How to use the Auto Review Loop (MiniMax) agent skill

What is an agent skill?

An agent skill is a folder of instructions, usually centred on a SKILL.md file, that an agent such as Claude Code loads when a request matches it. The skill tells the agent a fixed routine to follow, so the agent does not have to work out the process from scratch each time.

This skill is a single SKILL.md file. Its description says it runs an autonomous multi-round research review loop using the MiniMax API.

What does the skill make the agent do?

The skill runs a loop: send the research work to an external reviewer, parse the score and verdict, implement fixes or run experiments, then request another review. The reviewer is MiniMax-M3, called through the MiniMax API.

The loop stops when the reviewer gives a score of 6/10 or higher and a verdict of "ready" or "almost", or after MAX_ROUNDS, which is 4. Progress is written to review-stage/REVIEW_STATE.json after each round, so an interrupted run can resume. A cumulative log is kept in review-stage/AUTO_REVIEW.md.

• Phase A: send context and a review prompt to the external reviewer
• Phase B: save the full raw review and extract score, verdict and action items
• Phase C: implement fixes, highest priority first
• Phase D: wait for any experiments launched in the round
• Phase E: document the round and write the state file

How do you install the skill?

Download the ZIP from the skill page on this site and unzip it. You can install it in one of two places. For a personal install available to all your projects, unzip it into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder.

If you use the Claude apps instead, open Settings, go to Capabilities, then Skills, and upload the ZIP there.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

What do you need before running it?

The skill needs Claude Code or a compatible agent that has Bash, Read, Grep, Glob, Write, Edit and Skill tools. It also needs a MiniMax API key, stored as MINIMAX_API_KEY either in ~/.claude/settings.json under env or as an environment variable. Network access to api.minimax.io is required.

Optionally, if you have the mcp__minimax-chat__minimax_chat MCP server, the skill uses it first. Otherwise it falls back to curl. If the loop waits on running experiments, you will also need screen sessions or a similar tool.

• Claude Code or a compatible agent with the listed tools
• MINIMAX_API_KEY in ~/.claude/settings.json (env) or the environment
• Network access to api.minimax.io
• Optional: mcp__minimax-chat__minimax_chat MCP server
• Optional: screen or similar, if experiments run in the background

How do you trigger the skill?

Use a natural request that includes one of the trigger phrases from the skill description: "auto review loop minimax" or "minimax review". The skill also accepts an optional argument describing the topic or scope, shown in its argument hint as [topic-or-scope].

Do not wrap the skill in /loop, /schedule or CronCreate. The skill already loops internally, and an external timer would discard the accumulated context between rounds.

• "Run auto review loop minimax on my draft paper claims."
• "minimax review of the experiments in this project"
• "Do a minimax review of the ablation section, scope: ablations only."

What happens in a typical run?

On the first round the agent gathers the current claims, methods, results and known weaknesses, then sends them to the reviewer with a request for a 1-10 score, ranked weaknesses, the minimum fix for each, and a ready/almost/not ready verdict. The agent saves the full response before parsing it.

If the verdict is not positive, the agent works through the action items and then, in later rounds, includes a summary of earlier feedback in the next prompt. If the run is interrupted, the next invocation checks the state file. A state file marked in_progress and updated within the last 24 hours lets the run resume at the next round; older ones are discarded.

What safety points should you check?

This skill runs scripts and uses the network. Review prompts include your research context, which is sent to MiniMax, an external service. Do not use it on confidential or unpublished data you are not willing to send to a third-party API.

The reviewer is a language model. Its scores are advisory, not a substitute for human peer review. The skill's allowed tools include Bash(*), so read the SKILL.md before installing and be aware of what the agent executes.

• Review prompts include research context, sent to an external API
• Scores are advisory and not human-quality peer review
• Skill allows Bash(*), Write and Edit, so check outputs and commands