Module 7 · Day Seven
Workshop
Build day · bring something to ship
Workshop
How the workshop works

No prepared material. The day is yours.

  • Pitch what you'll build in the first 30 minutes.
  • Pair up if you want — solo is also fine.
  • Instructor available throughout for design, debugging, and code review.
  • Short demos at the end of the day — one slide or one screen-share each, whatever you have.

Polish is not the goal. Shape is — having built one end-to-end LLM app so you know what it feels like.

Workshop
Suggestions, in case you want them
Extend an earlier project
  • Add a tool registry to your chat app — let Claude call weather, math, or web search
  • Wrap your extractor in an MCP server — invoke from Claude Desktop
  • Add a query-rewrite step to your RAG app for follow-up handling
  • Add citation renumbering + UI polish to your RAG chat
  • Add streaming to any of them — backend SSE + frontend incremental render
  • Add conversation memory to your chat app — full history on every turn
Greenfield
  • An MCP server for an API you use daily (Notion, Linear, your own DB)
  • A multi-model router for a workload — classify cheap, answer expensive
  • A subagent-style explorer — read all PRs in a repo and summarise
  • A file-watching agent — does work in response to filesystem changes
  • A computer-use agent that automates one task in an app with no API
  • Anything else — describe it Monday morning and we'll scope it together

Pick something small enough to finish. A working tiny thing beats a half-built ambitious one. You can always extend it later.

Go build.
Workshop
Final Quiz
Why it works — the whole week, no code
Final Quiz · 1 of 2 — Foundations → Tools → Context
Why it works

Answer in a sentence or two — from understanding, not notes.

Answers locked
keyword: tokens
  1. The Messages API is stateless — you resend the entire conversation on every turn. Why design it that way, and where does the cost of that choice land?
    Answer
    🔒 Enter the passcode to reveal this answer.
  2. Why is "answering from its training data" the main way hallucination happens, and why does grounding the model in retrieved sources or tools reduce it?
    Answer
    🔒 Enter the passcode to reveal this answer.
  3. Why force a schema with structured outputs / tool definitions instead of just asking the model to "return JSON" in the prompt?
    Answer
    🔒 Enter the passcode to reveal this answer.
  4. Why is MCP a meaningful shift over hand-writing an integration for each API — what changes about who owns the tool schema?
    Answer
    🔒 Enter the passcode to reveal this answer.
  5. Why does RAG retrieve a handful of chunks instead of stuffing the whole corpus into the prompt? Name the two forces at play.
    Answer
    🔒 Enter the passcode to reveal this answer.
Final Quiz · 2 of 2 — Context → Agents → Production
Why it works
Answers locked
keyword: guardrails
  1. Why use semantic (embedding) search for retrieval instead of keyword matching?
    Answer
    🔒 Enter the passcode to reveal this answer.
  2. Why is the agent loop only ~20 lines while Claude Code is enormous? What's the distinction between the loop and the harness?
    Answer
    🔒 Enter the passcode to reveal this answer.
  3. Why hand work to a subagent instead of doing it in the main loop — what does context isolation buy you?
    Answer
    🔒 Enter the passcode to reveal this answer.
  4. Why is the real danger of an agent usually "10,000 small failures" rather than one catastrophic action, and what two classes of control address it?
    Answer
    🔒 Enter the passcode to reveal this answer.
  5. Why replace "it looks good" with a golden set + LLM-as-judge wired into CI as a regression gate — and why treat all retrieved or external text as hostile?
    Answer
    🔒 Enter the passcode to reveal this answer.