Paper Replicate Agent Demo

Paper Replication Agent | Claude Code 论文复现工作流

作者： 朱 晨 | 遗传社科研究 Chen Zhu | China Agricultural University (CAU)

最后更新： 2026-05-22

Tutorial:

A structured Claude Code workflow for empirically replicating published research, created by 朱晨 @ China Agricultural University. Currently used with UK Biobank (UKB) and Chinese survey datasets (CHNS, CHARLS). Special thanks to Pedro H. C. Sant'Anna for the claude-code-my-workflow repository, which inspired this workflow.

You describe a paper; Claude plans the replication approach, writes R scripts, validates outputs against published targets, documents discrepancies, and saves a validation report — like a research assistant who handles the full pipeline.

中文介绍

论文复现智能体 — 面向经济学、流行病学等实证研究的结构化论文复现工作流，基于 Claude Code 构建。

这是什么？

本项目将 Claude Code 配置为一个论文复现研究助手：你提供论文 PDF 和数据集，Claude 自动完成从阅读论文、整理目标指标、编写 R 脚本、运行验证，到生成复现报告的完整流程——无需手动管理每一步。

适用于：
- 经济学实证论文（DID、RD、IV、固定效应、工具变量等计量方法）
- 流行病学与遗传社科研究（UK Biobank、CHNS、CHARLS 等大型队列/调查数据）
- 中文社科文献复现（中国国家健康与营养调查、中国健康与养老追踪调查等）
- 任何提供回归表格、样本描述统计或因果识别策略的实证论文

核心工作流（六阶段复现流程）

容差标准

| 零效应系数（\|t\| I want to replicate [Paper Author (Year)]. The PDF is in papers/[PaperName]/. Data is in data/[path]. Please enter plan mode, read the paper, identify all empirical targets, and plan the replication.

Claude will read the paper, inventory the data, draft a step-by-step plan, wait for your approval, then implement — running scripts, checking outputs against tolerance thresholds, and saving a validation report.

How It Works

Contractor Mode

For any non-trivial task, Claude enters plan mode first. You approve the plan, then it implements → verifies → reviews → fixes → scores. Say "just do it" for autonomous execution.

Replication Pipeline

Tolerance Thresholds

4 agents · 11 skills · 15 rules (click to…