How to use the Research Review via Claude agent skill

What an agent skill is and what research-review makes the agent do

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. This one is called research-review, and its SKILL.md is the only file in the package.

The skill is built on a Codex setup. A Codex agent first gathers context: it reads project narrative documents such as STORY.md, README.md and paper drafts, plus memory or notes files that record key findings and experiment history. It then sends that briefing to Claude through a local reviewer bridge, the claude-review MCP server. Claude is asked to act as a senior ML reviewer at NeurIPS/ICML level and to look for logical gaps, missing experiments, narrative weaknesses and whether the contribution is enough for a top venue.

The agent then runs follow-up rounds with Claude, responding to criticisms, asking targeted questions and requesting deliverables such as experiment designs or a mock review with scores. When the discussion converges, it saves the full exchange and conclusions to a review document in the project root and updates project notes.

Install the research-review skill from ZIP or Claude settings

The skill is an overlay. Install the base Codex-native skills first, then this package, which overwrites same-named skills.

Option 1, copy files manually: download the ZIP from this page and unzip it. Copy the skills/skills-codex-claude-review/* contents into ~/.codex/skills/ (user-wide) or into a project's .claude/skills/ folder, as your setup requires. Put the base skills from skills/skills-codex/* in ~/.codex/skills/ first.

Option 2, upload in the Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there.

Then register the reviewer bridge. The source describes this as a command you run once in your terminal:

• Install base skills from skills/skills-codex/* into ~/.codex/skills/
• Install this overlay from skills/skills-codex-claude-review/* into ~/.codex/skills/, allowing it to overwrite

Trigger the skill with natural requests

The skill activates when your request matches its description. The source lists phrases such as "review my research", "help me review" and "get external review", and also covers requests for critical feedback on research ideas, papers or experimental results.

Be specific about the material. Name the draft or results file and say what you want checked, so the agent knows which context to gather.

• "Review my research on the sparse attention method and tell me where the claims are weak."
• "Get an external review of my paper draft in drafts/paper.md before I submit."
• "Help me review these experimental results and list missing experiments."
• "Please write a mock NeurIPS review with scores for this draft."

What the review loop looks like

The first call starts a review and returns a jobId. The agent polls status until the job is done, then reads the response. Follow-up rounds reuse the saved threadId. The source notes that the reviewer gets no tools by default, so follow-up prompts that pass file paths ask for read-only access explicitly.

Expect the review to be adversarial. The prompt tells Claude to assume the work is broken somewhere and to verify rather than trust the author. The source also says Claude should state plainly when something is correct and should not manufacture findings.

• Round 1: full context plus specific questions, sent with review_start
• Rounds 2 to N: revised files and focused questions, sent with review_reply_start
• Stop when both sides agree on core claims, a concrete experiment plan exists, and the narrative is settled

Know the safety, network and data notes

This skill runs scripts and uses the network. The bridge server runs locally with python3, but the content it sends goes to Claude. Your project files and prompts are sent to Claude, so do not use this skill on confidential or unpublished work you are not allowed to share with a third-party model.

The skill is a research-workflow tool, not a security audit. Its prompt tells the reviewer to assume a cooperating operator on their own machine and not to treat malicious local users as in scope. The source also says that code which runs remote commands, starts a network service or installs an MCP server is held to a strict trust-boundary standard.

• Runs scripts: yes. Uses network: yes.
• Requires Python 3 and the claude-review MCP server at ~/.codex/mcp-servers/claude-review/server.py
• Optional: set CLAUDE_REVIEW_MODEL to override the Claude model used
• Your files and prompts leave your machine for Claude; check your rules first

Tips for getting a useful review

Give the reviewer a clear briefing. The skill's first step is to compile core claims, methodology, key results and known weaknesses, so write those down before you start. Vague context produces vague criticism.

Respond with evidence rather than just rebutting. The workflow asks you to answer each criticism with evidence or a counterargument, then ask targeted follow-ups on the most actionable points. Use the saved review document to track what you have addressed.

• Name concrete files with absolute paths when sending revisions
• Ask for deliverables: an experiment package ranked by acceptance lift per GPU week, a results-to-claims matrix, or a paper outline
• Keep the review document; it records round-by-round criticisms and the final TODO list with estimated compute costs