Use Claude Code Like a Pro

A workshop on building the process your agents follow, not just prompting better.

You build a shared expense splitter, splitwise-lite, and along the way you build the machine that builds it: a spec, a dependency-aware backlog, four subagents, hooks that enforce instead of ask, and a loop that works the whole backlog while you're not watching.

ADLC is the process the agents follow. Agentic OS is the machine that runs it. Claude Code is what both are built from.

What's here

worksheet.md	The workshop. Follow this.
slides/index.html	Slides for the room. Open in a browser.
starter	Starter project with all the agents, hooks etc.,

The worksheet is the real artifact. Every section ends with a checkpoint, so if something breaks you can skip to the next heading and still be caught up.

Setup

Check each of these before the session starts, not during it.

claude --version     # Claude Code, authenticated
python3 --version    # 3.11+
uv --version
git --version        # 2.5+
jq --version
node --version       # for npx ccusage
rtk --version
gh auth status       # GitHub CLI, logged in

You also need a GitHub account with permission to create a repo, for the MCP and PR sections.

If rtk --version errors, check which rtk. There's an unrelated tool with the same name (reachingforthejack/rtk, Rust Type Kit).

The three segments

1. Token usage. What a task actually costs, and how to cut that without cutting what you get. ccusage, rtk, context headroom.
2. Quality and consistency. Spec before code. Layered context. PM, engineer, QA and reviewer subagents. Hooks as enforcement. GitHub over a markdown file.
3. Autonomy and reproducibility. One task, one worktree, one PR. The whole backlog in parallel. An issue that fixes itself while you're at lunch, inside a sandbox.

By the end of Workshop

.claude/
  rules/          # conventions loaded only when relevant
  agents/         # pm, engineer, qa, reviewer
  skills/         # /task
  hooks/          # guard-deps, run-tests, require-green
  settings.json   # hook registration, sandbox
scripts/
  watch-issues.sh # the unattended trigger
CLAUDE.md         # always-on context
plans/
  spec.md
  backlog.md