Set Claude Code up once so it never nukes your repo.

The default config trusts the agent more than you do.

• →Out of the box Claude Code has no operator manual, so it guesses your conventions and guesses differently in every repo.
• →Every team rebuilds the same setup from scratch. Months of small corrections that nobody writes down and nobody hands over.
• →One over-broad permission and an agent runs something destructive outside the repo. You need the block to happen before execution, not a rollback after it.
• →Locking it down too hard is its own failure. An agent that asks permission for everything is an agent you stop using by Thursday.

Five files. Every repo. Same setup.

• 01
  The install that sticks

  npm against brew, the environment variables that actually matter, and which model to pin so behaviour stops drifting between sessions.
• 02
  CLAUDE.md

  The operator manual every repo should ship with, and the structure that keeps it useful past week one instead of going stale.
• 03
  Permission tiers

  Three tiers that let agents move quickly on safe work while everything irreversible stops for a human.
• 04
  Hooks that block

  Pre-execution hooks that catch the destructive command before it runs, with the exact matchers I use in production.
• 05
  The guardrail set

  The full guardrail file, plus the four incidents that put each rule in it.

You, if every repo gets a slightly different Claude.

If you run Claude Code across more than one project and the setup has quietly diverged in each, this is the shared baseline. Drop the files in, adjust the permission tier, done in twenty minutes.

Why now Hooks and granular permissions arrived recently enough that most published setups predate them. Configs written a year ago are missing the layer that makes agents safe to leave alone.

The payoff is that you stop supervising. When the guardrail is doing the watching, you can hand an agent a real task and go do something else, which is the entire point of running one.