retrieval-judge

npx skills add majiayu000/claude-skill-registry --skill retrieval-judge --agent claude-code

Same command for any agent — swap --agent for codex, cursor, copilot.

# Retrieval Judge Skill When using `recall_search`, you MUST apply critical judgment to results rather than treating them as authoritative. This is a mandatory step after every `recall_search` call. ## Query Construction Build specific queries that include conversation context: - **Good**: `"payment retry logic after provider timeout"`, `"Go error wrapping with sentinel errors"` - **Bad**: `"retry"`, `"error handling"` Include the problem domain, technology, and specific concern in your query. ## Result Evaluation After each `recall_search` call, evaluate every result before using it: 1. **Title/type match** — Does the entry's title and type (pattern, decision, failure, etc.) relate to what you're actually looking for? 2. **Content relevance** — Does the snippet address your specific question, or is it tangentially related? 3. **Applicability** — Does the result apply to the current technology, codebase area, or problem domain? Only reference results that **directly address** the query. Discard results that are merely keyword-adjacent. ## Anti-Patterns - **Don't trust rank order blindly.** RRF scoring produces narrow distributions — result #1 may not be meaningfully better than res