A foundation for AI-assisted academic work using Claude Code. You describe what you want — lecture slides, a research paper, a data analysis, a replication package — and Claude plans the approach, runs specialized agents, fixes issues, verifies quality, and presents results — goal-first and gate-enforced. Like a contractor who handles the entire job, with you as the auditor of the disagreements the review loop surfaces. It is deliberately not an autonomous daemon: every loop is started by you or a skill, never on its own.
Producing academic work is no longer the slow part. Checking it is. An AI can draft a lecture, run a regression, and write the paragraph describing the result faster than you can read it — and every one of those outputs can be confidently, fluently wrong in a way that looks exactly like being right.
Twenty bugs were deliberately planted in a working codebase, and the review agents were asked to check it again. They reported everything was fine.
Nothing in the output distinguished that from a genuinely clean run.
So this template is built around one idea: a check earns trust by catching something, not by existing. Every one of its gates has been shown a planted defect and confirmed to go red — recall and false-alarm rate recorded in a qualification ledger that is itself one of the gates: a check with no row in it fails the build. What has not yet been measured is listed there by name as visible debt, because an unqualified check is not weak evidence — it is none.
| If you are… | Start here |
|---|---|
| Writing a paper | /review-paper --peer aer simulates an editor and two dispositioned referees.
/verify-claims checks every citation and number against source.
/challenge asks whether your result survives the specifications you didn't run. |
| Running an analysis | /data-analysis for the end-to-end R pipeline,
/diagnose when a number is wrong
and you don't know why. |
| Teaching a course | /syllabus → /create-lecture → /scaffold-exercises, with
Beamer and Quarto kept in sync and reviewed by pedagogy, layout, and proofreading agents. |
| Replicating or refereeing | /audit-reproducibility checks claims against outputs;
/differential-audit compares two implementations;
/replication-package builds the AEA DCAS deposit. |
Seven rungs, each because a cheaper one let something through. The first is the one most workflows skip:
/vaccinate../scripts/backtest.sh.Two things this template states plainly because they are measured, not assumed: AI peer review does not reduce analytical-choice variance (it catches errors; robustness is a separate rung), and agreement between models is not confirmation — they correlate on the same wrong answer.
The most common way AI-assisted empirical work goes wrong is not a wrong number. It is evidence for one question being read as evidence for another.
Reproducibility (does the code run and produce the reported numbers?) → implementation fidelity (does it implement the estimator it claims?) → statistical performance (does it behave in finite samples?) → measurement validity (does the variable capture the construct?) → identification (is the causal claim warranted?).
Evidence for one never clears another. When an agent reports "all checks pass", it is almost always speaking about the first and being heard about the last.
18 agents, 60 skills, 37 rules, 8 hooks — and ten gates that run on every commit.
# Fork this repo on GitHub, then:
git clone https://github.com/YOUR_USERNAME/claude-code-my-workflow.git my-project
cd my-project
./scripts/install-hooks.sh # activates the gate suite on every commit
Start Claude Code, describe your project in two or three sentences, and it will read the
configuration and adapt the workflow to your use case. You do not need all of this on day one
— begin with CLAUDE.md and two or three skills, and add the rest as you hit the
problems they solve.
Everything is on GitHub: pedrohcgs/claude-code-my-workflow. MIT licensed — use freely for teaching, research, or any academic purpose.