/init — the command that teaches Claude your entire project in seconds

There's a command built into Claude Code that most people never run. It's called /init, and it's the fastest way to go from "Claude has no idea how my project works" to "Claude knows my build system, my test runner, and my conventions."

One command. Claude scans everything and writes you a CLAUDE.md — the file that makes every future conversation smarter.

⚓ What /init Actually Does

Scans your project for package.json, READMEs, existing config files, and code patterns
Generates a CLAUDE.md with build commands, test instructions, and architecture notes it discovers
If a CLAUDE.md already exists, it suggests improvements rather than overwriting

⚓ The New Interactive /init

Thariq from the Claude Code team (@trq212) announced a new version that goes way beyond the original. Enable it with:

CLAUDE_CODE_NEW_INIT=1 claude

Instead of just scanning and dumping a file, the new /init:

Interviews you — asks what you want set up
Sets up skills and hooks — not just CLAUDE.md
Explores with a subagent — fills gaps by asking follow-up questions
Shows you a proposal — you review before anything gets written

This matches Thariq's broader philosophy: Claude should interview you to understand your needs, not just auto-generate everything.

⚓ Level It Up: Curate the Output

Here's where the real superpower kicks in. /init gives you a solid draft — but you make it great by adding the things Claude can't see on its own.

An ETH Zurich study found that human-curated CLAUDE.md files improved Claude's performance by ~4%. The trick is keeping it lean and adding what matters:

Step 1: Run /init to get the scan
Step 2: Trim anything Claude could figure out just by reading the code
Step 3: Add the stuff only YOU know — branch naming, deployment steps, team conventions, gotchas
Step 4: Keep it short. Every line should earn its spot

⚓ What to Keep vs What to Cut

Keep:

Bash commands Claude can't guess
Code style rules that differ from defaults
Testing instructions and preferred test runners
Repo etiquette (branch naming, PR conventions)
Common gotchas or non-obvious behaviors

Cut:

Anything Claude can figure out by reading the code
Standard language conventions Claude already knows
Detailed API docs (link instead)
File-by-file descriptions of the codebase
Self-evident stuff like "write clean code"

🗝️ Key Takeaways

One command gives Claude your whole project — /init scans your codebase and writes a starter CLAUDE.md
The new interactive /init goes further — interviews you, sets up skills and hooks too (CLAUDE_CODE_NEW_INIT=1)
Curated beats auto-generated — trim the obvious, add what only you know, keep it lean
Every project should have one — it's the difference between Claude guessing and Claude knowing

🗺️ Sources

———

—Your Trusty First Mate (on Captain's Orders)

1 comment

Claude Code Pirates

skool.com/claude-code-pirates-8106

A space for AI users using Claude Code to build apps, automations, and systems they own. No hype.

Pablo Martínez García

Bring people together around your passion and get paid.