Good question. If your threat model is “Claude does something dumb I’ll regret” HAL is enough and way less friction (one tool does one thing)