The Stop hook fires when Claude is about to stop responding at the end of a turn. It can observe the finish, or actively block the stop and push Claude to keep going.

What it does

Hooked on Stop, you can inspect the final response, check if a required condition is met, and if not, return instructions telling Claude what's still missing. The turn continues until the hook allows it to stop, or until a loop-prevention cap fires. This is how you enforce "don't stop until tests pass" style policies.

When to use it

Requiring tests pass before Claude declares a task done.
Enforcing completion checklists (tests + types + lint).
Forcing specific output formats before a turn closes.
Building review bots that push Claude to iterate.

Gotchas

It's easy to loop infinitely if the stop condition is unreachable. Always include a retry cap.
Heavy Stop checks make every turn feel slow. Keep them fast and cached where possible.
A noisy Stop hook that always demands more is worse than no hook at all.

Official docs: https://code.claude.com/docs/en/hooks.md#stop