Upgrade to GPT-5.5 but still using old checklist prompts? Shift from commanding every step to defining a clear results contract — what done looks like, evidence rules, and when to stop — so the model delivers, not just tries. What's one prompt habit you'd change today?