Claude-ECON_RA

An AI-powered applied economics research pipeline for Claude Code.

Six specialized agents. One research proposal in, full project out.

---

What This Is

An open-source Claude Code workflow that turns your terminal into a full-service applied economics research assistant — from literature review to journal submission.

Inspired by:

• Hugo Sant'Anna's Clo-Author — multi-agent research architecture
• Pedro Sant'Anna's claude-code-my-workflow — academic Claude Code template

Tailored for: Applied microeconomists using Stata, LaTeX, and causal inference methods.

---

Quick Start

# 1. Clone this repo
git clone https://github.com/Ghosh-Anupam/Claude-ECON_RA.git
cd Claude-ECON_RA

# 2. Start Claude Code
claude

Then type:

I am starting a new applied economics research project. Read CLAUDE.md and help me set up the project. Here is my research proposal:

[PASTE YOUR 300-500 WORD PROPOSAL]

---

The Six Agents

#	Agent	What It Does	Key Output
1	Lit Survey	Two-stage Google Scholar search (X+Y+Z joint, then X+Z relaxed) with gap analysis	PDF with literature table, ID strategy comparison
2	Identification	Interactive causal inference advisor — brainstorm, refine, and stress-test your research design	Strategy memo with estimating equation, assumptions, threats
3	Data Cleaning	Sequential download, clean, merge of data sources (one at a time)	Unified .dta files + .do scripts per source
4	Data Analysis	Full empirical analysis with replication package	Tables, figures, event plots, robustness checks
5	Presentation	Beamer LaTeX slide deck using custom style.sty	Seminar/job market/conference presentation
6	Journal Article	Complete manuscript using custom Class.cls	Publication-ready LaTeX article

How They Work

Each agent can be run independently or as part of the full pipeline.

Full pipeline (sequential):

Agents run 1 → 2 → 3 → 4 → 5 → 6

Independent (any order, any time after project setup):

Run the Lit Survey Agent          ← runs anytime, no dependencies
Run the Identification Agent      ← runs anytime (reads lit survey if available)
Run the Data Cleaning Agent       ← runs anytime, no dependencies
Run the Data Analysis Agent       ← needs data in Data/cleaned/
Run the Presentation Agent        ← uses whatever Tables/Figures exist
Run the Journal Article Agent     ← uses whatever outputs exist

Each agent checks what inputs are available and works with what exists. If something is missing, it tells you what's needed and asks how to proceed — it never refuses to run or silently triggers another agent.

---

Lit Survey: Two-Stage Search Protocol

The Lit Survey Agent uses a structured keyword approach built around three components:

• Y = Outcome variable (e.g., "Migration Rate")
• X = Treatment / shock (e.g., "Natural Disasters")
• Z = Unit of analysis (e.g., "US County level")

Stage 1 (Strict): Searches "Y" "Impact of X" "Z" jointly — finds papers studying the same treatment, outcome, and unit. Fewer results, but most directly relevant.

Stage 2 (Relaxed): Searches "Impact of X" "Z" only — finds all papers studying the treatment at the same unit level, regardless of outcome. Captures the broader literature.

Both stages include synonym expansion, targeted journal searches (AER, NBER, SSRN), and citation chaining.

---

Identification Agent: Interactive Causal Inference Advisor

Unlike other agents, the Identification Agent operates as a chatbot. It:

1. Asks about your ideal experiment and source of variation
2. Maps your design to a strategy (DiD, IV, RDD, synthetic control, etc.)
3. Explains why the strategy works in plain English before any equations
4. Shows how to implement it in Stata step by step
5. Stress-tests your design against threats a referee would raise
6. Compares your approach to what the literature survey found
7. Saves a strategy memo when you're satisfied

Reference textbooks (place PDFs in templates/textbooks/):

• Mostly Harmless Econometrics — Angrist & Pischke (2008)
• Causal Inference: The Mixtape — Cunningham (2021)
• Design of Observational Studies — Rosenbaum (2010)
• Causal Inference for Statistics, Social, and Biomedical Sciences — Imbens & Rubin (2015)

---

Project Structure

Claude-ECON_RA/
├── CLAUDE.md              # Project configuration (the brain)
├── MEMORY.md              # Cross-session learning
├── .claude/
│   ├── agents/            # 6 agent definitions
│   ├── skills/            # 6 slash commands
│   ├── rules/             # 5 always-on rules
│   └── references/        # On-demand reference docs
├── Paper/                 # Journal article (Class.cls)
│   └── 1-Sections/        # Modular .tex sections
├── Presentation/          # Beamer slides (style.sty)
│   └── 2-Presentation/    # Modular slide files
├── Data/                  # Raw + cleaned datasets
├── Scripts/               # All .do files
├── Tables/                # Regression output
├── Figures/               # Plots and graphs
├── LitSurvey/             # Literature review PDF
├── quality_reports/       # Plans, session logs, strategy memos
└── templates/             # LaTeX templates + textbooks
    ├── style.sty           # Custom Beamer style
    ├── Class.cls           # Custom article class
    └── textbooks/          # Reference PDFs (local only, not uploaded)

---

Key Design Principles

Interactive by Design

Every agent asks questions before doing work. You never get a 50-page output that misses the point.

Plain English First

The Identification Agent explains causal inference concepts in plain English before showing equations or code.

Sequential Data Processing

Data sources are cleaned one at a time. You approve each one before moving to the next.

Stata-First (Flexible)

Default output is .do files, but R and Python are supported when specified.

Custom LaTeX Integration

Uses your own style.sty (Beamer) and Class.cls (article) with modular \input{} structure.

Quality Gates

Scoring thresholds (80/90/95) prevent low-quality outputs from advancing through the pipeline.

Reproducibility Built In

Every analysis produces a self-contained replication package that runs from raw data to final outputs.

---

Prerequisites

Tool	Required For	Install
Claude Code	Everything	npm install -g @anthropic-ai/claude-code
Stata	Analysis & data cleaning	stata.com
TeX Live / MiKTeX	LaTeX compilation	tug.org/texlive

Optional: R (with fixest, modelsummary), Python (with pandas, statsmodels)

---

Daily Workflow

# Start of day
cd "path/to/your/Claude-Econ RA"
claude

# Inside Claude Code
Run the Lit Survey Agent on: [your topic]
Run the Identification Agent
Run the Data Cleaning Agent
# ... etc.

# End of day (after exiting Claude Code with Ctrl+C)
git add .
git commit -m "End of day: [what you accomplished]"
git push

---

License

MIT License. Fork it, customize it, make it yours.

---

Acknowledgments

This project is inspired by and builds upon:

• The Clo-Author by Hugo Sant'Anna
• claude-code-my-workflow by Pedro Sant'Anna

The multi-agent architecture, quality gates, and plan-first workflow patterns originate from those projects.