ai-peer-review-skill

A Claude Code skill that runs a multi-reviewer peer review of an academic paper.

Drop a PDF in, get back N independent reviews + a synthesized meta-review + a CSV table of which reviewer raised which concern.

What it does

Given a PDF (or DOCX / .txt / .md) of a paper, the skill:

1. Extracts the text.
2. Spawns N reviewer subagents in parallel with anonymized NATO codenames (alfa, bravo, charlie, …). Each subagent sees only the paper and produces an independent, structured review (summary → major concerns → minor concerns → verdict).
   • By default, one of the panel slots is filled by an AI Alignment Forum-style critic that follows Neel Nanda's Highly Opinionated Advice on How to Write ML Papers — hard red-teaming on narrative, novelty, baselines, ablations, post-hoc analysis, p-value rigor, reproducibility, and an explicit "what did this update in my beliefs?" check. Disable with alignment_critic=false.
3. Synthesizes a meta-review in the main thread, identifying common vs unique concerns, ranking the reviewers by usefulness, and producing a final verdict.
4. Extracts a concerns table — a boolean matrix of concern × reviewer — and saves it as CSV.
5. Bundles everything into results.json.

Output layout:

papers/<paper-stem>/
├── review_alfa.md
├── review_bravo.md
├── review_charlie.md
├── review_delta.md
├── review_echo.md
├── meta_review.md
├── concerns_table.csv
└── results.json

Where it came from

This is a Claude Code skill port of poldrack/ai-peer-review by Russ Poldrack — a Python tool that calls 6 different proprietary LLMs (GPT-4o, GPT-4o-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro, DeepSeek R1, Llama 4 Maverick) to peer-review a paper, then synthesizes a meta-review.

The port differs in two ways:

	Original	This skill
Reviewers	6 different proprietary LLMs	N parallel Claude subagents (default 5)
API keys needed	OpenAI + Anthropic + Google + Together	None beyond Claude Code itself
Diversity	True cross-model diversity	Independent generations, single model family
Domain	Hard-coded to neuroscience	domain argument, inferred from paper if omitted

The skill keeps the original's artifact layout (review_*.md, meta_review.md, concerns_table.csv, results.json) and the NATO-codename anonymization scheme so outputs are interchangeable.

If you actually need cross-model diversity (e.g. for a methods paper about AI peer review), use the original poldrack/ai-peer-review Python tool instead. The SKILL.md documents this fallback explicitly.

Install

git clone https://github.com/AlexWortega/ai-peer-review-skill.git
ln -s "$(pwd)/ai-peer-review-skill" ~/.claude/skills/paper-review

(or symlink into <project>/.claude/skills/paper-review for project-scoped install.)

Restart your Claude Code session so the skill is picked up.

Use

In Claude Code:

Peer-review this paper: ~/Downloads/manuscript.pdf

or

/paper-review ~/Downloads/manuscript.pdf

Optional knobs (just say them in plain language):

• domain — "neuroscience and brain imaging", "reinforcement learning", etc.
• num_reviewers — 3 to 8 (default 5)
• output_dir — defaults to ./papers/<paper-stem>/
• skip_meta — only individual reviews, no synthesis
• overwrite — regenerate existing review_*.md files

Layout

.
├── SKILL.md                              # frontmatter + workflow Claude follows
├── prompts/
│   ├── reviewer.md                       # generic reviewer template
│   ├── reviewer_alignment_forum.md       # AAF-style critic (Nanda framework)
│   └── metareview.md                     # synthesis template
└── README.md                             # this file

Credits

• Original tool and prompt design: Russell Poldrack — poldrack/ai-peer-review
• Skill adaptation: this repo

License

MIT (skill adaptation only). The upstream poldrack/ai-peer-review repo is unlicensed at the time of this port; only the design and workflow are referenced here, no upstream code is redistributed.