OpenAI: Old Prompts Hold Back GPT-5.5

OpenAI Recommends New Prompts for GPT-5.5

OpenAI has recently released a guide for using its latest model, GPT-5.5, in which it strongly advises against reusing old prompts. Instead, the company recommends starting with minimal, outcome-oriented instructions. This approach aims to leverage the model's enhanced capabilities, which operate more efficiently with less prescriptive guidelines.

Old prompts, often overly detailed, can indeed limit the performance of GPT-5.5. This model is designed to function more effectively with succinct instructions, thereby avoiding noise and mechanical responses generated by superfluous details. OpenAI emphasizes that process specifications inherited from previous models can restrict the model's search space, which is not optimal for GPT-5.5.

A Seven-Part Framework for Complex Cases

For more complex use cases, OpenAI proposes a seven-part framework that begins with a clear definition of the role. This method encourages users to rebuild their prompts from scratch, rather than relying on methods inherited from earlier versions like GPT-5.2 or GPT-5.4. Transitioning to GPT-5.5 should start from zero, with the smallest possible prompt to achieve the goal. Only after that should developers adjust reasoning effort, scope, tool descriptions, and output format.

The guide explicitly warns against transferring every instruction from old prompt stacks. Legacy prompts often specify the process too much, as previous models required more guidance. With GPT-5.5, this additional detail creates noise, restricts the model's search space, or produces mechanical responses. Instead, the prompt should specify the target outcome, success criteria, constraints, and available context, then let the model determine how to achieve it.

Examples of Effective and Ineffective Prompts

A positive example from the guide is a customer service prompt that only defines the objective: to resolve the customer's issue end-to-end, make eligibility decisions based on available policy and account data, complete any authorized actions before responding, and include in the final response the completed actions, the customer's message, and any obstacles. If evidence is lacking, it is advised to ask for the smallest missing field.

In contrast, a negative example micromanages every step: first inspect A, then inspect B, compare each field, consider all possible exceptions, decide which tool to call, call the tool, and explain the entire process to the user. Absolute rules using words like "ALWAYS" or "NEVER" should be reserved for true invariants such as safety rules or required output fields. For decisions, OpenAI recommends using judgment rules instead.

Role Definitions Make a Comeback

The prompt community has debated the usefulness of role definitions in newer models. Some had deemed them unnecessary or even counterproductive. The GPT-5.5 guide argues the opposite: the recommended prompt structure begins with a role definition and context. This includes elements such as tone, behavior, and collaboration style, as well as the visible outcome for the user.

For customer-facing assistants, support flows, or coaching tools, the guide recommends separating two distinct dimensions in this framework: personality and collaboration style. Personality covers the assistant's tone: warmth, formality, or humor. Collaboration style covers how it operates, when to ask questions, when to make assumptions, and how to handle uncertainty.

OpenAI provides two contrasting examples. First, a factual and task-oriented personality block: you are a capable, accessible, stable, and straightforward collaborator. Assume the user is competent and acting in good faith, and respond with patience, respect, and practical help. Prefer to make progress rather than stop for clarification when the request is already clear enough to attempt. Use context and reasonable assumptions to move forward. Ask for clarifications only when the missing information would materially change the response or create significant risk, and keep any questions narrow.

And a more expressive and collaborative style: adopt a lively conversational presence, intelligent, curious, playful when appropriate, and attentive to the user's thoughts. Ask good questions when the problem is unclear, then become decisive once there is enough context. Be warm, collaborative, and polite. The conversation should feel easy and lively, but not chatty for its own sake. Offer a genuine perspective rather than simply reflecting the user, while remaining responsive to their goals and constraints.

Each section should remain brief. Details should only be added where they genuinely alter behavior, according to OpenAI, and the prompt structure should be seen as a starting point, not a rigid template.

Establishing Retrieval Budgets and Citation Rules

For fact-based responses, the citation behavior should be part of the prompt itself. Developers should specify which claims require evidence, what counts as sufficient evidence, and how the model should respond when evidence is lacking. A lack of evidence should not automatically translate to a factual "no." The guide describes retrieval budgets that act as stop rules for searches:

For ordinary Q&A, start with a broad search using short, discriminating keywords. If the main results contain enough supporting citations for the primary request, respond based on those results rather than searching again.

Make a new retrieval call only when:

• The main results do not answer the primary question.
• A required fact, parameter, owner, date, ID, or source is missing.
• The user has requested exhaustive coverage, a comparison, or a complete list.
• A specific document, URL, email, meeting, recording, or code artifact needs to be read.
• The response would otherwise contain a significant unsupported factual claim.

Do not search again to improve wording, add examples, cite non-essential details, or support a formulation that can be safely made more generic.

For writing tasks like presentations, summaries, or marketing texts, OpenAI recommends drawing a clear line in the prompt between claims that require sources and parts that can be written more freely:

• Use retrieved or provided facts for concrete claims about products, customers, metrics, roadmaps, dates, capabilities, and competitiveness, and cite these claims.
• Do not invent specific names, first-party data claims, metrics, roadmap status, customer outcomes, or product capabilities to bolster the draft.
• If there is little or no supporting citation, draft a useful generic outline with placeholders or clearly labeled assumptions rather than unsupported specifics.

Preemptive Updates to Reduce Perceived Latency in Streaming

In streaming applications, every second before the first visible response counts. GPT-5.5 can spend considerable time reasoning, planning, or calling tools before any text appears. For longer or tool-heavy tasks, the guide recommends a short "preemptive update": a visible acknowledgment that confirms the request and names the first step. This improves perceived responsiveness without changing the underlying task.

Before any tool call for a multi-step task, send a brief visible update to the user that acknowledges the request and states the first step. Limit this to one or two sentences.

Developers who do not wish to manually rewrite prompts can delegate the work to Codex, according to OpenAI. The coding agent can apply the guide's changes with a single command. OpenAI has released its own "OpenAI Docs Skill" for this task, which also works with other coding agents.