GPT-5.5 prompting guide

The Catalyst: New Model, New Rules GPT-5.5 has arrived, and OpenAI has simultaneously released its official prompting guide. What makes this significant isn't just another model version update, but OpenAI's explicit warning in the guide: treat GPT-5.5 as an entirely new model family, not a "drop-in replacement" for GPT-5.2 or GPT-5.4. This means the meticulously crafted prompts optimized for older models are likely to perform poorly, or even fail, on GPT-5.5. It颠覆了 many developers' expectation that "models only get better and old prompts will naturally be compatible," forcing everyone to rethink the fundamental approach to collaborating with AI. Deconstruction: Start with the 'Minimum Viable Prompt' OpenAI's core advice is: Start from scratch during migration. Specifically, they recommend beginning with "the smallest prompt that preserves the product contract." This is a crucial shift in mindset. The "product contract" refers to your application's core promise to the user, such as "accurately summarize documents" or "generate harmless code." In the past, we might have stacked detailed instructions, examples, and formatting requirements to constrain older models. Now, for GPT-5.5, OpenAI suggests retaining only the most essential instructions first, and then re-tuning aspects like "reasoning effort, verbosity, tool descriptions, and output format" based on that foundation. It's like assigning a task to a new colleague; you wouldn't just hand over the detailed manual you gave to your predecessor. Instead, you'd first clarify the core objective and then adjust the communication details based on the new colleague's characteristics. The guide also mentions a very practical interaction tip: for multi-step, time-consuming tasks, it's advisable to send a brief user-visible status update before calling tools (e.g., "I've received your request and am analyzing the data. The first step is..."). This seems simple but dramatically improves the user experience, avoiding the anxiety of users wondering "has the program crashed?" during waits. Simon Willison observed that OpenAI's own Codex app already uses this pattern with notable success. Trend Insight: Prompt Engineering Evolves from 'Tricks' to 'Systematic Migration' This reveals a deeper trend: prompt engineering is evolving from a collection of scattered "tricks" and "incantations" into an engineering activity requiring systematic methods. When models undergo generational updates, prompting strategies also need versioned migration. The fact that OpenAI provides a path to assist upgrades via the Codex tool ($openai-docs migrate this project to gpt-5.5) itself indicates that managing prompt configurations for different model versions is becoming part of the developer workflow. In the future, we might not only design prompts for different tasks but also maintain different prompt libraries for the same task across various model versions. Practical Value and Counter-Intuitive Insights For developers and product managers, the most immediate action point is: Do not assume old prompts remain effective on new models. When evaluating or migrating to GPT-5.5, you must allocate time for re-validating and re-tuning prompts. You can use OpenAI's upgrade guide and tools as a starting point, but the core is to adopt the mindset of "starting with the smallest prompt and iterating step by step." The most counter-intuitive aspect here is that as models become more capable, the demand for "fundamental skills" in prompt engineering might actually increase. GPT-5.5 might be "smarter," but it could also be more "sensitive," potentially having lower tolerance for ambiguity and redundant information in instructions. The crude prompting style of the past, which relied on stacking details to "coax" results out of the model, may become less effective than concise, precise instructions on the new model. This forces us to return to the essence of communication: clearly defining objectives and adjusting expression based on the counterpart's characteristics.