← ClaudeAtlas

reviewlisted

Two focused scopes: (1) Reflection Gate — after Distributor routes to a skill, verify the model actually executed it (catches 'mentioned but skipped'); re-inject on miss. (2) Lifecycle Governance — review/merge/split/retire skills in INDEX.yaml. Horizontal; force-triggered by Distributor after every task and on lifecycle events.
AVA-2568/MY_SKILL · ★ 1 · AI & Automation · score 75
Install: claude install-skill AVA-2568/MY_SKILL
# Review (审查) Two scopes only. Everything else (code review, task recheck) is the platform's or the ecosystem's job — not ours. ## Scope 1: Reflection Gate (anti-skip) This is the third layer of the anti-hallucination triad: | Cause | Owner | |---|---| | Model never loaded skill description | Distributor (session-start injection) | | Model loaded but mis-routes | Distributor (keyword + fuzzy routing) | | **Model routed but skipped execution** | **Review (Reflection Gate)** ← this | ### Mechanism After Distributor routes a task to skill X and the model produces output: 1. Check whether the model actually **invoked** skill X (look for skill-specific behavior / artifacts in the output). 2. If yes → pass. No action. 3. If no (the model said "I'll use X" but then did something else, or silently dropped X) → **inject the reflection**: ``` [Reflection Gate] You were routed to skill <X> but there is no evidence you executed it. Either (a) execute <X> now, or (b) explain why <X> does not apply and what you did instead. Do not silently skip. ``` 4. The model must respond to the reflection before the task is considered complete. ### Why this is the only per-task gate Code Review and Task Recheck were removed because: - **Code Review** — superpowers and platform-native review tools cover this better. - **Task Recheck** — the model's own verification-before-completion handles this. The Reflection Gate survives because **no other layer catches "routed but skipped"*