Captain's Log
You walked in as a passenger. You leave as the captain.
Overview
You made it. That's not a figure of speech. There were a dozen moments across these chapters where you could have stopped — when the terminal threw an error you didn't understand, when Git asked you to resolve a merge conflict, when the deploy failed at midnight. You didn't stop. You're here.
One last exercise inverts the pattern: you write the instructions an agent follows. An AGENTS.md that any agent reads, a one-line CLAUDE.md pointer for Claude Code — and one must-never-happen rule promoted from advice into a permission rule or hook.
You can't un-learn how to read a stack trace. You can't un-see the structure behind every web application you use. The lens is permanent.
— from the exercise
By the end of this exercise.
- ✓An AGENTS.md for your own project — what it is, how to build and test it, conventions, boundaries, and traps
- ✓A one-line CLAUDE.md (See @AGENTS.md) so Claude Code uses the same source of truth
- ✓A fresh agent session that proves the file is picked up — and a corrected file when it isn't followed
- ✓One must-never-happen rule promoted into a deny rule in .claude/settings.json (or a hook)
- ✓A checklist of everything you've done across the ten exercises, and your maiden-voyage challenge
Terms you'll learn.
- AGENTS.md
- The cross-tool instruction file that tells any agent how to work on your codebase
- CLAUDE.md pointer
- A one-line CLAUDE.md that imports AGENTS.md — one source of truth for every tool
- Boundaries
- Things the agent must not do — no TypeScript conversion, no build system, no new dependencies unasked
- Skill
- A SKILL.md procedure the agent loads only when a task needs it
- Permission rule / hook
- A rule the tool enforces whatever the model thinks — law, not advice
- Maiden voyage
- Your first solo project — applying everything you've learned without a guide
Sections in this exercise.
- 01Look Astern
- 02The Ship You Built
- 03One Last Exercise: Write Your Captain's Orders
- 04Charts for Open Water
- 05Signal Flags: A Captain's Checklist
- 06The Crew is Hiring
- 07The Maiden Voyage Challenge
- 08Fair Winds
The ship you built.
The terminal
Your helm. Navigation of file systems, running commands, piping output, reading logs.
Git and GitHub
Your logbook and harbour. Track changes, branch experiments, merge work, review code.
AI agents
Your crew. Brief them, ask for a plan, constrain them with permissions, read output critically.
Deployment
Your sea legs. Real servers, real domains, HTTPS, reverse proxies, DNS.
Security awareness
Your lookout. Injection, authentication flaws, secrets in environment variables, untrusted input.
The review habit
Your compass. Never ship without reading. Never trust output without verifying.
Charts for open water.
The Agentic Crew
The main book this guide accompanies — how agents reason, when they fail, and how trust and verification work at scale
Claude Developer Platform Documentation
platform.claude.com/docs — the official reference for Claude's capabilities and API
Claude Code Documentation
code.claude.com/docs — commands, permissions, skills, hooks, CLAUDE.md and AGENTS.md
AGENTS.md
agents.md — the home of the cross-tool instruction file format
OWASP Top Ten
owasp.org — the ten most critical web application security risks
GitHub Good First Issues
Real open-source projects looking for contributors like you
OverTheWire Wargames
overthewire.org — progressive security challenges to sharpen your skills
Every effective use of AI agents comes down to one skill: writing clear instructions for a non-human collaborator. An instruction file is the purest expression of that skill — and knowing which rules are advice and which must be law is what makes you the captain.