A hands-on, two-day workshop teaching researchers to work with Claude Code as a research agent — not a chatbot that does the work for you, but a directed collaborator across the full empirical research cycle: literature review, data collection, analysis, and reproduction.
Four labs, ~2.5 hours each, one research question apiece — together they form a complete empirical research cycle directed entirely from the terminal.
Appraise a search strategy, turn screening criteria into an executable protocol, validate AI screening against a human gold standard, and catch a model that hallucinates references.
Force planning before collection, externalise a codebook and protocol, run incrementally with provenance on every observation, and diagnose failures at the level of the pipeline, not the row.
Direct an entire analysis without writing or reading code — onboard the agent, separate mechanical from judgment decisions, and resolve a live data crisis purely through supervision.
Reverse-engineer a published result, reproduce it independently, bound a defensible specification space, and referee a robustness vs. fragility debate between agents.
Everything for a pair to get started: the repo, each lab's starter folder as a ready-to-download zip, and every guide, environment file, subagent and skill described in full on the resources page.
The devcontainer, all four labs' starter content, and the reference subagents/skills — one zip, unzip and open in VS Code. Instructor-only material is deliberately excluded.
All four labs, the infrastructure design notes, and the reference subagents/skills — versioned, browsable, and where updates land first.
~440-record corpus, dedup log, gold standard, open-access full texts. Semantic Scholar reference verification runs live via the proxy below.
18 real, curated European companies (7 clear AI-governance / 6 vague / 5 none). The offline mirror is currently the fully-synthetic fallback set — see the resources page for status.
2,500-row synthetic survey engineered so a naive merge silently drops exactly 417 observations — the mid-lab crisis.
A self-contained replication package whose headline result is genuinely fragile — significant without sector controls, gone once you add them.
No participant ever handles an API key or crawls the open web — both run centrally and are described in full on the resources page.
Looking for the six reference subagents, the referee-review skill, or a specific CLAUDE.md? → Full resource index
Set these up before day one. No Anthropic API key needed — you'll log in to Claude Code with your workshop seat.
The agent you'll direct for all four labs.
curl -fsSL https://claude.ai/install.sh | bash
Setup docs →
Runs the pinned devcontainer (Python + R + Claude Code, identical for every pair).
Download →New to devcontainers? → Setup guide: Option A (VS Code + Docker Desktop) and Option C (devcontainer CLI, no VS Code)