Peer Review
Work in Progress. This plugin is still under active development by a university student. For now, it is an experiment on the utility of Claude Code for peer-reviewing academic papers. If you have any feedback or suggestions, reach me on X at @felpix_.
A Claude Code plugin that runs an adversarial multi-agent peer review on an academic paper PDF. Spawns a Red Team of specialized critics (Foundations-Critic, Empirical Auditor, Procedural Auditor, Collector, Omissions Auditor), a Blue Team that writes honest defenses, an Assessor that adjudicates between them, fact-checkers that verify quotes and citations against the PDF, and — in writer mode — an author-facing editor, proofreader, and copyeditor. Produces a structured Credibility Assessment / Bottom Line / Potential Issues / Future Research report.
The pipeline's effectiveness comes from the verification cascade: aggressive prompting drives careful reading, and downstream verifiers strip out the LLM hallucinations that aggressive prompting also produces.
Quick Start
1. Install the plugin
Use Claude Code to add the marketplace and install the plugin.
/plugin marketplace add Felpix-Studios/peer-review
/plugin install peer-review@felpix-research
2. Install Python dependencies
The plugin extracts PDFs to plain text once per run (so a 15-agent pipeline doesn't re-render the same pages 15 times as images). This requires Python 3.10+ on PATH as python and the pypdf package.
pip install pypdf reportlab bibtexparser
reportlab is optional — it's only used by the compile-code-to-pdf.py helper when you want to pack a replication-code directory into a single dense PDF for the code-audit agents. bibtexparser is optional — it's only used by the parse-bib.py helper if you supply a .bib file at run start (the citation-checker uses it to verify cited references against your bibliography). The plugin's regex .bib parser handles common entries without bibtexparser; install it if your bibliography uses unusual concatenations or escape sequences. Installing all three now avoids a second failure later.
Preflight check. If Python or pypdf is missing when you run /peer-review, the skill halts at preflight and prints the install command — nothing is extracted, no agent is invoked, no directory is created.
3. Run /peer-review
Point the skill at a PDF. The invocation accepts free-form prose — there are no flags to parse:
/peer-review path/to/paper.pdf
/peer-review my paper is named "working-paper.pdf"
Claude will then ask a short follow-up covering:
- Math audit — auto-detect (default), force, or skip
- Code audit — whether to audit a replication directory (and if so, its path)
- Writer mode — include the author-facing Editor's Note + Copyediting (default on)
- Supplementary PDFs — any appendices or data supplements to include
Accept the defaults for a straightforward base run.
The final report is written to ./peer-review-report.md (top of your cwd). If pandoc and a LaTeX distribution with xelatex are installed, a companion ./peer-review-report.pdf is produced in the same directory (all LaTeX intermediates are cleaned up automatically). Every intermediate stage file — plus a paper_text.txt plain-text dump of the PDF and a generated README.md mapping each NN_stage.txt to its contents — is cached under ./peer-review-output/<slug>/, a visible folder in your cwd. A partial run can be resumed by re-invoking the command: only stages whose cached file is missing will re-run.
Prerequisites
| Tool | Required For | Install |
|---|
| Claude Code (with plugin support) | Everything | claude.ai/download |
Python (>= 3.10) on PATH as python | PDF text extraction (every run) | python.org |
pypdf | PDF text extraction (every run) | pip install pypdf |
reportlab (optional) | compile-code-to-pdf.py helper for code-audit runs | pip install reportlab |
bibtexparser (optional) | parse-bib.py helper for citation validation when a .bib is supplied | pip install bibtexparser |
pandoc (optional) | PDF export of the final report | brew install pandoc (macOS) / apt install pandoc |
LaTeX distribution with xelatex (optional) | PDF export of the final report | MacTeX / TeX Live / MiKTeX |
| Access to Claude Opus + Sonnet | 26 agents use Opus; 9 use Sonnet | Claude Code plan |
How It Works
The orchestrator drives a 30+ stage pipeline of subagents through a verification cascade designed to remove LLM hallucinations:
