Turn "finish—get criticized—revise" into a fixed workflow: the agent first submits a draft, then dispatches a review sub-agent that only looks at the final output—not its own reasoning process—to assign a score of 1–10 and list issues, then revises based on the feedback until the quality threshold is met.
skills.sh install count 106836 (+2750 / 24h)
Security Audit: Audited (no risks found)
Install: npx skills add https://github.com/2dmurali/review-loop-skill --skill review-loop
Rationale
- Default to at least 2 rounds, at most 4 rounds, and a passing score of 8/10; these can be adjusted in the prompt, e.g., "use review-loop, gate 9, max 5 loops"
- The review prompt template requires categorizing issues as Critical / Important / Minor, providing file:line references where possible, and explaining why changes are needed and what to change; report scores and key points to the user each round, and append a summary of the entire loop at the end
- Specify that the review model must not be weaker than the working model; if scores do not improve for 2 consecutive rounds and feedback remains identical, suggest switching to a stronger model, temporarily lowering the threshold, or breaking the task into smaller parts
- Specify when not to use it: simple edits, tasks where tests are the sole criterion, exploratory work prioritizing speed, and tasks without acceptance criteria
- skills.sh installs 2750 in 24 hours, cumulative 106836 (indicating attention only)
Usage
- npx skills add https://github.com/2dmurali/review-loop-skill --skill review-loop
- Add "use review-loop" to your request, or say "run review-loop on the file I just wrote" to review an existing file
- Requires a client capable of dispatching sub-agents; if unsupported, the author recommends opening a new conversation and pasting only the final output for scoring
Risks
- Audit result is pass; SKILL.md contains only textual workflow, no scripts, network requests, or keys, consistent with the content
- Each round requires running an additional review agent, doubling token consumption and time
- Scores are generated by the model; passing does not guarantee absence of defects; the repository contains only a SKILL.md, with no example outputs