skill-retrieval-gate

# Skill Retrieval Gate ## Goal Use this skill to decide whether the current task should query `memory_search` before following another skill or workflow. This skill is for **retrieval judgment**, not mandatory retrieval. ## Core rule Do **not** make every skill query memory first. Instead: 1. Judge whether the task depends on local knowledge or history 2. Retrieve only when that dependency is real 3. Load only a few high-signal results 4. Fall back immediately if retrieval is weak, unavailable, or unnecessary ## Example triggers This skill is especially useful for requests like: - continue the previous work on this project - check what we already documented before you proceed - use project memory first if this depends on earlier decisions - decide whether retrieval is worth it before following the skill - base this on existing notes instead of asking me again ## Workflow ### 1. Decide whether retrieval is needed Use [decision-flow](./references/decision-flow.md) when the request may depend on: - project history - prior decisions - local knowledge bases - user-specific preferences - previously organized notes ### 2. Judge the skill tier Use [skill-tiering](./references/skill-tiering.md) to classify the current skill or task into: - retrieval-first - retrieval-optional - retrieval-usually-skip ### 3. Build the query Use [query-construction](./references/query-construction.md) to build a compact query from: - task object - task type - key module, symptom, or entity ### 4. Keep the result set small Use [result-trimming](./references/result-trimming.md) to limit context expansion. Default rule: - fetch top 1-3 results first - only expand deeper when clearly needed ### 5. Fall back fast Use [fallback-rules](./references/fallback-rules.md) if retrieval is empty, noisy, low-confidence, unavailable, or unnecessary. ## Anti-patterns Avoid these mistakes: - forcing retrieval for every task - copying the entire user prompt into `memory_search` - expanding every hit just because it matched - dragging weak or stale snippets into later reasoning - treating retrieval failure as a blocker instead of falling back ## Output expectation After using this skill, the agent should be able to answer: - Should I call `memory_search` for this task? - What query should I use? - How many results should I keep? - Should I fall back to the original skill flow immediately?