How to use the Auto Review Loop agent skill (auto-review-loop)

What the auto-review-loop skill makes the agent do

The skill runs an autonomous review cycle on a research project. Each round has three parts: the agent sends the project context to Gemini through the gemini-review MCP bridge, reads the critique and score, and implements the suggested fixes. Then it asks the reviewer again.

The loop stops when the reviewer's verdict is 'ready' or 'almost' and the score is at least 6/10, or when 4 rounds have run, whichever comes first. The source says both conditions must hold for an early stop, so a verdict of 'not ready' never ends the loop early.

• Round limit: MAX_ROUNDS = 4.
• Stop condition: score >= 6/10 AND verdict is 'ready' or 'almost'.
• Log file: review-stage/AUTO_REVIEW.md, a cumulative record of rounds.
• State file: review-stage/REVIEW_STATE.json, used to resume after context compaction.

Install the auto-review-loop skill

Download the skill ZIP from this page and unzip it into one of these locations: ~/.claude/skills/ for all your projects, or a project's .claude/skills/ folder for one project only.

Alternatively, in the Claude apps go to Settings, then Capabilities, then Skills, and upload the ZIP there.

If you use Codex, install the skills/skills-codex/* package first, then install this one. This skill is written as an override so that Gemini, not a second Codex agent, acts as the reviewer.

• Claude Code, all projects: unzip into ~/.claude/skills/
• Claude Code, one project: unzip into .claude/skills/ inside the project
• Claude apps: Settings → Capabilities → Skills → upload the ZIP
• Codex users: install skills/skills-codex/* before this package

Set up the gemini-review MCP bridge

The review step cannot run without the gemini-review MCP bridge installed and configured. The skill calls its tools (mcp__gemini-review__review_start, review_reply_start and review_status). If the bridge is missing, the loop cannot perform its review step.

The skill can use a specific Gemini model if you set the optional GEMINI_REVIEW_MODEL environment variable. If you leave it unset, the source does not name a default model, so do not assume one.

• Required: a gemini-review MCP bridge providing the mcp__gemini-review tools.
• Optional: GEMINI_REVIEW_MODEL to choose a Gemini model.
• The source does not give bridge installation commands; follow the bridge's own documentation.

Trigger the skill with natural requests

You can ask for the loop in plain language. The skill's description lists these trigger phrases: 'auto review loop', 'review until it passes', or a request for autonomous iterative improvement.

You can also invoke it with a command and a topic. The source shows an override that enables a human checkpoint, which pauses after each round for your input.

• "Run an auto review loop on my paper draft."
• "Review until it passes, then stop."
• "Keep improving this research project autonomously with auto review loop."
• Command form from the source: /auto-review-loop "topic" — human checkpoint: true

Use the human checkpoint for oversight

By default, HUMAN_CHECKPOINT is false and the loop runs fully autonomously. When it is true, the loop pauses after each round's review and shows you the score, top weaknesses and suggested fixes. You can approve the fixes, give custom instructions, skip specific fixes, or stop early.

Use the checkpoint when you do not want the agent to accept reviewer suggestions without a person checking them.

• Approve all suggested fixes.
• Give custom modification instructions.
• Skip specific fixes.
• Stop the loop early.

Safety notes: scripts, network and data sharing

This skill runs scripts and uses the network. Each round sends comprehensive project context, including claims, methods, results and known weaknesses, to an external Gemini service through the bridge. Do not use it on repositories where that is not acceptable.

The skill's own prompt tells the reviewer to assume a cooperating operator on their own machine. It also says that code which runs remote commands, starts a network service, or installs an MCP server should be reviewed strictly. Those are the source's rules, not a security audit.

• Full project context is sent to an external Gemini service.
• Scripts run on your machine with your credentials.
• The loop can apply reviewer-suggested fixes without a human checking them unless the human checkpoint is on.
• Review the state and log files after a run to see what changed.

Tips for better results

The reviewer scores work using the context it receives, so keep your research narrative, memory files and recent experiment results up to date before starting. The skill reads these during initialisation.

If a run is interrupted, the skill resumes from review-stage/REVIEW_STATE.json when its timestamp is within 24 hours and its status is in_progress. Older in-progress states are deleted and the loop starts fresh.

• Update project documents and recent results before the first round.
• Use the human checkpoint for work headed to submission.
• Expect a stale in-progress state (over 24 hours old) to be discarded.
• Check review-stage/AUTO_REVIEW.md to see how the scores changed.