name: auto-empirical-research-skills description: Route empirical-research requests through the Auto-Empirical Research Skills catalog when this whole repository is installed as one skill in Codex, CodeBuddy, Claude Code, or another IDE. Use to choose and load the right vendored AERS skill for causal inference, econometrics, replication, data acquisition, manuscript writing, peer review and referee responses, citation checking, de-AIGC editing, or full empirical-paper workflows, or an end-to-end run from raw data all the way to a finished Word (.docx) manuscript, without reading the entire repository at once. license: CC-BY-SA-4.0

Auto-Empirical Research Skills Router

Use this root skill when the full AERS repository has been installed as a single skill folder. Treat it as a router and catalog, not as a request to load every vendored SKILL.md.

The catalog holds 1,107 skills across 77 vendored collections. Never read them all — route to one, then load only that skill's SKILL.md.

Workflow

1. Classify the user's empirical-research task by stage, then load the single best-matching skill:

   • Full pipeline or orchestration: start with skills/69-Paper-WorkFlow/ or the skills/00* flagship analysis skills — skills/00-Full-empirical-analysis-skill_StatsPAI/ (StatsPAI), skills/00.1-Full-empirical-analysis-skill_Python/ (Python), skills/00.2-Full-empirical-analysis-skill_Stata/ (Stata), skills/00.3-Full-empirical-analysis-skill_R/ (R). Note the StatsPAI flagship has no dot in its prefix, so a skills/00.* glob misses it.
   • Causal inference and econometrics: pick by method from the table below, or search catalog/skills.json / docs/TAXONOMY.md.
   • AER or top economics journal work: start with skills/50-brycewang-aer-skills/.
   • Replication, citation, or peer review: use docs/SKILL_CATALOG.md and docs/GOLDEN_WORKFLOWS.md to choose a focused skill.
   • Academic de-AIGC (English or Chinese) or academic rewriting: start with skills/48-de-AIGC-skills/ or nearby writing skills in the catalog.

2. Read only the selected child skill's SKILL.md, then follow its progressive-disclosure instructions for references/, scripts/, assets/, or templates.

3. If no child skill clearly matches, run the ranked catalog search first — it takes English or Chinese task text:

   python3 scripts/find-skill.py "staggered difference-in-differences event study"
   python3 scripts/find-skill.py "降低中文论文 AIGC 率" -k 5
   

   It ranks by relevance, then by the routing tier in catalog/curation.json: first-party core skills first, a single preferred copy for each duplicated name (the other copies are folded away), and natural-science / CS guides (out-of-domain) last unless the query names them. Take the top result unless its description clearly misfits, then read only that SKILL.md. Its accuracy is measured by python3 scripts/check-routing.py against evals/routing-cases.json.

   If you need raw data instead, catalog/skills.json has path, name, description, line_count, and a globally-unique qualified_name; catalog/skills-enriched.json adds tier, tags, quality_score, license, and commercial_use. Avoid broad recursive reads of skills/.

   • Both catalog JSON files are large (roughly 1 MB / 20k lines each) — query them instead of reading them whole. Example:

     python3 -c "import json; [print(s['qualified_name'], '->', s['path']) for s in json.load(open('catalog/skills.json'))['skills'] if 'synthetic control' in (s['name'] + ' ' + s['description']).lower()]"
     

     A plain grep -in "synthetic control" catalog/skills.json works too when a rough match is enough.

4. For installation help, use docs/INSTALL.md for Codex-style copy installs and INSTALL.md for Claude Code marketplace/plugin installs.

5. If editing this repository, keep parent and nested repos separate. In particular, inspect git status inside skills/69-Paper-WorkFlow/ (a git submodule) before touching it.

Method → where to start

Match the user's identification strategy or task to a starting collection, then confirm against catalog/skills.json.

This table is a shortcut to the most common starting points, not a complete index — it names fewer than half of the vendored collections, and the rest are reachable only through catalog/skills.json. A task missing from this table is not a task without a skill: fall through to step 3 and search the catalog before concluding nothing matches.

Full-pipeline trigger

If the user is asking for a complete empirical paper from idea to submission, route to skills/69-Paper-WorkFlow/. The orchestrator loads the right skill at the right stage and stops for human decisions at the two hard gates (Method Gate after Stage 3, Draft Quality Gate after Stage 7).

Trigger phrases (any one is enough to dispatch to the orchestrator):

• /paper-workflow
• "帮我写一篇实证论文"
• "从选题到投稿"
• "end-to-end empirical paper"
• "完整复现"
• "from proposal to submission"
• "从数据到 docx 论文全文" / "一条龙" / "出一份 Word 版论文"
• "raw data to a finished Word manuscript"

The orchestrator is not the right entry point for a single-task ask (e.g. "fit a DiD", "recode this variable", "write a referee report") — those are listed in the Method → where to start table above.

Coverage Notes

• skills/69-Paper-WorkFlow/ is a git submodule. If its folder is empty, the copy or clone skipped submodules (git submodule update --init fixes a clone); fall back to the skills/00* flagship pipeline skills, which are vendored directly. Those end at publication-ready tables and figures plus a Step 8.5 handoff contract (exhibits_index.md + results_summary.json); pair them with a writing skill for the manuscript itself, since assembling and gating the full .docx lives in the orchestrator.
• The vendored ARIS collection (skills/42-wanshuiyin-ARIS/) also ships its skill set as OpenAI Codex CLI runtime ports (skills-codex* subtrees). Those stay on disk but are excluded from catalog/skills.json (see scripts/skill_discovery.py) — route Claude agents to the primary skills/ tree only.

Install Notes

• Whole-repo imports are supported by this root SKILL.md as a lightweight compatibility entry point.
• Individual skill installs are still preferred when a runtime expects one folder per skill. Copy the folder that directly contains the target SKILL.md.
• Do not copy the repository root into a runtime and expect every child skill to become individually registered unless that runtime explicitly supports recursive skill discovery.
• Do not flat-install the whole catalog (e.g. symlinking every child folder into ~/.claude/skills). Every registered skill's description is loaded at session start — about 64k tokens for all 1,107 here — and runtimes truncate long skill listings, so matching gets worse, not better. Use a plugin, this router, or a handful of copied skills.
• Name collisions: the catalog contains 47 bare names shared across collections (e.g. data-analysis, lit-review, proofread); catalog/curation.json names one preferred copy of each, which scripts/find-skill.py returns. When a runtime registers skills by flat name, install one collection at a time, or disambiguate with the globally-unique qualified_name field in catalog/skills.json (<collection>::<name>, e.g. 12-pedrohcgs-claude-code-my-workflow::data-analysis), or the full skills/<collection>/.../SKILL.md path.

Key Files

• catalog/skills.json: machine-readable list of vendored skills.
• catalog/skills-enriched.json: same list plus tier, tags, quality_score, license, and commercial_use for filtering.
• catalog/curation.json: hand-curated routing tiers (core collections, the preferred copy of each duplicated name, out-of-domain prefixes). Ranking only — nothing is removed.
• scripts/find-skill.py: ranked search over the catalog (the router's search step).
• docs/SKILL_CATALOG.md: human-readable skill index.
• docs/TAXONOMY.md: task and method taxonomy.
• docs/GOLDEN_WORKFLOWS.md: ready-to-use empirical-research prompts.
• docs/INSTALL.md: runtime installation guidance for single-skill and whole-repo use.
• docs/CONTENT_ZH.md and README-zh-CN.md: Chinese-language collection index and entry point. Prefer these when the user is working in Chinese — several collections (de-AIGC, SSCI/CSSCI polishing, Chinese academic writing) are documented there in more detail than in the English docs.

Auto-Empirical Research Skills (AERS)

🔥 力荐 · StatsPAI — Stata 与 R 的 Python 平替，为 AI Agent 而生

• 🔁 Stata / R 平替：reghdfe、csdid、rdrobust、synth、fixest、did 等常用命令在 Python 中一站式替代，一个 import statspai as sp 搞定 DID / RD / IV / SCM / DML / matching
• 🤖 面向 Agent 设计：结构化结果对象 + 内置 MCP Server，Claude / Codex 等智能体可直接调用、诊断、审计与出表
• ⚡ 运行速度更快：无需启动 Stata / R 会话，免许可证，纯 Python 流水线端到端提速
• 🏛️ 同一团队维护：StatsPAI 与 AERS 由同一团队（Stanford REAP × CoPaper.AI）开发和持续维护，两者深度集成、同步迭代

github.com/brycewang-stanford/StatsPAI · pip install statspai · MIT 开源 · 欢迎 ⭐ Star

🌐 语言: English | 简体中文（默认） | 繁體中文 | 日本語 | 한국어

Stanford REAP × CoPaper.AI · 实证研究 AI 工具的学术工业级产品 由斯坦福实证研究方法论团队打造，覆盖从数据清洗到顶刊投稿的完整工作流

🚀 New here? Open the Skill Search → to filter all 1,107 skills by method, stage, language, and license. The 5-minute tour (make quickstart) prints the same picture in your terminal.

🇨🇳 中文用户从本文件开始（流水线速览 + 77 行总表），每个合集的完整描述见 docs/CONTENT_ZH.md。📖 English readers: see README-en.md.

信任面 · Trust surface (rigor stats)

Full trust overview: docs/TRUST.md · docs/RIGOR_COVERAGE.md

🏁 带上你自己的 agent 来考同一份卷子：pip install -e . 后用 aers-score 给自己打分，成绩发布在 docs/EXTERNAL_SCOREBOARD.md（规则见 docs/SCOREBOARD_RULES.md）。榜上的数字是我们用同一套评分器重算出来的，不是提交者自报的。

⚡ 安装与使用（30 秒上手）

最省事的一招：把 URL 丢给 Agent

把项目 URL 地址 https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills 丢给 Claude Code / Codex，并指定是目录 / 项目 / 全局安装 —— 剩下的让它自己做。例如：

帮我安装 https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills
装到「全局」（~/.claude/skills/），我想在所有项目里都能用

把最后一行换成你要的作用域即可：

手动安装（两种，任选其一）

A. 插件市场（Claude Code v2.1+，推荐，可升级）

claude plugin marketplace add brycewang-stanford/Auto-Empirical-Research-Skills
claude plugin install aer-skills@auto-empirical-research-skills                 # 顶刊投稿全流程（9 skills）
claude plugin install empirical-analysis-python@auto-empirical-research-skills  # Python 计量流水线
claude plugin install empirical-analysis-stata@auto-empirical-research-skills   # Stata 计量流水线
claude plugin install empirical-analysis-r@auto-empirical-research-skills       # R + Quarto 流水线

B. 只要某一个 skill —— 直接拷文件夹

git clone --recurse-submodules https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills.git
cd Auto-Empirical-Research-Skills

cp -R skills/00.1-Full-empirical-analysis-skill_Python  .claude/skills/   # 项目级
cp -R skills/00.1-Full-empirical-analysis-skill_Python  ~/.claude/skills/ # 全局

拷进去的文件夹必须自带 SKILL.md（部分合集的 SKILL.md 在下一层，拷那一层）。

怎么用

新开一个会话，直接用自然语言说要做什么，Agent 会按 description 自动挑 skill；说不动就点名方法或 skill：

用面板数据跑一个 Callaway–Sant'Anna 事件研究，并出 HonestDiD 稳健性和期刊级表格

完整安装说明（Codex / CodeBuddy 整库导入、--plugin-dir 单次加载、常见故障排查）见 INSTALL.md。

中文文档结构

中文内容分两级维护，各司其职：

• 本文件（README.md，GitHub 默认入口）：banner、badges、信任面、9 阶段流水线速览、77 行合集总表。
• docs/CONTENT_ZH.md（扩展正文）：每个合集的完整描述（#skill-NN 锚点）、按用途分组、精确数字、2 分钟验证、三层信任、旗舰流水线详解、贡献与引用。总表行内的 → 直接跳到对应锚点。
• 其他语言：README-en.md · README-zh-TW.md · README-ja.md · README-ko.md

[!NOTE] 维护规则： 改合集总表 → 本文件与 CONTENT_ZH.md 的锚点表两处同步；改合集详情 / 分组 / 数字 → 只改 docs/CONTENT_ZH.md。统计数字（合集数 / skill 数）以 catalog/skills.json 为准，由 make validate 的 readme-stats 检查器守护。

贡献者（Contributors）： 提交前请在本地跑通完整门禁 make check（catalog 校验 + 链接 + 单元测试 + eval-harness + benchmark）。详见 CONTRIBUTING.md。

旧版归档： README-zh-CN.md 已弃用，仅作向后兼容的重定向占位。

🚀 从一个 idea 到一篇论文：社科实证研究 · 端到端流水线（全自动、可介入）

AERS 不只是 77 个散装 skill —— 它能陪你走完一篇论文。 从模糊 idea → 选题精炼 → 文献综述 → 数据获取 → 识别策略 → 估计建模 → 稳健性审计 → 出版级表格 / 图形 → 写作与同行评审 → 降 AIGC → 投稿。端到端、全自动、每一步都可被人介入（中间任何一步你都可以接过去手工改方法、补变量、加稳健性，再让流水线自动接上跑）。

9 阶段流水线 · 每一步都覆盖到具体 skill