auto-empirical-research-skills: 1,096 Econometrics Skills Backed by Benchmarks

Ask a coding agent to "run a diff-in-diff" and it will happily produce something that looks like econometrics and may be quietly wrong — wrong clustering, wrong comparison group, a standard error that does not mean what it says. brycewang-stanford/Auto-Empirical-Research-Skills (AERS, 3.6k stars) is a Stanford REAP × CoPaper.AI collection that packages empirical-research methods as agent skills — and, unusually for this ecosystem, ships the benchmark harness to check that agents run them correctly.

Why This Skill Matters

The catalog is large — the project's skill search counts 1,096 skills across 76 collections, covering the workflow the README describes as running from data cleaning to top-journal submission. What separates it from a prompt dump is the trust surface: 19 numeric benchmark tasks whose gold values are recomputed from real data on every run, plus 42 behavioral eval scenarios with 217 rubric items — 9 of those scenarios proven to distinguish correct agent behavior from incorrect. The repo even invites you to score your own agent with aers-score and publishes the results, recomputed by the same scorer, on an external scoreboard.

The skills lean econometrics-heavy: the example prompt on the README asks for a Callaway–Sant'Anna event study on panel data with HonestDiD robustness checks and journal-grade tables. Dedicated plugins cover Python, Stata, and R + Quarto pipelines.

Installation

The README's fastest path is to paste the repo URL into your agent and say where to install it:

Install https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills
globally so I can use it in every project

Scopes map to directories: the current folder (.claude/skills/), a shared project (committed to git), or global (~/.claude/skills/, or ~/.codex/skills/ on Codex). For Claude Code v2.1+, the plugin marketplace is the upgradable route:

claude plugin marketplace add brycewang-stanford/Auto-Empirical-Research-Skills
claude plugin install aer-skills@auto-empirical-research-skills

aer-skills is the full top-journal workflow (9 skills); swap in empirical-analysis-python, empirical-analysis-stata, or empirical-analysis-r for a single pipeline. To grab just one skill, clone with --recurse-submodules and copy the folder — the README notes some collections keep their SKILL.md one level deeper, so copy the layer that has it.

Real Workflow: An Event Study You Can Defend

1. Install the Python analysis plugin and open a new session with your panel dataset in the working directory.
2. Prompt in natural language — the agent picks skills by their descriptions:

Run a Callaway–Sant'Anna event study on this panel data, with HonestDiD
robustness and journal-grade tables

3. Check the output against the benchmark discipline: the skills are written to produce the estimation, the robustness checks, and formatted tables as separate verifiable steps rather than one opaque script.

If you want to know whether your agent actually follows the methods, run the project's own harness: pip install -e . gets you aers-score, which grades the agent against the benchmark tasks and prints where it went wrong.

Tips

• Documentation is Chinese-first; English, traditional Chinese, Japanese, and Korean versions sit alongside it in the repo.
• make quickstart prints a five-minute tour of the whole catalog in the terminal.
• The license is CC BY-SA 4.0 — different from the MIT/Apache norm in this ecosystem, so check attribution terms before redistributing.
• When a skill folder copy seems inert, verify the copied directory actually contains the SKILL.md — the README calls this out as the common failure.

When Not to Use This

The depth is in empirical economics and statistics workflows; if your research is wet-lab biology or qualitative fieldwork, coverage will be thin. And the scale cuts both ways — installing all 76 collections into one agent is exactly the context bloat that skill authors warn about, so install the pipelines you use.

See the leaderboard for more skills.