The Claude Code Ninja Manual

THE CLAUDE CODE

NINJA 外伝

MANUAL

Eleven stages from your first prompt to your first shipped artifact.

► PRESS START ◄

TABLE OF CONTENTS

Eleven stages. Two scrolls of reference.

STAGE 01 Orient — CLAUDE.md is the steering wheelPG 03
STAGE 02 Cheatsheet — the 7 jargon terms PG 04
STAGE 03 Git Discipline — commit = save PG 05
STAGE 04 GitHub — the social layer PG 06
STAGE 05 Build a Skill — /wrap PG 07
STAGE 06 Build a Hook — SessionStart PG 08
STAGE 07 Chrome MCP — drive a real browser PG 09
STAGE 08Gmail MCP — the bridge email PG 10
STAGE 09Compounding Memory — the typed-wiki PG 11
STAGE 10Deploy — Vercel from GitHub PG 12
STAGE 11Targeted Research — the boss move PG 13
SCROLL AJargon & Glossary PG 14
SCROLL BCommands & Hook Events PG 15

STAGE 01

Orient — CLAUDE.md is the steering wheel

Change one line in CLAUDE.md. Change Claude.

What: Edit CLAUDE.md, restart Claude, watch behavior shift, undo.

Why: Every project lives or dies by its CLAUDE.md. No file = generic assistant. Good file = your collaborator.

Time: ~20 minutes.

Exit: One round-trip edit and a clean undo. Behavior shift observed.

Watch-out: Long CLAUDE.mds cost tokens. Most projects want 30 to 100 lines.

Open CLAUDE.md. Add a bullet to Voice + posture: End every response with "Aboveboard." Save. /exit. claude. Ask anything. Watch the word land. Now undo.

STAGE 02

Cheatsheet — the 7 jargon terms

Front-load the vocabulary. No more Googling mid-stage.

What: Run the /cheat skill. Walk all seven terms.

Why: The next nine stages throw these words at you. Knowing them on the way in makes every later stage feel familiar.

Time: ~15 minutes.

Exit: You can say out loud what makes a skill different from a hook.

The seven: skill, hook, sub-agent, MCP, context, slash command, .claude/.

Type /cheat. Walk the seven. When you can answer "skill vs hook?" out loud, you are done.

STAGE 03

Git Discipline — commit = save

You cannot break anything that cannot be un-broken.

What: Make a small edit. Stage. Commit. Read the log.

Why: Git is the save system. Trust it, you move fast. Distrust it, you tiptoe around every change.

Time: ~30 minutes.

Exit: git log shows your commit at the top with your name and a real message.

Watch-out: On Windows, paths with spaces need quotes. Always. git add "my file.md".

Edit README.md. Add your name and today's date. git add README.md. git commit -m "first commit from <your name>". git log. You are in the system now.

STAGE 04

GitHub — the social layer

Every paid software you will ever use has this ritual. Learn it once.

What: GitHub account. Install gh CLI. gh auth login. (If you have a private companion repo, clone it.)

Why: Git was local. This is the network. Auth, private repos, the developer login ritual you will see again and again.

Time: ~30 to 60 minutes.

Exit: gh auth status shows you are logged in. me.json.github_username is set.

Watch-out: "gh command not found" after install means a PATH issue, not a broken install. Tell Claude your OS.

Run gh auth login. Pick HTTPS. Pick browser. Paste the code. Approve. gh auth status. Tell Claude your username.

STAGE 05

Build a Skill — /wrap

When a prompt works, save it as a skill. Stop re-typing.

What: Author /wrap — a three-question session-summary skill that saves to wraps/<date>.md.

Why: Skills are the unit of team reuse. They embed prompt-shape, output-format, and file-naming so every fire is consistent.

Time: ~30 minutes.

Exit: .claude/skills/wrap/SKILL.md exists. You ran /wrap after restart. A real wrap file lives at wraps/<today>.md.

Watch-out: Claude does not see new skills until you restart. /exit then claude after creating a skill.

Create .claude/skills/wrap/SKILL.md. Paste the template from the stage file. /exit. claude. Fire /wrap. Answer three questions. Watch the file land.

STAGE 06

Build a Hook — SessionStart

Hooks fire on events. You do not trigger them. Claude Code does.

What: Wire a SessionStart hook that reads your latest wrap and surfaces it on every new session.

Why: Solves the worst Claude Code problem — amnesia between sessions. Your future Claude wakes up knowing what you did yesterday.

Time: ~35 minutes.

Exit: Script exists in .claude/hooks/. Referenced in .claude/settings.json (project-shared, ships with the kit — not settings.local.json). Restart shows "Picking up from last session."

Watch-out: Schema mismatch is the #1 silent failure. The hook block needs three-level nesting: event array → matcher group → hooks array → handler. A flat {"command": "..."} looks plausible but will not fire.

Write the script. Wire the hook in .claude/settings.json under SessionStart using the three-level schema. Restart Claude. Your last wrap surfaces. Done.

STAGE 07

Chrome MCP — drive a real browser

WebFetch sees what is public. Chrome MCP sees what you log into.

What: Install Chrome MCP. Verify a public read. Read an auth-gated page that WebFetch cannot.

Why: Auth-gated pages, JS-heavy SPAs, interactive form-fills. The web you actually use, not the public scraps.

Time: ~30 minutes.

Exit: Claude reads a real auth-gated page (your LinkedIn profile works). Reports specific content. No hallucination.

Watch-out: Cookie popups block screenshots. Click Accept first. LinkedIn company pages are screenshot-timeout-prone — use the text extractor.

Install the Chrome extension. Wire the MCP config. Restart. Tell Claude to read your LinkedIn and summarize your headline + recent posts. Verify it matches reality.

STAGE 08

Gmail MCP — the bridge email

Read the inbox. Draft. Send. The bridge into real comms.

What: Wire Gmail MCP. Authenticate. Send the bridge email to your engagement partner (or a test email to yourself, if solo).

Why: Most useful AI work routes through email. Drafting replies, inbox triage, weekly status, outbound. All compose from this MCP.

Time: ~25 to 40 minutes.

Exit: Bridge email sent with your GitHub username + private repo URL. (Solo path: test email landed in your own inbox.)

Watch-out: OAuth flow blocked by pop-up blockers. Disable for the auth tab and retry.

Add the Gmail MCP server config to .claude/settings.json under mcpServers. Run /mcp. Authenticate Gmail. Verify with a search read. Then ask Claude to compose the bridge email. Confirm the draft. Send.

STAGE 09

Compounding Memory — the typed-wiki

Wraps are chronological. Memory is structured. Both compound.

What: Write your first typed memory file (person, project, decision, or concept). @import it from CLAUDE.md.

Why: Wraps tell you about time. Memory tells you about an entity. Together they make Claude's knowledge accrete instead of evaporate.

Time: ~40 minutes.

Exit: memory/<type>/<name>.md exists with real content. CLAUDE.md imports it. Claude can answer about that entity after restart.

Watch-out: Decisions are append-only. Never edit a past decision's body. Write a new dated file and reference the old.

Pick one real entity. Write the file with YAML frontmatter. Add @memory/<type>/<name>.md to CLAUDE.md. Restart. Ask Claude about it.

STAGE 10

Deploy — Vercel from GitHub

Commit. Push. The world can see it.

What: Write a graduation page. Commit. Push. Connect Vercel. Visit the live URL.

Why: Every modern web project uses this loop. Vercel, Netlify, Cloudflare Pages, GitHub Pages, AWS Amplify. Same shape. Different brand.

Time: ~30 minutes.

Exit: graduation.html lives at your Vercel URL. curl returns 200. You would not be embarrassed to share the link.

Watch-out: 404 usually means you forgot to push. Run git status to verify a clean tree before blaming Vercel.

Write graduation.html. git add, git commit, git push. Connect Vercel to your fork. Visit the URL. curl -s -o /dev/null -w "%{http_code}\n" <url> → 200.

STAGE 11

Targeted Research — the boss move

Do not ask one agent five questions. Spawn five agents on one question each.

What: Pick a real topic. Decompose into 3 to 5 sub-questions. Spawn N agents in parallel. Synthesize into one polished HTML page. Deploy.

Why: The pattern that separates "Claude is helpful" from "Claude is the engine behind serious analytical work." Competitive briefs, prospect scans, due-diligence, all the same shape.

Time: ~45 to 60 minutes.

Exit: N agent files in research/. One polished HTML synthesis deployed. URL returns 200.

Watch-out: Do not over-brief. Do not under-verify. The synthesis is in YOUR voice, not the agents'.

Pick a topic you actually want to know about. Write a tight per-agent brief. Spawn N. Read their files. Synthesize. Deploy via Stage 10's pipeline. Share the URL.

The words. What they mean.

CLAUDE.MDThe project file that shapes Claude's behavior. The steering wheel.
SKILLA saved prompt invoked with /<name>. Lives in .claude/skills/<name>/SKILL.md.
HOOKCode that fires automatically on session events. SessionStart, PreToolUse, Stop.
SUB-AGENTA Claude spawned by your Claude. Its own context window. Used for parallel reads.
MCPModel Context Protocol. A bridge that adds tools Claude can use (Chrome, Gmail, others).
CONTEXTEverything Claude can see right now. CLAUDE.md + open files + this conversation.
SLASH COMMAND/ + name. Fires a skill or a built-in command.
.CLAUDE/Project-local folder for skills, hooks, settings. Lives in the repo root.
MEMORYTyped files Claude reads at session start to compound knowledge across sessions.
REPOA folder tracked by git. Has a hidden .git/ inside.
ME.JSONYour local progress file in this kit. Gitignored. Tracks what stage you are on.
WRAPEnd-of-session summary file. Lives in wraps/<date>.md. Built in Stage 05.

The slash arsenal. The keys. The events.

CLAUDEStart (or restart) a Claude Code session in the current folder.
/EXITLeave Claude. Type claude again to come back.
/CLEARWipe the current session. Start fresh.
/HELPShow Claude Code help.
/MCPList MCP servers and their auth status.
/CHEATWalk the 7 jargon terms. Stage 02's first skill.
/WRAPThree-question session summary. Saves to wraps/<date>.md. Built in Stage 05.
ESCInterrupt whatever Claude is doing right now.
GH AUTH LOGINAuthenticate the GitHub CLI. Browser flow. Stage 04.
SESSIONSTARTHook event. Fires when a Claude session opens. Used in Stage 06.
USERPROMPTSUBMITHook event. Fires every time you hit enter on a prompt.
PRETOOLUSE / POSTTOOLUSEHook events. Fire before or after Claude uses a tool (Bash, Edit, Write).
STOPHook event. Fires when the session ends.