AlexWortega/ai-peer-review-skill

README preview

# ai-peer-review-skill

A [Claude Code](https://claude.com/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](https://www.alignmentforum.org/posts/eJGptPbbFPZGLpjsp/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:

```

View full repository on GitHub →