Getting Started
A sequenced guide for economists — from zero to working AI workflow
The resources already exist — they’re just scattered. This page puts them in order.
Most economists don’t need a course. They need someone to say: do this first, then this, then this. That’s what this page is.
How to use this page: Work through the layers in order if you’re starting from scratch. Jump to wherever you got stuck if you’re partway there. Each layer ends with a checkpoint so you know when you’re ready to move on.
Short on time? Layers 1–3 are the core. You can be set up and doing useful work in an afternoon. Layers 4–5 are for when you’re ready to share your work with others.
Layer 1: The Terminal
You already use a terminal — Stata’s command window is one. The difference is that your computer’s terminal (Terminal on Mac, PowerShell on Windows) can control everything, not just Stata. AI coding tools live here, and learning a handful of commands makes the rest of this guide possible.
You need about 10 commands: ls, cd, pwd, mkdir, cp, mv, rm, cat, grep, which. That’s it for now.
Primary resource: Terminal Basics for Economists (~75 min, hands-on with economics data) — Our module. Uses a Kenya household survey dataset. Includes exercises with answer keys.
Supplement: The Unix Shell (Software Carpentry) — More thorough, with a biology-flavored dataset, but the shell commands are identical.
Windows users: Use PowerShell (not Command Prompt). Your prompt should start with PS C:\.
You can open a terminal, cd to a project folder, ls its contents, and grep for a variable name in a .do or .R file.
Layer 2: Version Control (Git & GitHub)
If your version control system is analysis_final_v2_REAL_USE_THIS.do, you need Git. Even if you work alone — because AI tools generate a lot of code, and you need a way to track what changed and undo what broke.
The core workflow is five commands: clone, add, commit, push, pull. Learn those and you’re functional.
Primary resource: Git & GitHub Essentials for Economists (~75 min, hands-on) — Our module. Covers the mental model, the 5-command workflow, reading diffs, and common mistakes. Includes an econ-specific .gitignore template.
Supplement: Git: A Guide for Economists (Frank Pinter) — The gold standard econ-specific git reference. More comprehensive than you need right now, but excellent to return to.
If you don’t want the terminal: GitHub Desktop is a legitimate choice. It does everything above through a GUI. Not a crutch — just a different interface.
You can clone a GitHub repo to your machine, make a change, commit it with a meaningful message, and push it back.
Layer 3: Your AI Setup
Now you need the actual tool. The two main options for economists right now are Claude Code (Anthropic) and Codex (OpenAI). Both are terminal-based, both can read and edit your files, both cost money beyond free tiers.
Not sure you need the terminal? Read When to Move to the Terminal — a short guide for people who are productive with desktop apps and wondering whether the terminal is worth it.
Installation: Follow our step-by-step install instructions for Claude Code or Codex on Mac and Windows.
Getting oriented after install:
- Getting Started with Claude Code (Paul Goldsmith-Pinkham) — A thorough walkthrough from a fellow economist. Covers setup, first tasks, and an explicit privacy framework.
- Claude Blattman: Essentials (Chris Blattman) — “I’ve never coded in my life, so if I can do this you can.” Beginner-friendly, operational focus.
On cost: Claude Pro ($20/mo) or ChatGPT Plus ($20/mo) are enough to get started. You’ll know if you need more. Don’t overthink this — pick one and try it for a month.
These tools send your files to remote servers during your session. Published data, your own code, and course materials are fine. IRB-protected data, student records, and restricted-use datasets are not. Paul GP’s rule of thumb: “If you wouldn’t email it to a collaborator, don’t point the AI at it.” Use .claudeignore to exclude sensitive directories.
You can learn Layers 1–2 (terminal + Git) for free — no subscription needed. For Layer 3, UVM Copilot is free with a UVM account (web chat, not terminal). Claude and ChatGPT both have free tiers for chat. The $20/mo subscription adds the terminal tools (Claude Code / Codex), which is where the real leverage is — but you can learn the foundations first and decide later.
You can open your AI tool in a terminal, cd to a project folder, and ask it to do something small — summarize a file, rename some variables, draft an email. It works, and you understand what it read to do the task.
Layer 4: Deploying Things
This is the layer nobody teaches. AI can build you a website, write your replication package, or generate a slide deck — but if you can’t get it off your laptop and onto the internet, it doesn’t matter. The gap between “AI made me a thing” and “the thing actually works in the world” is deployment. It’s not hard, but you have to learn it once.
Three things economists actually deploy:
If you use Quarto (for slides, course sites, or personal pages), you can publish for free on GitHub Pages. The workflow: write .qmd files → render → push to GitHub → site is live.
- Publishing with GitHub Pages (Quarto docs) — The official guide. Clear and complete.
- Quarto + GitHub Pages walkthrough (UBC DSCI 521) — A lecture-style walkthrough if you want more hand-holding than the official docs.
Sharing a GitHub repo is better than emailing a zip file. But a repo someone else can actually use needs: a README that explains what’s in it, a .gitignore that excludes junk, and file paths that aren’t hardcoded to your machine.
The config_local.do pattern: Instead of cd "/Users/emily/Dropbox/myproject/", create a config_local_template.do with placeholder paths. Each collaborator copies it to config_local.do (which is .gitignore’d) and fills in their own paths. Your master script includes it. Now the code runs on anyone’s machine.
If you’re assembling a replication package for a journal, the standard is a master script that runs everything from raw data to final tables. The AEA Data Editor’s guidance is the reference.
You’ve published something — a Quarto page, a GitHub repo with a README, or a shared project folder with config_local.do. The bar is: someone else could find it and use it without emailing you.
Layer 5: Connecting the Pieces
The previous four layers are independent skills. This section is about how they combine into workflows you’ll actually use.
Workflow A: “Make a course website”
Terminal (Layer 1) → create a Quarto project → AI (Layer 3) generates page content from your syllabus and notes → Git (Layer 2) tracks versions → GitHub Pages (Layer 4) publishes it. Time to live site: under an hour once you’ve done it once.
Workflow B: “Build a replication package”
AI (Layer 3) drafts a master do-file and README from your existing scripts → you review and fix what it gets wrong → Git (Layer 2) tracks the package → GitHub (Layer 4) hosts it. The AI handles the tedious parts (file headers, path management, README boilerplate). You handle the judgment (does this code actually reproduce the results?).
What comes next
Once you’re comfortable with the basics, Skills & Workflows shows specific research and teaching workflows. Resources covers integrations and advanced tools for when you’re ready to go deeper.
If you’re attending one of our webinars — this page is your pre-work. You don’t need all five layers before the session, but Layers 1-3 will help you follow along.