Capstone: set up your own project end to end

The steps

One step at a time. When a step’s check matches, tick it and the next one opens.

1. 01

   Step 1 of 5: Write the CLAUDE.md

   Let `/init` draft it, then cut it down.

   Start Claude Code in your repository and run /init. If a CLAUDE.md already exists, it suggests improvements instead of overwriting. Keep the file under 200 lines.

   In Claude Code

   – What happens: Started with claude in your repository.

   You type: › /init

   Check it worked

   CLAUDE.md is at the repository root, under 200 lines, and /context lists it with the memory files.

   It worked — next step Not yet
2. 02

   Step 2 of 5: Set permission rules

   Allow your test command and deny reading secrets.

   Put these rules in .claude/settings.json, using your own test command. The file is shared with your team when you commit it.

   .claude/settings.json

   permissions.allow: ["Bash(npm test)"]

   Runs your tests without asking

   permissions.deny: ["Read(./.env)", "Read(./.env.*)"]

   Claude’s file tools can’t read your env files

   Check it worked

   /permissions shows both rules.

   It worked — next step Not yet Back
3. 03

   Step 3 of 5: Add one skill and one subagent

   Ask Claude to write both, then read what it wrote.

   Paste into Claude CodeCopy

   Look at how we work in this repository, then create two files. Show me each one before you save it.
   
   1. A project skill in .claude/skills/ for the task I repeat most: <describe the task>.
   2. A project subagent in .claude/agents/ that reviews my uncommitted changes for bugs. It may read files but must not edit them, and it reports findings with file and line.

   your-repo/

   • folder .claude
     • folder skills
       • folder your-skillNew
         • SKILL.mdNew
     • folder agents
       • reviewer.mdNew

   Check it worked

   /skills lists your skill, and /context lists the subagent with the file it loaded from.

   If it didn’t work

   The subagent doesn’t show up.

   Restart Claude Code. A running session doesn’t notice a newly created agents folder.

   It worked — next step Not yet Back
4. 04

   Step 4 of 5: Add a hook that runs your tests

   Make the test run a guarantee, not a request.

   This is the Stop hook from Day 14, pointed at your own tests.

   Paste into Claude CodeCopy

   Add a Stop hook to .claude/settings.json that runs <your test command> when you finish a reply.
   - If the tests fail, the hook exits with code 2 and prints the failures to stderr, so you go back and fix them.
   - If stop_hook_active is true in the hook's JSON input, it exits 0, so it can't loop.
   Show me the script and the settings change before you save them.

   In order:

   1. Claude

      Claude finishes a reply

   2. Hook

      The Stop hook runs your tests

   • If the tests fail

     Claude keeps working

     The failures go back to it

   • If they pass

     Claude stops

   Check it worked

   /hooks lists a Stop hook. Break the code on purpose, ask any question, and Claude goes back to fix it.

   It worked — next step Not yet Back
5. 05

   Step 5 of 5: Run the checkup

   `/doctor` finds what you missed.

   It flags invalid settings, unused extensions and CLAUDE.md lines Claude could work out from the code.

   In Claude Code

   You type: › /doctor

   – What happens: Claude proposes fixes and applies them only after you confirm.

   Check it worked

   No invalid settings files are reported, and git status shows your new files ready to commit.

   It worked — finish Not yet Back