Skip to content
Get packs

Guide · Agent skill

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

Updated · written with AI from the project's own documentation and checked against our catalogue data

Auto Review Loop is an agent skill that makes an AI agent review its own research work with a separate Claude model, fix the problems that reviewer finds, and ask again. It is aimed at researchers who want a second model to stress-test a paper or experiment, and at Codex users who want Claude Code to act as the reviewer. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before it will run.

Auto Review Loopwanshuiyin/Auto-claude-code-research-in-sleep · MIT

Quick start

  1. 1Download and unzip. Download the auto-review-loop ZIP from this page and unzip it so the auto-review-loop folder with SKILL.md is in place.
  2. 2Place the skill. Move the folder to ~/.claude/skills/ for personal use, or to .claude/skills/ in your project. Alternatively, upload it under Settings → Capabilities → Skills in the Claude apps.
  3. 3Install the reviewer bridge. Set up the claude-review MCP bridge with review_start and review_reply_start, and confirm you have Claude access through it.
  4. 4Set optional model. If you need a specific Claude model, set the CLAUDE_REVIEW_MODEL environment variable.
  5. 5Start the loop. Ask the agent for an auto review loop, or use the slash command with any overrides you want.
  6. 6Check the log. When the loop ends, open review-stage/AUTO_REVIEW.md and review-stage/REVIEW_STATE.json to see the rounds and final status.

What an agent skill is and what auto-review-loop does

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent loads when a request matches the skill's description. The agent then follows those instructions for that task.

The auto-review-loop skill runs a repeated cycle: review, implement fixes, re-review. The reviewer is a Claude model reached through a local claude-review MCP bridge. The agent reads the score and weaknesses the reviewer returns, makes changes, and sends the work back. The loop stops when the reviewer gives a positive verdict or after a fixed maximum of 4 rounds.

  • Reviews are sent to the reviewer through the review_start and review_reply_start tools.
  • A positive verdict needs a score of at least 6/10 and a verdict of 'ready' or 'almost'.
  • Progress is logged in the review-stage/ folder, so the loop can resume after the conversation is compacted.

Install the auto-review-loop skill

Download the ZIP for auto-review-loop from the skill's page on this directory. Unzip it so the auto-review-loop folder, containing SKILL.md, sits in one of these locations.

For a personal install, use ~/.claude/skills/. For a single project, put it in the project's .claude/skills/ folder. Alternatively, you can upload the skill in the Claude apps under Settings → Capabilities → Skills.

Codex users who want Claude Code as the reviewer should install this package after skills/skills-codex/*, because it overrides the reviewer choice.

  • Personal: ~/.claude/skills/auto-review-loop/
  • Project: .claude/skills/auto-review-loop/ inside your project
  • Claude apps: Settings → Capabilities → Skills, then upload the skill

Set up the claude-review MCP bridge

The loop cannot run without the claude-review MCP bridge. It is a local MCP server that exposes the review_start and review_reply_start tools. You also need Claude access through that bridge.

The skill's default reviewer model is claude-review. If you need a specific Claude model, set the optional CLAUDE_REVIEW_MODEL environment variable. The source does not list further setup steps for the bridge, so follow the bridge's own documentation for that part.

  • Required: claude-review MCP bridge with review_start and review_reply_start
  • Required: Claude access through that bridge
  • Optional: CLAUDE_REVIEW_MODEL to override the Claude model

Trigger the skill with a natural request

The skill loads when you ask for an automatic review cycle. The description matches phrases such as 'auto review loop', 'review until it passes', or a request for autonomous iterative improvement.

You can also invoke it with a slash command and override settings inline. The example below sets compact mode, a human checkpoint, and hard difficulty for one run.

  • 'Run an auto review loop on my draft in this project.'
  • 'Review until it passes, then summarise the remaining weaknesses.'
  • 'Start an autonomous iterative improvement pass on the experiment write-up.'
/auto-review-loop "topic" — compact: true, human checkpoint: true, difficulty: hard

Choose the settings that fit your project

The defaults run fully unattended at medium difficulty. Each setting changes how much you are involved or how hard the reviewer pushes.

HUMAN_CHECKPOINT = true pauses after each round's review, shows you the score and weaknesses, and waits for your decision. COMPACT = true makes the loop read summary files on recovery and append key findings after each round. RENDER_HTML is true by default and renders the review log to HTML when the loop ends, using --no-review.

  • REVIEWER_DIFFICULTY = medium (default): normal high-rigor Claude review.
  • REVIEWER_DIFFICULTY = hard: adds Reviewer Memory and a Debate Protocol, where you rebut objections and the reviewer rules on them.
  • REVIEWER_DIFFICULTY = nightmare: adds an independent reviewer that reads the repository directly and checks the author's claims against code, logs and results.
  • MAX_ROUNDS is fixed at 4 in this version of the skill.

Read the outputs after a run

The skill writes its files into review-stage/. The cumulative log is review-stage/AUTO_REVIEW.md. The current state is in review-stage/REVIEW_STATE.json, which is overwritten each round and marked completed when the loop finishes. Reviewer notes are kept in review-stage/REVIEWER_MEMORY.md.

Positive verdicts are also recorded in review-stage/ACQUITTAL_LOG.jsonl. This file is append-only, so earlier entries are never removed. Each entry links to a trace in .aris/traces/.

  • review-stage/AUTO_REVIEW.md: cumulative log of rounds
  • review-stage/REVIEW_STATE.json: latest round state, status and last score
  • review-stage/REVIEWER_MEMORY.md: reviewer memory across rounds
  • review-stage/ACQUITTAL_LOG.jsonl: permanent record of positive verdicts

Safety and privacy notes

The skill does not run scripts, but it does use the network. Each review sends your work to an external Claude model service through the bridge. If your project is confidential, do not run this skill on it.

The skill's own acceptance markers are written by the agent. Treat a passing result as a signal to check, not as final approval. A person should read the final log and the work itself before any submission decision.

  • Runs scripts: no
  • Uses network: yes, reviews are routed through the Claude reviewer bridge
  • Do not use on confidential content that must not leave your machine

What people use it for

  • Stress-testing a paper draft with a second model before submitting it to a venue.
  • Checking an experiment write-up for unsupported claims or metric mistakes using hard difficulty.
  • Running an unattended improvement pass overnight and reviewing the log in the morning.
  • Using a human checkpoint to approve each round's suggested fixes before the agent applies them.

Frequently asked questions

Is the auto-review-loop skill free?

The skill is published under the MIT licence. The source does not state pricing for the Claude access the reviewer bridge uses, so check your own Claude access terms.

What do I need to run auto-review-loop?

You need the claude-review MCP bridge exposing review_start and review_reply_start, plus Claude access through that bridge. CLAUDE_REVIEW_MODEL is optional. Without the bridge the loop cannot run.

Where do I install the auto-review-loop skill?

Unzip the downloaded ZIP into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. You can also upload it in the Claude apps under Settings → Capabilities → Skills.

Can Codex use auto-review-loop?

Yes. Codex users who want Claude Code as the reviewer should install this package after skills/skills-codex/*. The reviewer is then a different model family from the Codex executor.

How many rounds does the loop run?

The loop runs at most 4 rounds. It stops earlier if the reviewer gives a score of at least 6/10 with a verdict of 'ready' or 'almost'.

What should I do if the loop stops or will not resume?

Check review-stage/REVIEW_STATE.json and review-stage/AUTO_REVIEW.md for the last round and status. Confirm the claude-review bridge is running and reachable, since the loop cannot run without it.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Use auto-review-loop if you want a second Claude model to stress-test research work and you can set up the claude-review bridge and accept that reviews leave your machine. Pick something else if your content is confidential, you lack the bridge, or you need a final approval decision that the skill's own acceptance markers can settle.

Download the skill