Learn About AI Coding Agents - Awesome Econ AI Stuff

🤖

What Are AI Coding Agents?

AI coding agents are intelligent assistants that can read, understand, and write code. Unlike simple autocomplete tools, these agents can:

Understand context — Read your entire project structure, documentation, and existing code
Execute tasks autonomously — Run commands, create files, and make changes across multiple files
Follow instructions — Adhere to project-specific conventions and coding standards
Learn from feedback — Improve based on corrections and preferences you provide

Popular AI Coding Agents

🤖

Claude Code Terminal-based agent from Anthropic

⌨️

Cursor AI-native code editor

✨

Gemini CLI Google's command-line agent

🐙

OpenAI Codex OpenAI's coding agent CLI

Example: Using an AI Agent for Economics Research

                        $ claude

> Analyze this panel dataset using a difference-in-differences 
  approach with state and year fixed effects. Use robust 
  standard errors clustered at the state level.

✓ Reading data from treatment_data.dta
✓ Generating Stata code with reghdfe
✓ Adding clustered standard errors
✓ Creating event study plot
✓ Exporting results to LaTeX table
                    

🛠️

What Are Skills?

Skills are reusable instruction sets that teach AI agents how to perform specific tasks. Think of them as "expert knowledge packages" that you can plug into your agent.

The Agent Skills standard defines a portable format using SKILL.md files that work across multiple AI tools.

Without Skills

You must explain every detail each time: "Use Stata's reghdfe package with clustered standard errors, format output as LaTeX, follow AER style guidelines..."

With Skills

Just say: "Run a DiD analysis on this data" — the skill handles all the domain-specific knowledge automatically.

Skill Structure (SKILL.md)

A skill is a folder containing a SKILL.md file with metadata and instructions:

                    # Folder structure
r-econometrics/
├── SKILL.md          # Required: instructions + metadata
├── scripts/          # Optional: helper scripts
├── references/       # Optional: documentation
└── assets/           # Optional: templates, data
                

                    # Example SKILL.md
---
name: r-econometrics
description: Run econometric analyses in R using fixest, modelsummary
workflow_stage: analysis
compatibility:
  - claude-code
  - cursor
  - gemini-cli
tags: [regression, panel-data, did]
---

# R Econometrics

## Purpose
Use this skill for econometric analysis in R, including 
difference-in-differences, instrumental variables, and 
regression discontinuity designs.

## Instructions
1. Use the fixest package for fixed effects estimation
2. Always report heteroskedasticity-robust standard errors
3. Create coefficient plots using ggplot2
4. Export tables in LaTeX format using modelsummary
                

How Skills Work: Progressive Disclosure

Discovery — Agent loads only skill names and descriptions at startup
Activation — When a task matches, the full instructions are loaded
Execution — Agent follows instructions, using bundled scripts/assets as needed

📋

What Are Rules?

Rules are project-level configuration files that provide persistent instructions to AI agents. They define coding standards, project conventions, and preferences that should apply to all interactions.

🔧 Project Rules

Live in your repository. Shared with all contributors. Version controlled.

AGENTS.md

👤 Personal Rules

Apply to all your projects. Your individual preferences.

~/.claude/CLAUDE.md

🏢 Team Rules

Organization-wide standards managed by IT/DevOps.

Managed policies

What to Include in Rules

Build commands — How to install dependencies, run tests, build the project
Code style — Formatting preferences, naming conventions, patterns to use
Project structure — Where files should go, how modules are organized
Domain knowledge — Project-specific terminology, data sources, methodologies
Security rules — Files to never modify, secrets handling, compliance requirements

Example: Rules for an Economics Research Project

                        # AGENTS.md

## Project Overview
This repository contains replication files for our JEL submission
on minimum wage effects in developing countries.

## Data
- Raw data is in `data/raw/` — NEVER modify these files
- Processed data goes in `data/processed/`
- All data cleaning must be reproducible via `make data`

## Code Style
- Stata: Use `reghdfe` for fixed effects, cluster at state level
- R: tidyverse style, use `fixest` for estimation
- All tables in LaTeX format following AER guidelines

## Testing
- Run `make test` before committing
- Ensure all figures reproduce exactly
                    

📄

AGENTS.md vs CLAUDE.md

Both are rule files, but they serve different purposes and have different scopes:

Aspect	AGENTS.md	CLAUDE.md
Scope	Project-specific	Personal (all projects)
Location	Repository root	`~/.claude/CLAUDE.md`
Version Control	✓ Committed to git	✗ Local only
Shared With Team	✓ Yes	✗ No
Compatibility	60k+ projects, many agents	Claude Code only
Best For	Build commands, project structure, team conventions	Personal preferences, communication style

💡 Pro Tip: Use Both

Put project-specific instructions (build commands, data handling) in AGENTS.md so your collaborators benefit too. Keep personal preferences (response style, your favorite packages) in CLAUDE.md.

Other Tool-Specific Files

Different tools may have their own configuration files that work alongside AGENTS.md:

.cursorrules — Cursor-specific rules (being unified with AGENTS.md)
.gemini/settings.json — Gemini CLI configuration
.github/copilot-instructions.md — GitHub Copilot instructions

🔀

Skills vs Subagents

These terms are sometimes confused, but they serve fundamentally different purposes:

🛠️ Skills

What: Instruction sets (text files) that teach an agent HOW to do something

Analogy: A recipe book — provides knowledge and procedures

When loaded: On-demand when relevant to the task

Execution: Main agent follows the instructions

🤖 Subagents

What: Separate AI instances spawned to handle specific subtasks

Analogy: Hiring a contractor — delegates work to another worker

When spawned: For parallelizable or isolated tasks

Execution: Independent agent runs, returns results

Example: When to Use Each

Use a Skill

"Run a DiD analysis following best practices"

→ The agent needs domain knowledge on HOW to do DiD correctly

Use a Subagent

"Search the codebase for all files that use this dataset"

→ An independent search task that can run in parallel

Use a Skill

"Format this regression output as a LaTeX table"

→ Needs knowledge of formatting conventions

Use a Subagent

"Read all 50 do-files and summarize what variables each creates"

→ Large, parallelizable reading task

Key Differences

	Skills	Subagents
Purpose	Add knowledge	Parallelize work
Context	Shares main context	Isolated context
Can edit files	✓ (via main agent)	Usually read-only
Persistent	✓ Reusable files	✗ Ephemeral

🔗

How They Work Together

All these components form a hierarchy that guides how AI agents understand and work on your project:

Organization Level Managed Policies

Company-wide standards, security requirements

↓

User Level ~/.claude/CLAUDE.md

Your personal preferences across all projects

↓

Project Level AGENTS.md

Project rules, conventions, build commands

↓

Task Level Skills (SKILL.md)

Domain expertise loaded on-demand

↓

Immediate Your Prompt

Specific instructions for this task (highest priority)

Resolution Order

When instructions conflict, more specific ones win:

Your prompt > Nearest AGENTS.md > Project root AGENTS.md > User CLAUDE.md > Org policies

Real-World Example: Economics Research Workflow

1

You open your research project

Agent reads AGENTS.md: learns about your data structure, Stata conventions, and that raw data must never be modified.

2

You ask for a regression analysis

Agent activates the stata-econometrics skill: now knows to use reghdfe, cluster standard errors, and format output for AER.

3

You specify "use probit instead"

Your prompt overrides the skill's default of OLS — agent adapts while keeping all other conventions.

4

Agent spawns a subagent to search for similar regressions

Parallel search through your do-files to check for consistency with existing analyses.

📚

Ready to Get Started?

Browse our curated collection of economics-specific AI skills.

Browse Skills Submit a Skill

Understanding AI Coding Agents