Specifying global rules for the AI (collected from experienced practitioners and refined by observing what the AI repeatedly gets wrong) keeps the AI 'on the rails' by constraining its behavior to do things a certain way and avoid known failure modes.