Hook exit codes: 0 success (stdout is context on UserPromptSubmit and SessionStart); 2 block, with stderr fed back to Claude; anything else is a non-blocking error, and exit 1 is the classic trap. A PreToolUse hook can instead exit 0 and print hookSpecificOutput with permissionDecision set to deny, allow, ask or defer (Chapter 22).

Skill frontmatter (Chapter 18): name, description, when_to_use, argument-hint, arguments, disable-model-invocation, user-invocable, allowed-tools, disallowed-tools, model, effort, context, agent, background, hooks, paths, shell. Subagent frontmatter (Chapter 20): name, description, tools, disallowedTools, model, permissionMode, maxTurns, skills, mcpServers, hooks, memory, background, effort, isolation, color, initialPrompt, plus omitClaudeMd and autoCompactWindow. Unknown keys are ignored silently in both, so lint them: Chapter 25 has a free script.

Each of these was observed while building the hands-on chapters, not copied from a blog. The chapter link has the evidence.

1. A broken skill can simply be missing. Two deliberately broken skills were absent from the session's skill list, and I could not establish why for the misspelled-key one (Ch 18).
2. A hook can fail open. With jq missing, the Chapter 12 hook allowed rm -rf and exited 0 (Ch 22).
3. Exit 1 does not block. Only exit 2, or a JSON deny, does (Ch 22).
4. PostToolUse sees only successes. An audit hook on it misses every blocked or denied attempt (Ch 22).
5. MCP permissions are per tool. Allowing one tool did not allow its sibling (Ch 23).
6. claude mcp list prints every connector you have, including work URLs. Filter before sharing (Ch 23).
7. Validation is not a test. claude plugin validate passed a plugin whose server path was wrong (Ch 24).
8. Project hooks and servers run under -p with no trust dialog, unless you use --bare (Ch 25).
9. A capped or failed run has a null result, and caps are checked after a step, so they overshoot (Ch 25).
10. A load error does not fail a run. A missing plugin or a bad MCP entry exits cleanly unless you gate on the start-up event (Ch 25).
11. A subagent starts blind, and a confident reviewer can be wrong: one of three findings had a false premise (Ch 20, Ch 26).
12. The raw event stream is not publishable. The start-up event lists your installed skills, servers and local paths (Ch 17).

The examples repository keeps a log of every captured run: model, turns and the cost Claude Code estimated. Totalled by part:

Three notes on reading that table. The costs are client-side estimates, not a bill. Most of Part 8's cost is the larger model answering about three dozen prompts; Part 11 was almost free because hooks, servers and plugins can be tested without a model and the sessions that remained used the small one. And what was not run matters as much: agent teams, HTTP and OAuth MCP servers, marketplace installs, --bare, GitHub Actions and the SDK packages appear on pages as documentation, with the pages saying so.