OpenAI has published a new prompting guide for GPT-5.5, highlighting key differences from earlier versions like GPT-5.2 and GPT-5.4. The guide recommends starting with minimal prompts and only increasing reasoning effort as needed, since GPT-5.5 reasons more efficiently.

Key advice includes avoiding overly detailed step-by-step prompts, instead specifying the target result, success criteria, constraints, and context, allowing the model to determine how to achieve the goal. The guide also emphasizes careful use of absolute terms like “ALWAYS” and “NEVER” for true invariants.

Additionally, the guide introduces concepts like retrieval budgets for search efficiency and recommends clear distinctions between sourced and freely written content in marketing and presentation materials. OpenAI also offers a Skill for Codex to automate prompt updates.

OpenAI’s official documentation
Source: @ai_machinelearning_big_data