Claude Code intro

Several CEL repos contain a .claude/ folder and files like CLAUDE.md, MEMORY.md, and SESSION_REPORT.md. These come from Claude Code, the AI-assisted tooling used to build and document the repos.

You can ignore all of it to run the analysis

The pipeline is ordinary Stata. The .claude/ configuration is not part of the pipeline — you never need it to clean data, estimate, or produce tables and figures.

What the files are, if you're curious

• CLAUDE.md — project map and conventions; the AI reads it each session. Often the most complete single description of a repo.
• MEMORY.md — accumulated learnings and gotchas discovered during the work. Genuinely useful when debugging, even if you never run the AI tooling.
• SESSION_REPORT.md / quality_reports/session_logs/ — chronological work logs.
• .claude/ — rules, agent definitions, and hooks that shape how the AI assists. Safe to ignore.

The ideas behind it (if you're curious)

The .claude/ setup isn't just "let the AI write code." A few ideas are what make it produce work I can actually trust — and they're the genuinely interesting part:

It's a capable RA, not a senior researcher. I think of the AI as a sharp first- or second-year RA: technically proficient (writes Stata/R that runs, drafts text, reads and summarizes papers), but with textbook-level subject knowledge and no five years on my question. So it handles the execution; the judgment — what's correct, novel, and worth doing — stays mine. Most of the scaffolding exists to force evidence so I catch the gaps an RA wouldn't catch on their own. (See where the workflow fits, and where your judgment is irreplaceable.)

Four rules against making things up. The most distinctive piece — and the answer to the obvious "but doesn't AI just hallucinate?" worry. Four rules each block a different category of fabrication, so everything the AI claims is auditable: you can ask "where did you get that?" and the answer is always a tracked file, a recorded grep, or an explicit assumption.

Rule	What it stops the AI from doing
no-assumptions	Guessing your deadline, target journal, or preferences instead of asking
primary-source-first	Citing or paraphrasing a paper it hasn't actually read
derive-don't-guess	Inventing a filepath, variable, or macro the repo already defines
adversarial-default	Claiming "it works" / "it's compliant" without showing the check

(The full four-rule stack — you'll recognize these as the same rules this hub holds its docs to.)

Every creator has a critic. Any agent that produces something is paired with a separate critic agent that only reviews it — never edits — and scores it, because the agent that wrote something is biased toward "looks fine" and shares its own blind spots. Splitting the roles surfaces issues self-review misses. You're looking at it right now: this hub uses a docs-writer / docs-critic pair from the same template. (See worker–critic pairs.)

If you want to use it

Claude Code is an optional productivity tool, not a lab requirement. The agents, rules, and hooks you see in these repos come from my own template, Claude Code Research Workflow — a fork-and-adapt foundation for AI-assisted empirical research (discovery, data engineering, analysis, writing, peer review). If you want to work this way on a new project, fork it and adapt; its docs/ cover getting started, the agents and skills, and customization. It's an evolving preview release, so expect it to change.