface prior thinking), writer directly. Read-only — no writes.

Workflow

Search corpus for query:
- [ ] Step 1: Parse query → topic tag, keyword, analogy term, or author
- [ ] Step 2: Grep corpus/{seeds,drafts,published}/**/*.md
- [ ] Step 3: Rank matches by signal
- [ ] Step 4: Format top 10 as a ranked list with id, status, density, excerpt

Query forms

The skill accepts three query shapes:

1. Topic tag (e.g., dropout): grep frontmatter topics: for the tag; fallback to body keyword.
2. Freeform keyword (e.g., kv cache): grep body + title for the phrase.
3. Structured (e.g., {topics: [attention-mechanism], status: published}): direct filter.

Ranking

Rank by:

1. Exact tag match > freeform body match.
2. Density score descending.
3. Recency (created date descending).
4. Status priority: published > draft > seed > dead.

Exclude corpus/dead/ always unless query includes include_dead: true.

Output format

Top 10 matches, one per line:

N. {id} | {status} | density={score} | "{first-sentence excerpt, ≤120 chars}"

If >10 results, show 10 and append: ... N more matches — narrow the query.

If 0 results: No matches. Candidate related searches: {suggest 2-3 alternate topic tags}.

Worked example

Query: dropout

Matches:

1. 2026-04-21-dropout-as-ensemble-thinned-networks | seed | density=7 | "had a thought while running — dropout is secretly an ensemble method."
2. 2026-02-08-bagging-in-deep-nets | draft | density=6 | "bagging is the thing dropout is trying to be."
3. 2025-11-14-noise-as-regularization | published | density=5 | "adding noise at training time prevents the model from memorizing."

Query: KV cache (freeform)

Matches:

No matches. Candidate related searches: attention-mechanism, inference, context-engineering

Guardrails

1. Read-only. Never mutates seeds.
2. Never returns seeds from corpus/dead/ unless query explicitly opts in.
3. Excerpt is verbatim from body — never a summary.
4. If >10 matches, truncate and tell the caller.
5. If 0 matches, suggest related searches; do not invent matches.
6. Respect manual_edits: true — do not reveal private-looking content beyond first-sentence excerpt without caller explicitly requesting full seed read.

Quick reference

• Input: query string or structured filter.
• Output: top-10 ranked matches with id, status, density, excerpt.
• Read-only, no side effects.