reflection-review-loop

# Reflection Review Loop ## Overview Use a generator-critic loop inside the current task: produce a concrete draft, review it against explicit success criteria, fix only actionable issues, then verify the revised result. ## When to Use Use this skill: - Before final delivery of non-trivial code, docs, reports, plans, prompts, or workflows. - After implementing a fix but before claiming it is complete. - When the task has checkable requirements, known failure modes, or user-facing impact. - When an answer could be incomplete because context is large, ambiguous, or multi-step. Do not use it for one-line shell outputs, trivial translations, or read-only facts where normal verification is enough. ## Workflow 1. Define the success criteria in 3-7 bullets. - Include the user's explicit request. - Include project or local instructions that constrain the result. - Include verification evidence expected before completion. 2. Produce the first draft or implementation. - Keep the work scoped to the request. - Do not polish around unclear requirements before checking the core behavior. 3. Run a critic pass. - Check for requirement gaps. - Check for factual or code correctness issues. - Check for missing tests, lint, build, visual QA, or source verification. - Check for safety, data, secret, auth, irreversible, or security boundaries. - Check whether the final answer will tell the user what changed, what passed, and what remains unknown. 4. Convert critique into actions. - Fix concrete defects. - Drop speculative improvements. - If the critic finds a real blocker, stop and report it clearly. 5. Re-run the narrowest relevant verification. - Use `verification-loop` after meaningful file or workflow changes. - Use domain skills such as `pdf`, `webapp-testing`, or `documentation-lookup` when the artifact requires them. 6. Finalize only when the critic has no remaining actionable issue. ## Critic Prompt Use this checklist as the review lens: ```text Review this result against the user's request and local instructions. Find only actionable issues. Prioritize correctness, completeness, verification, safety boundaries, and user impact. Ignore style preferences unless they affect clarity or usability. Return: blocking issues, non-blocking issues, missing verification, and final answer risks. ``` ## Using Subagents If the user explicitly allowed subagents or parallel agents, a reviewer subagent can perform the critic pass. Otherwise, perform the critic pass locally. ## Common Mistakes - Treating reflection as extra prose instead of a defect-finding step. - Accepting vague critique such as "improve clarity" without a concrete change. - Running the loop forever. Two passes are usually enough unless verification exposes new failures. - Claiming completion before the revised result has been checked.