Automated paper review powered by Claude Code and edsl.
roboree reads your manuscript, runs a battery of issue spotters (math errors, overclaiming, logical gaps, etc.), investigates and deduplicates findings, votes on severity with multiple models, and produces a structured review — all from a single command.
pip install roboree
cd my-paper-repo/
roboreeOn first run, roboree will prompt you to log in to Expected Parrot if you don't already have an API key set.
Inside the Claude Code session, run:
/review-paper
to kick off the full pipeline, or run individual phases:
| Phase | Command | What it does |
|---|---|---|
| 0 | /setup-review |
Parse the manuscript, build a symbol table, copy issue spotters |
| 1a | /broad-sweep |
Run every chunk x spotter combination |
| 1b | /holistic-review |
Big-picture referee report |
| 2 | /investigate |
Gather evidence for each flagged issue |
| 3 | /deduplicate |
Cluster duplicate findings |
| 4 | /vote-severity |
Multi-model severity voting |
| 5 | /draft-responses |
Draft author responses |
| 6 | /publish-review |
Generate final REVIEW.md |
roboree installs Claude Code skills and agents into your project, then launches claude with everything pre-loaded. Heavy lifting (the broad sweep, severity voting) is farmed out to edsl for parallel multi-model execution.
All review state lives in an append-only .review/ directory — each phase writes new files, nothing is ever modified or deleted.
Default spotters cover: mathematical errors, undefined symbols, logical gaps, overclaiming, methodology errors, identification strategy, internal contradictions, and threats to validity.
Add your own from the CLI:
roboree add-spotter "p-hacking" "Look for signs of p-hacking or data dredging"
roboree add-spotter "selection bias" # creates a template to fill in
roboree list-spotters # see what's active
roboree remove-spotter p_hacking # remove oneOr from inside a roboree session:
/add-spotter check for missing confidence intervals
You can also edit the markdown files in .review/issue_spotters/ directly.
- Claude Code (
npm install -g @anthropic-ai/claude-code) - Python 3.10+
- An Expected Parrot account (free sign-up)