Show and tell

Day 1

Show and tell: one research question

If time allows: each member takes one question below, pastes it verbatim into Claude Code, and interrogates the answer - push back, question further, verify.

  1. When you are happy with the answer, ask Claude to build an artifact that presents your finding.
  2. Present it within your group, then share your artifact on the course Padlet so everyone else can read it.
  3. If time permits, turn it into slides too: start from the MARP template, optionally with the marp-slides skill.

The questions

  1. How do you wrap an API in an agent skill?
  2. How do you drive browser actions from Claude Code - the Chrome extension, a browser MCP, or a CLI tool? When is a browser the wrong tool for the job?
  3. How do you run an alternative model - Kimi K3, GLM 5.2, DeepSeek V4 Pro - inside Claude Code, and what degrades when you do?
  4. Pi ships a system prompt of about a thousand tokens and four tools; Claude Code ships far more of both. What do you gain and lose at each end of that spectrum?
  5. When should you use MCP vs agent skills vs CLI vs API?
  6. Explain the “smart zone” and “dumb zone” of the context window.
  7. Why would one use CLAUDE.md as a directory of links to other .md files instead of one large CLAUDE.md?
  8. What are the pros and cons of running coding agents (and models) on your own computer vs in the cloud?
  9. Find agent skills that review code quality, technical architecture, or security vulnerabilities. Compare and contrast them: what does each one actually look at, and where do they disagree?
  10. How should you format what you feed an agent - markdown, HTML, JSON, XML-style tags? Find a case where the format changed the answer.
  11. What is a sound decision-making process for choosing LLMs and inference providers?
  12. What happens when a tool result contains instructions aimed at the agent? Find a real prompt-injection case, and how harnesses defend against it.
  13. When does handing work to a subagent beat doing it in the main context - and when does it just cost more?
  14. What would you have to be sure of before letting an agent run unattended for an hour?
  15. How do you tell whether a prompt change helped, rather than just felt better?