Open-source Agent Skills review research papers claim by claim against their evidence
paulnovosad · x · 2026-10-27
The Catalogue of Errors Ltd open-sourced three MIT-licensed agent skills on GitHub for reviewing research papers:
- /review: point an agent at a folder with the paper, its supplementary materials, and replication package; it assesses the article claim by claim against the evidence, tracing earlier drafts and auditing code.
- /audit-review: mechanically verifies the initial review for accuracy and completeness and corrects it.
- /review-digest: converts the audited review into a ranked, human-readable issue list.
The skills were developed across models from Anthropic, Cognition, Google, and OpenAI; the author typically runs them on Opus 5.5 High (untested on Sol 6.1). They burn heavy tokens, include built-in verification steps, and results should still be verified by humans.
More from coding & agent
- Consumer AI agents found inside 6 major banks; 77% of users won't grant money-transfer access — focusTrevor · 2026-10-27
- HarborRL open-sources agentic RL: 200+ tasks, 40+ harnesses, one reproducible pipeline — Dear_Treat3688 · 2026-10-27
- Voxel-city demo shows how to prompt Ling-3.1-flash for orbit, zoom, hover and focus interactions — Mrs-Grumpy23 · 2026-10-27
- Are technical design reviews disappearing in the age of coding agents? — Suspicious_Orchid770 · 2026-10-27
- Solo dev ships browser multiplayer 3D racer Voltane with Opus, Tripo, Suno — MacXou34 · 2026-10-27
- Compound Engineering 3.33.0 ships PR workflow upgrades to curb scope-creeping agents — kieranklaassen · 2026-10-27