What it does This repo is a curated index of 119 GitHub repositories containing 23,000+ “Skills” — structured prompts and workflows that teach AI agents how to do empirical social science. It covers the full pipeline from data cleaning through causal inference (DID, IV, RDD) to robustness checks, tables, and submission formatting. The maintainers also ship a commercial tool, CoPaper.AI, that claims to produce a full paper in 20 minutes using these same skills.

The interesting bit The project treats methodology as reusable infrastructure. Instead of prompting an LLM fresh each time, you load a pre-built “Skill” that encodes, say, a complete difference-in-differences workflow — including what robustness checks a referee would expect. There’s even a vendored “AER-skills” stack for top-5 economics journal submission, with specific constraints like the 100-word AER abstract and AEA replication requirements.

Key highlights

• 23,000+ skills across 8 disciplines: economics, political science, sociology, psychology, public health, education, management, and public policy
• Machine-readable catalog at catalog/skills.json with local search and validation via make catalog && make validate
• Security-audited: 52/52 skills clean, 2,940+ files scanned across 13 risk categories
• Backed by Stanford REAP / SCCEI and powered by open-source StatsPAI (900+ causal inference functions, MIT-licensed)
• Weekly auto-sync workflow keeps vendored skill sets (like AER-skills) current

Caveats

• The repo is primarily an index and catalog; most actual code lives in the 119 upstream repositories
• CoPaper.AI is a separate commercial product; this repo is the “landscape” compiled while building it
• README makes strong claims about 20-minute papers but provides no independent benchmarks or sample outputs

Verdict Worth bookmarking if you’re building research agents or teaching LLMs to run econometric workflows. Less useful if you just want a drop-in analysis tool — you’ll still need to wire skills into your own agent runtime.