How I Use AI as an Economist
Notes from using AI in real, months-long research projects. Not a tutorial — what I have learned, what works, and what doesn't. Written for researchers who already use AI and want tricks beyond the basics.
Last updated: Jun 1, 2026. Download PDF version.
Principles
Three ideas that shape how I use AI day to day.
-
Teach the assistant once, benefit forever.
Every
CLAUDE.md line, skill, and memory compounds across projects and machines. The first task is expensive; the tenth is cheap. Not sure how to teach it? Just ask: "I want Claude Code to do X better in the future. What's the best way — a skill, a hook, a memory, or something else?"
-
Use structure, not willpower.
Plan mode, hooks, skills, and review passes keep the AI from charging ahead in the wrong direction. Unstructured prompts produce unstructured work. When a task is non-trivial, I reach for a workflow rather than typing harder.
-
The reviewer should not be the builder.
The agent that wrote something is the least likely to notice what it missed — like an author proofreading their own paper. Dispatch a second agent, or a second model, with a blank slate.
Dispatches
Notes I add as I learn. Sorted by most recently updated. Each card carries a permanent number (#N) and a floating Updated date.
Git, Dropbox, and AI agents
With agentic tools such as Codex or Claude Code, Git becomes less optional. Not because economists suddenly need to become software engineers, but because agents make many edits, branch off into parallel attempts, and need a clean way to review, merge, or discard their work. The awkward part is that many economics projects still live in Dropbox, the traditional tool for team folders, shared data, and co-author workflows. Dropbox is useful, but it can conflict with Git when both are trying to track the same changing files. Adrien Matray has a wonderful thread on X about exactly this GitHub-and-Dropbox trap for economists. If you are starting to use AI agents for research code, read it before your project folder becomes a sync puzzle.
Empirical Work in the Age of AI: three points that stayed with me
Stanford's Empirical Work in the Age of AI is worth watching from beginning to end. The whole event has useful material, but the part I keep returning to is Matthew Gentzkow's comments, starting around 55:56. They gave me three particularly sharp ways to think about research in the age of AI. First, maybe one yardstick for important work is whether the paper could not have been written before AI. Second, management becomes a central skill: researchers used to work almost like self-employed professionals until they were very senior, but AI now gives everyone something closer to a team of agents, and productivity depends on directing that team well. Third, reputation and trust become more important, not less. When the cost of analyzing a dataset and producing an answer falls sharply, the fundamental question is who we trust to produce answers worth believing.
AI Brain Drain
I recently read HBR's article on AI "brain fry": the mental fatigue that comes from overseeing more AI agents than our attention can comfortably hold. The part that stayed with me is the productivity warning. More agents are not automatically better if supervising them fragments attention and turns the day into multitasking across outputs. For research, the deeper cost may be even more important: too many parallel agents can crowd out the slow, uninterrupted thinking needed for research design, interpretation, and writing.
Digitizing historical records with AI? Read Noah Dasanaike's post.
Fellow econ-history nerds — if you want to use AI to digitize handwritten historical records (censuses, parish registers, tax rolls, ledgers, court books — anything in cursive), save yourself the false starts and go read Noah Dasanaike's writeup. It is the practical, end-to-end account I would have wanted before starting my own pipeline: which models actually read 19th-century handwriting, where the pipeline breaks, how to validate at scale. Better than anything I could add here.
Research runs for months and most trials fail. Keeping every failed attempt on main pollutes Claude's context; running parallel agents on one tree creates file conflicts. One flag — claude -w <name> — isolates each exploration in its own worktree, so only the winners ever touch main.
Read →
Principle 3 in practice. Review the plan when re-running is expensive (GPU training, LLM API calls). Review the execution when it is cheap to replicate. The mechanism is the same: pick option 4 on the "Would you like to proceed?" prompt and type /document-review, /codex:adversarial-review, or /ce:review.
Read →
We want Claude Code to run as an autonomous agent, but it keeps stopping to ask for permission on every bash or MCP call. Two recent features — auto mode and the /fewer-permission-prompts skill — silence the interruptions without resorting to --dangerously-skip-permissions.
Read →
Most usage overruns come from a handful of very long sessions, not many short ones. A practical framework — from personal experience and Anthropic's session management guide — for knowing when to keep going, compact, start fresh, or spin off sub-agents.
Read →
Four commands — /ce:brainstorm, /ce:plan, /ce:work, /ce:review — turn a drifting AI session into a structured workflow. How I use them on research projects that span weeks and months.
Read →
Turn one hard-won AI session into reusable instructions, skills, memories, or settings so the next similar task starts further ahead.
Read →
When one model gets stuck, ask it to package the problem for another model, then bring the second opinion back into the coding session.
Read →
A small skill for detecting common AI writing tics and rewriting text so it sounds more natural.
Read →
Use Claude Code remote control to keep long tasks moving when you are away from your desk.
Read →
After a conference, I use a custom Claude Code skill to turn receipts, policies, and templates into a reimbursement package.
Read →
Use hooks to inject context, block dangerous actions, and automate repetitive checks every time Claude Code runs.
Read →
A Dropbox-backed setup keeps shared Claude Code context synchronized while machine-specific settings stay local.
Read →
Use AI to build a study plan, ask better questions while reading, and keep the exchange candid rather than agreeable.
Read →
Resources