Claude & Claude Code

Claude & Claude Code

Reasoning LLM plus agentic coding CLI. If one tool touches every stage of this course, it's this one.

Two products, one course

• Claude (claude.ai) — the chat reasoning LLM: drafting, critiquing, summarizing.
• Claude Code — an agentic CLI that reads and edits files, runs commands, and follows the instructions in AGENTS.md.
• The course uses both, often in the same session: chat for a quick question, Claude Code when a task touches the repo.

Where it fits

Every stage in the course either runs through Claude Code directly, or is reviewed by it.

Getting access

• Sign up at claude.ai with an institutional or personal email.
• The Pro plan (≈ €22/mo incl. Dutch VAT) unlocks Claude Code.
• Install: npm install -g @anthropic-ai/claude-code
• Set ANTHROPIC_API_KEY for the session (D1·1).

AGENTS.md — the mechanism

• Standing instructions the agent reads on every session — not documentation for humans.
• The root file sets ground rules: no invented citations, no bare causal claims, disclose every AI-assisted step.
• D1·2: you extend it with rules specific to your own angle on the case — more specific rules win over the root file.

Skills — on-demand checklists

• A SKILL.md packages a repeatable review task an agent loads on demand, instead of retyping instructions every time.
• Reference: /referee-review — a 5-check gated pipeline over a manuscript draft (D1·4).
• You build your own narrow skill (a pre-trend checker, a data-provenance checker) as the exercise.

Subagents — narrow specialists

• Six reference subagents, each briefed narrowly and run independently (D1·5): literature-scout, identification-skeptic, data-integrity-checker, stats-reviewer, writing-editor, disclosure-auditor.
• Independent reviewers catch different things — one subagent's blind spot doesn't hide a problem another would have caught.

Running the Day-2 pipeline

• D2·1 Hypothesis — the generate → critique → refine loop, critiqued by identification-skeptic.
• D2·2 Design — the design memo, attacked by identification-skeptic before a single regression runs.
• D2·3 Data lab — running analysis.py/analysis.R, finding and fixing the pre-trend confound.
• D2·4 Results — writing up the naive vs. robust estimate honestly.

Writing & review

• D2·5 Manuscript — writing-editor tightens prose to match voice.md, never touching causal-confidence wording.
• D2·6 Referee & governance — the same referee-review skill applied to the real draft, then disclosure-auditor checks every AI-assisted step is logged.

What it's good at

• Fast iteration on a specific, well-scoped task.
• Reading and editing real files/code, not just describing them.
• Following a written rule consistently across a long session — if the rule is specific enough.

What it doesn't do on its own

• Doesn't verify a citation just because it sounds plausible — that's why literature-scout exists.
• Doesn't know if a regression spec is right without an adversarial pass — that's identification-skeptic's job.
• Doesn't disclose itself — the AI-use log is a human's responsibility, checked by disclosure-auditor, not generated by magic.

Recap

Every decision graph in this course describes what Claude Code should do at each gate. See the home page for the full pipeline and all eleven step-level graphs.