Make Claude Code follow your rules with hooks

The problem

You put “never touch .env files” in your CLAUDE.md. Claude Code respects it for a week, then one long session it edits the file anyway. That is not a bug. CLAUDE.md, and every skill you write, is text the model reads and usually follows. Usually is not always. Rules that must hold every single time need enforcement, not a polite request, and that is exactly what Claude Code hooks are for.

When does a rule belong in a hook?

A hook is a shell command Claude Code runs automatically at a fixed point in its lifecycle: before a tool call, after an edit, when a session starts. The trigger is guaranteed, which gives you the whole split in four words: skills suggest, hooks enforce.

The test I use: if the rule needs no judgment and must fire every time, it is a hook. If Claude should reason about how to apply it, it belongs in CLAUDE.md or a skill. Run the formatter after every edit: hook. Never touch protected files: hook. “Prefer small components”: CLAUDE.md. “Here is our deploy checklist”: skill.

An instruction is a request. A hook is a guarantee.

The hook to copy

This blocks every edit to any .env file. Paste it into .claude/settings.json in your project (commit it and the whole team gets the guardrail), or into ~/.claude/settings.json to cover all your projects:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "jq -r '.tool_input.file_path // empty' | grep -q '\\.env' && { echo 'Blocked: .env files are protected' >&2; exit 2; }; exit 0"
          }
        ]
      }
    ]
  }
}

Every hook receives the event as JSON on stdin. This one pulls the file path of the pending edit and, if it contains .env, exits with exit code 2: the action is blocked and your stderr message is fed back to Claude so it adjusts instead of retrying. Run /hooks inside Claude Code to confirm it registered.

The formatter version is the same pattern after the edit instead of before it: a PostToolUse hook with the same matcher and the command jq -r '.tool_input.file_path' | xargs npx prettier --write.

When a rule needs a judgment call at a guaranteed moment, there is a middle ground: a hook with "type": "prompt" sends the event to a small Claude model that returns ok or a reason to block. Deterministic trigger, model decision.

Do this now

1. Pick the one rule Claude keeps breaking in your project.
2. Paste the hook above into .claude/settings.json and point the grep at your protected path.
3. Run /hooks to confirm it registered, then ask Claude to edit the file and watch the block land.