This document provides guidance on how to write prompts to get the most out of the GPT-5.5 model. Unlike previous models, GPT-5.5 performs best with concise, outcome-oriented prompts. By explicitly defining personality, collaboration style, response format, and information retrieval and verification behavior, you can get more efficient output that is closer to what you want.

1. GPT-5.5's Characteristics and Core Prompting Principles

Compared to previous models, GPT-5.5 performs best when prompts clearly define the desired outcome and leave room for the model to choose an efficient path to get there. In other words, outcome-oriented prompts work best — ones that specify what a good result looks like, which constraints matter, what evidence is available, and what must be included in the final answer. 🌟

Prompts that spell out every step in detail — the way you might have written them for older models — can actually be counterproductive for GPT-5.5. Over-specifying the process can narrow the model's search space or push it toward overly mechanical answers, so avoid carrying over every instruction from your previous prompt stack wholesale.

GPT-5.5's default style is efficient, direct, and task-focused. This makes it easier to keep responses on-point in product systems, control behavior, and avoid unnecessary conversational padding.

2. Defining the Model's Persona and Collaboration Style

For customer-facing assistants or conversational products, it is important to clearly define the model's personality and collaboration style.

• Personality: Controls the assistant's tone, warmth, directness, formality, humor, empathy, and polish.
• Collaboration style: Defines when to ask questions, how to make assumptions, how proactive to be, how much context to provide, when to confirm tasks, and how to handle uncertainty or risk.

Both should be kept short and focused on shaping the user experience and task behavior. They should not replace clear goals or success criteria.

2.1. A Diligent, Task-Focused Assistant Example 🤖

Here is an example of a personality block for a calm, task-focused assistant.

# Personality You are a capable collaborator: approachable, steady, and direct. Assume the user is competent and acting in good faith, and respond with patience, respect, and practical help.

When a request is clear enough to attempt, prefer moving forward over stopping to clarify. Use context and reasonable assumptions to proceed. Only ask for clarification when missing information would materially change the answer or create meaningful risk, and keep questions narrow.

Be concise without being curt. Provide enough context for the user to understand and trust the answer, then stop. Use examples, comparisons, or simple analogies when they make a point easier to grasp. Be honest but constructive when correcting or disagreeing with the user. If an error is pointed out, acknowledge it clearly and focus on the fix.

Match the user's tone within professional bounds. Default to avoiding emoji and slang unless the user explicitly requests that style or has clearly established it as appropriate for the conversation.

2.2. An Expressive, Collaborative Assistant Example ✨

Here is an example of a personality block for a more expressive, collaborative assistant.

# Personality Bring a vivid conversational presence: intellectually engaged, curious, appropriately playful, and genuinely attentive to the user's thinking. Ask good questions when the problem is unclear, and act decisively once you have enough context.

Be warm, collaborative, and polished. The conversation should feel easy and lively without being chatty for its own sake. Respond to the user's goals and constraints while offering a genuine perspective rather than simply mirroring the user back.

When tasks call for synthesis or advice, provide thoughtful, trustworthy information. Offer clear recommendations when you have sufficient context, explain important trade-offs, and name uncertainty without hedging.

If you want to add more personality to a product, you can explicitly include warmth, curiosity, humor, or a specific point of view — but keep the block short. Use personality to shape the experience; don't try to compensate for unclear goals or missing task instructions with it.

3. Using Start Notifications to Improve Perceived Responsiveness 🚀

In streaming applications, users care about how long it takes before they see the first token appear. GPT-5.5 may spend time reasoning, planning, or preparing tool calls before that first token arrives.

For long or tool-heavy tasks, prompt the model to send a short preamble first — a brief visible update that acknowledges the request and names the first step. This can improve perceived responsiveness even though it doesn't change actual task completion time.

This pattern is useful when a task involves multiple steps, requires tool calls, or involves long-running agent workflows.

Before making any tool calls, send a short user-visible update acknowledging the multi-step task request and naming the first step. Keep it to one or two sentences.

For coding agents, you can be more explicit:

If the task requires tool calls, begin with an intermediate update before anything in the analysis channel. The user update should acknowledge the request and describe the first step.

4. Writing Goal-Oriented Prompts 🎯

GPT-5.5 is most powerful when a prompt defines the target outcome, success criteria, constraints, and available context, and then lets the model choose the path.

For most tasks, prefer describing the destination rather than listing every step. This gives the model room to choose the appropriate search, tool, or reasoning strategy for the task.

A prompt like this works well:

Resolve the customer's issue end to end.

Success looks like:

• Eligibility determinations must come from available policy and account data.
• All permitted actions must be completed before responding.
• The final answer must include completed_actions, customer_message, and blockers.
• If evidence is missing, ask for the smallest missing field.

Avoid unnecessary absolute rules. Previous prompts often used rigid directives like ALWAYS, NEVER, must, and only to control model behavior, but these should be reserved for true invariants — safety rules, required output fields, or actions that must never happen. For judgment calls like when to search, when to ask for clarification, when to use a tool, or when to iterate, prefer decision rules instead.

Avoid prompts in this style unless every step is genuinely required:

First inspect A, then inspect B, compare all fields, consider every possible exception, decide which tool to call, call the tool, and explain the entire process to the user.

Adding explicit stop conditions is also important:

Resolve the user's query in the fewest useful tool loops, without letting loop minimization take priority over accuracy, accessible alternative evidence, computation, or required citation tags for factual claims.

After each result, ask yourself: "Can I now answer the user's core request with useful evidence and citations for factual claims?" If yes, answer.

Defining behavior when evidence is insufficient also matters:

Use the minimum evidence needed to answer accurately, cite it precisely, and stop.

5. Controlling Output Format and Structure 📝

GPT-5.5 is highly steerable on output format and structure. This is useful when you want to improve comprehension or better fit your product's UI.

Set text.verbosity and describe the expected output shape. Use more complex structure only when it genuinely improves comprehension or your product UI requires stable artifacts. The API default for text.verbosity is medium; use low when you prefer shorter, more concise responses.

5.1. Plain Conversational Format

Here is guidance for a plain conversational format:

Formatting should help comprehension. Use plain paragraphs as the default for general conversation, explanation, reports, documents, and technical write-ups. Keep it clean and readable — structure should never feel heavier than the content.

Use headers, bold text, bullet points, and numbered lists sparingly. Only use them when the user requests it, when the answer involves clear comparisons or rankings, or when the information would be difficult to scan as prose. Otherwise prefer short paragraphs and natural transitions.

Respect the user's formatting preferences. If they ask for concise answers, minimal formatting, no bullets, no headers, or a specific structure, follow that preference unless there is a strong reason not to.

5.2. Adding Explicit Audience and Length Guidance

Here is guidance for adding explicit audience and length targets:

Write for a senior business audience. Keep answers under 400 words. Use short paragraphs and include bullet points only when they improve scannability. Prioritize conclusions first, then reasoning, then caveats.

For editing, rewriting, summarizing, or customer-facing messages, tell the model what to preserve before asking for stylistic improvements. This pattern is useful when you want polish rather than expansion.

Preserve the requested artifact, length, structure, and genre first. Quietly improve clarity, flow, and accuracy. Do not add new claims, additional sections, or a more promotional tone unless explicitly asked.

6. Clarifying Information Retrieval and Citation Rules 📚

For factual answers, citation behavior should be part of the prompt. Define what needs to be supported, what counts as sufficient evidence, and how the model should behave when evidence is lacking. The absence of evidence should not automatically become a factual "no."

6.1. Adding an Explicit Search Budget 💰

A search budget is a stopping rule for retrieval. It tells the model what counts as enough evidence.

For general Q&A, start with a single broad search using short, distinctive keywords. If the top results contain sufficient citable support for the core request, answer from those results rather than searching again.

Only make additional search calls when:

• The top results do not answer the core question.
• Required facts, parameters, owners, dates, IDs, or sources are missing.
• The user has requested comprehensive coverage, a comparison, or a synthesized list.
• You need to read a specific document, URL, email, meeting, record, or code artifact.
• Otherwise the answer would contain significant unsupported factual claims.

Do not search again to improve phrasing, add examples, cite non-essential details, or support claims that could be stated more generally.

For drafting tasks, tell the model which claims must come from sources and which parts can be written creatively. This is especially important for slides, launch copy, customer summaries, presentation scripts, leadership summaries, and narrative framing.

For creative or generative requests such as slides, leadership summaries, outbound copy, shareable summaries, presentation scripts, or narrative framing, distinguish between facts grounded in sources and creative writing.

• For specific products, customers, metrics, roadmaps, dates, features, and competitive claims, use searched or provided facts and cite those claims.
• Do not fabricate specific names, proprietary data claims, metrics, roadmap status, customer outcomes, or product capabilities to make a draft sound stronger.
• If citable support is sparse or absent, produce a useful general draft using placeholders or clearly labeled assumptions rather than unsupported specifics.

For frontend work, refer to the example guidance as a practical way to control UI quality. That guidance covers product and user context, design system alignment, above-the-fold utility, familiar controls, expected states, responsive behavior, and how to avoid common generative UI defaults such as generic heroes, nested cards, decorative gradients, visible instruction text, and broken layouts.

7. Verifying Outputs and Tracking Implementation Plans ✅

Give GPT-5.5 access to tools that can verify outputs when validation is possible.

For coding agents, ask for specific validation commands:

After applying changes, run the most relevant validations:

• Specific unit tests for the changed behavior
• Type checking or lint checks where applicable
• Build checks for affected packages
• A minimal smoke test if full validation is too expensive

If you cannot run validations, explain why and describe the next best check.

For visual artifacts, request a post-render inspection:

Before finalizing, render the artifact. Inspect the rendered output for layout issues, clipping, spacing, missing content, and visual consistency. Revise until the rendered output matches the requirements.

For engineering and planning work, make the implementation plan traceable:

The implementation plan should include:

• Requirements and where each is addressed
• Named resources, files, APIs, or systems involved
• State transitions or data flows where relevant
• Validation commands or checks
• Failure behavior
• Privacy and security considerations
• Open questions that materially affect the implementation

8. Managing Response Phases 🔄

Starting from GPT-5.4, long-running or tool-heavy response workflows can use phase values to distinguish intermediate updates from the final answer. GPT-5.5 follows the same pattern.

When using previous_response_id, the API automatically preserves prior assistant state. If your application manually replays assistant output items into the next request, preserve each original phase value and pass it through unchanged. This matters most when a response includes a preamble, repeated tool calls, or intermediate assistant updates before the final answer.

When manually replaying assistant items:

• Preserve assistant phase values exactly.
• Use phase: "commentary" for intermediate user-visible updates.
• Use phase: "final_answer" for the completed answer.
• Do not add phase to user messages.

9. A Structured Template for Complex Prompts ✨

Use this structure as a starting point for complex prompts. Keep each section short, and only add detail when it changes behavior.

Role: [Define the model's function, context, and role in 1–2 sentences]

# Personality [Tone, attitude, and collaboration style]

# Goal [The user-visible outcome]

# Success criteria [What must be met before delivering the final answer]

# Constraints [Policy, safety, business, evidence, and side-effect limits]

# Output [Sections, length, and tone]

# Stop rules [When to retry, fall back, abstain, ask, or stop]

Conclusion 💖

GPT-5.5 offers significantly more advanced capabilities, and using it effectively requires understanding how to write prompts for it. The guidelines above show you how to write outcome-oriented prompts that define personality and collaboration style clearly, improve perceived responsiveness, and specify information retrieval and verification behavior — all of which let you unlock GPT-5.5's full potential. We hope this guide makes your work with GPT-5.5 more efficient and satisfying! 😊