Requests versus enforcement

You'll finish with: You know which of your rules are actually holding.

Every rule you've written so far is a request. Anthropic's own documentation is blunt about it: these files are context, not enforced configuration. Claude reads them and tries to comply. There's no guarantee.

A request versus a rule

Instructiongets through

Hookblocked

The test

Ask of each rule: if this failed silently, what would it cost me?

Wrong tone in a draft → an instruction is fine
Money moved without approval → hook
Client emailed without review → hook
Production data deleted → hook

Plain English

A hook is a tiny script that runs automatically the moment Claude tries to do something. It answers one question — allow this, or stop it — before the action happens.

You want signs for most things and locks for four. Locking everything makes the setup unusable. Locking nothing makes it a liability.

Do this now

List every rule you've written so far — in CLAUDE.md, in your constitution.
Against each, answer: if this failed silently, what would it cost me?
Anything costing money, a client, or data goes on the hook list. Everything else stays an instruction.
You should end up with four or fewer. If you have twelve, you're over-locking.

Before you move on

You have a short written list of what needs a lock rather than a sign — and it's short.

All courses · Start here · Privacy