Model providers generally don't ask you to "throw away what you've built." But OpenAI has made an exception.
On April 26, OpenAI updated its official GPT-5.5 prompt guide with an unusually direct message: Don't just copy over your prompts from GPT-5.2 or 5.4. They aren't helping; they are holding GPT-5.5 back.
Why This Advice Is Unusual
Anyone familiar with the field knows that "prompt assets" are one of the biggest hidden costs when enterprises adopt large language models. You might have written hundreds of prompts for GPT-4, migrated them through 5.0, 5.2, and 5.4 with reasonable results, and now moving to 5.5 requires a rewrite—that's a real migration cost for businesses.
OpenAI is aware of this. Their wording this time is worth reading closely:
"Treat GPT-5.5 as a new model family to tune, not as a drop-in replacement for gpt-5.2 or gpt-5.4. Start over with the smallest prompt that gets the job done."
Translated, this means: Don't take shortcuts; re-evaluate your entire prompt stack.
Where Old Prompts Fall Short
OpenAI provided two very specific "anti-pattern" characteristics:
1. Over-specifying the process
Old prompts are often filled with instructions like "First do A, then check B, then call C, then output D." This was necessary because earlier models had weaker reasoning and needed step-by-step guidance. GPT-5.5 is different; give it the goal and let it choose the path.
OpenAI's comparison example:
| Approach | Effect |
| "Solve the customer's problem end-to-end" | Recommended |
| "First ask the question, second search the knowledge base, third..." | Not recommended |
The latter isn't wrong, but it limits the model's judgment space.
2. Forcing output schemas
This is especially relevant for developers. To ensure stable JSON output, prompts often include rigid instructions like "must include field X, field Y, field Z." GPT-5.5's advice: Keep it minimal; let the model organize the output—only hardcode when the product contract truly requires it.
Old "Role Definitions" Make a Comeback
Interestingly, over the past two years, the prompt engineering community has leaned toward treating role definitions like "You are an expert" as noise that is better removed.
OpenAI's new guide instead recommends a seven-part structure, with the first part being the role:
- Role
- Personality
- Goal
- Success criteria
- Constraints
- Output format
- Stop rules
How much detail for each? OpenAI suggests short and specific—for success criteria, don't write "user is satisfied"; write "problem is resolved and a next actionable step is provided."
A Detail: Default Reasoning Effort Changed
Another implicit change in GPT-5.5: the default reasoning effort has been changed from high to medium.
This means if you migrate your prompts directly, the model will "think" less than GPT-5.4 did. OpenAI recommends treating medium as the new baseline, only increasing it when medium proves insufficient. The reasoning behind this is latency and cost—GPT-5.5's medium is already smarter than GPT-5.4's high, so there's no need to default to high.
Retrieval Budget and Short Preamble
Two engineering details worth noting:
- Retrieval budget: For tasks requiring information lookup, explicitly tell the model the maximum number of searches. OpenAI's starting point is "use one broad search with short, distinctive keywords."
- Short preamble: In streaming output scenarios, having the model output a one- or two-word "acknowledgment" before a tool call makes the response feel faster to the user.
Both are designed to make long tasks feel smoother—not because the model is faster, but because the feedback rhythm to the user is faster.
Why This Guide Is So Direct
OpenAI's past prompt guides have mostly been "advisory." The GPT-5.5 guide has a much stronger tone. The reason is straightforward:
GPT-5.5 is priced the same as GPT-5.4, but its internal cost structure has changed. It responds best to short, sharp prompts and becomes less efficient with long, stacked prompts. If developers continue using old methods, OpenAI itself earns less.
In a way, this is the first time OpenAI has explicitly stated that "the prompt is part of the product." Models evolve, and not updating your prompt is like using an old version. Everyone has known this for years, but the product side has been vague about it, and no one has actively called it out.
Will this cause developer pushback? Probably. But once top customers run A/B tests and see data showing "accuracy up a few points, costs down a few percent after rewriting prompts," the tide will turn.
The next model upgrade will likely spark the debate: "Do we need to rewrite prompts again for GPT-5.6?"
Sources: OpenAI says old prompts are holding GPT-5.5 back and developers need a fresh baseline (The Decoder); GPT-5.5 prompting guide (simonwillison.net); CocoLoop; Using GPT-5.5 (OpenAI official documentation)