OpenAI's GPT-5.5 prompt guide has one instruction: stop micromanaging

OpenAI has released the GPT-5.5 prompt guidance, and its primary message is delightfully blunt: your old prompts are probably terrible now. The new documentation warns against dragging legacy prompt stacks into the new model. Where older systems required developers to hold their hand through every granular step of a process—inspect A, compare B, consider exceptions, call a tool—GPT-5.5 treats that level of over-specification as noise. Micromanaging the steps narrows the search space and forces overly mechanical answers. The new instruction is to describe the destination, define what available evidence looks like, set constraints, and then step back.

This means migrating to GPT-5.5 is not a festive find-and-replace exercise, but an audit of your prompt debt. Production prompts usually carry the scars of past failures: safety patches, workaround phrases, and mysterious paragraphs nobody deletes because a demo worked once after they were added. Treat GPT-5.5 as a clean slate. The guide also separates personality from collaboration style, advising that you dictate tone independently from how proactive the model should be. If you want a model to stop hallucinating, you define its stopping conditions, rather than just telling it to be a "highly accurate expert." The guide introduces the concept of explicit retrieval budgets, preventing the model from searching endlessly or making up details when evidence is missing. Ultimately, the era of the artisanal, four-page prompt preamble as a badge of honor is ending. The real test of the new model isn't what it can do with a massive prompt, but how much of your old prompt you can finally delete.