Query skills

Six skills turn a question into a scored recommendation.

Understand

Parses the user's request into a structured request.json. Extracts indication, sample type, modality, minimum N, budget, hard negatives. Classifies intent as access, commission, or mixed.

• Reads: user request text
• Writes: store/queries/{slug}/request.json
• Model: Sonnet

Source: .claude/skills/query/understand/SKILL.md

Search

Runs when the wiki is thin. Iterates PubMed and EuropePMC queries with deterministic coverage scoring and mechanical synonym rewriting. Each round persists to disk. Stops when coverage plateaus or a round limit is hit.

• Reads: request.json, wiki index
• Writes: store/queries/{slug}/search/round_{k}.json, coverage_{k}.json
• Model: runs inline in the orchestrator's context (not a subagent)

Source: .claude/skills/query/search/SKILL.md

Discover

Scans the wiki index for entities matching the request filters. Reads cohort entities that match. Produces a candidates list with a verdict: wiki_rich, wiki_partial, or wiki_empty.

• Reads: request.json, store/wiki/index/, matching entity files
• Writes: store/queries/{slug}/candidates.json
• Model: Sonnet

Source: .claude/skills/query/discover/SKILL.md

Score

Evaluates each candidate on three axes: Scale, Cost, Quality. Reads entity articles and references/pricing-data.md for cost evidence. Produces per-axis confidence ratings.

• Reads: candidates.json, entity articles, references/pricing-data.md
• Writes: store/queries/{slug}/scored_candidates.json
• Model: Sonnet

The three axes are independent. No composite score. The buyer chooses the weighting. See Scoring for the full axis specification.

Source: .claude/skills/query/score/SKILL.md

Deliver

Assembles the final outputs from scored candidates. Produces a human-readable recommendation and machine-readable listings for the web app.

• Reads: scored_candidates.json
• Writes: recommendation.md, listings.jsonl, delta.jsonl
• Model: Sonnet

Candidates are rendered in axis-confidence-weighted order. The most defensible candidate reads first. Every candidate gets all three axes side by side.

Source: .claude/skills/query/deliver/SKILL.md

Bounty format

Used by the bounty workflow. Takes a decided procurement chain and writes it as a wiki bundle entity under store/wiki/bundles/. Validates chain structure (2-7 links, per-link evidence tiers). Hook-gated.

Source: .claude/skills/query/bounty/format/SKILL.md