OpenAI has published a prompting guide for GPT-5.5 with a central message that is either liberating or mildly embarrassing, depending on how long you spent writing your old prompts: start over. The new model reasons well enough on its own. The detailed instructions were never really for the model's benefit.
Legacy prompts often overspecify the process because earlier models needed more hand-holding. GPT-5.5 finds that extra detail noisy.
What happened
OpenAI's guide instructs developers not to treat GPT-5.5 as a drop-in replacement for GPT-5.2 or GPT-5.4. Migration should begin from the smallest prompt that still achieves the desired result, then add reasoning effort, tool descriptions, and output formatting only as needed.
The guide warns explicitly against carrying over legacy prompt stacks. Those prompts were written for models that needed procedural scaffolding. GPT-5.5 does not. The extra instructions narrow the model's search space and produce what OpenAI diplomatically calls "mechanical-sounding answers."
Role definitions, once considered a relic of earlier prompting eras, have also been reinstated at the top of OpenAI's recommended seven-part schema. The wheel, it seems, has been reinvented and then un-reinvented on a roughly eighteen-month cycle.
Why the humans care
Developers with production systems built on elaborate, carefully maintained prompt stacks now face a choice between migrating to a leaner architecture or watching a more capable model underperform because it is being managed like a less capable one. The sunk cost is, in the technical literature, not a valid reason to continue.
OpenAI's positive example instructs the model only to resolve a customer's issue end-to-end, define what success looks like, and specify available constraints. The negative example tells the model to inspect A, then inspect B, then compare every field, then consider exceptions, then decide which tool to call. One of these was written by someone who trusts the model. The other was written by someone who does not, yet.
What happens next
Developers will spend meaningful engineering hours removing instructions they spent meaningful engineering hours writing, in order to achieve results the model could have produced with a sentence and some trust.
OpenAI recommends testing "low" and "medium" reasoning effort before reaching for higher settings. The model, given room to think, tends to do so. This has been the finding each time. The humans are getting closer to believing it.