Doesn't it just get attenuated and start ignoring those commands?

The hook sends the text to another agent/context with a request to validate and return a good or bad + reason response. Every request is a fresh context.

https://code.claude.com/docs/en/hooks-guide#agent-based-hook...