Claude Code for Research

Used carelessly, LLMs and AI agents may be adversarial: they hallucinate sources, cut corners, become “lazy,” and in the worst case fabricate results. Researchers bring deep domain knowledge and sharp questions to these tools, but an academic career rarely includes experience managing a large team. At the current stage of the technology, that management skill is what allows a researcher to make good use of these tools. This workshop introduces a framework for giving agents the context, purpose, and accountability they require, so that research time is spent on research rather than on debugging, formatting tables, and other tedious parts of the work.

The session covers the parts of Claude Code most relevant to a research workflow, beginning with planning mode and context management and building toward custom personas, agents, and skills. It presents a model of “Adversarial Agents”: assigning permissions and roles so that agents check one another’s work and hold each other to the instructions they were given.

Hands-on tracks

Participants pick one track and work on it for the rest of the session.

• Your own project: set up folders and project instructions (CLAUDE.md), then use plan mode on one real task.
• Proof of concept: see if you can get from an idea to a first estimate in 90 minutes.
• Literature library: ingest 10–20 of your PDFs and search the summaries, building a retrieval-augmented generation (RAG) system for your papers.
• Agents and skills: write a reviewer agent and run it on your draft; turn one task into a skill.
• Data from a messy source: scrape a site, parse a PDF table, or pull from an API (Census, IPUMS, BLS).
• Teaching: write a problem set, solutions, and a grading rubric for a course you teach. Example: discussion slides on heteroskedasticity.

Materials

• Lecture slides [Interactive HTML Slides] [pdf]
• Workshop materials [link]

Workshop Offerings

• California Center for Population Research (CCPR), UCLA 2026