Day 1

Tools and the walking skeleton

Day 1

Day 1 starts with a one-hour Claude Code warm-up, then moves from modern coding-agent practice to the general agent loop that every HADR agent runs on, plus the report and incident stores. The group should finish with one reviewed branch: green loop and store scenarios, a --no-llm skeleton run, and the loop exercised live on toy tools. The dispatcher that turns this into a full episode is Day 2’s first session.

Morning

The morning checks that everyone can build with Claude Code, then develops the habits needed to plan, delegate, verify, and review agent-written code. The HADR game becomes the running problem after the warm-up.

  1. Background: trace the milestones that made today’s coding agents possible.
  2. Claude Code warm-up: build a small, collaborative CRUD app in one hour, then explain and demonstrate it.
  3. The HADR game: meet the course project, its hidden state, and the dispatch agent your team will build.
  4. Plan, build, review: learn how to code agentically through scoping, planning, verification, validation, worktrees, and review - and build the team’s report and incident stores.

Afternoon

The afternoon connects the tools to the game and ends with the agent loop running live.

  1. Skills, tools, and MCPs: connect Claude Code to the game, solve it interactively, and package the procedure as a skill.
  2. Tokens and costing: inspect context growth, cost, and prompt caching.
  3. Evaluation: match each check to its evidence layer, and manufacture labeled intake data with Haiku subagents under /goal.
  4. Agentic loop: build the agentic tool loop every agent in the kit runs on, prove it on toy tools, and run it live against a real model.
  5. Checkpoint: merge, verify, and divide Day 2 work.
  6. Show and tell: share what your team built and saw.

Learning outcomes

You will learn how to:

  • Turn a narrow idea into a working, database-backed app with Claude Code.
  • Scope, plan, and review agent-written code, keeping requirements and acceptance with the developer.
  • Use plan mode, CLAUDE.md, worktrees, artifacts, and PR review loops as everyday habits.
  • Connect an agent to an external system through skills, typed tools, and MCP.
  • Read and manage context, token spend, and prompt caching.
  • Match each check to its evidence: deterministic scenario tests, model evals, or full game runs.
  • Build a complete agent loop from parts: standing orders, memory, tools, inner loop, and tick runner.