OpenAI releases GPT-5.6 prompt guidelines: only give the results and the red lines; leave the rest to the model

OpenAI releases the GPT-5.6 prompt guidelines. Internal testing found that after significantly shortening a verbose system prompt, scores actually rose by 10% to 15% instead of dropping. Token usage was also reduced by 41% to 66%. The guidelines suggest developers should only provide model results, constraints, and stopping criteria—no need to specify how to take every step.
（Background: GPT-5.6 secret tests are spreading fast: ChatGPT suddenly got smarter, and user testing crushes Fable 5）
（Background add-on: Harness Engineering (AI driving engineering) beginner article: OpenAI’s latest programming standards, teaching you how to reach Lv.1 easily）

Table of Contents

Toggle

• Does the model get more accurate with fewer instructions?
• Stop writing steps—just write the “end point”
• From “stacking instructions” to “measurement-based adjustments”

OpenAI’s internal tests show that after substantially simplifying an originally lengthy system prompt, the score didn’t fall—instead it increased by 10% to 15%, while token usage dropped by 41% to 66%. In other words, when you tell the model to do less “preparation actions,” it actually runs faster and answers more accurately.

This is the key point described in OpenAI’s official GPT-5.6 prompt guidelines. Below, the author summarizes three takeaways: why simplifying instructions is more effective, what new writing style OpenAI recommends, and how developers should actually change what they do.

Does the model get more accurate with fewer instructions?

Engineers’ intuition when writing prompts used to be: “The more detailed you are, the more obedient the model will be.” They’d write every step and every edge case, fearing the model would miss something. But OpenAI calls out in the guidelines that this intuition fails with GPT-5.6. The engineering team’s tests found that a system prompt—which, in simple terms, is the hidden instruction set developers insert into a model before a user conversation begins—when filled with repeated rules, behavioral-impact-light style reminders, unnecessary examples, and process guidance the model can already do, removing it actually improves performance.

The specific approach in the guidelines is: “Start from a working version and delete step by step.” Keep the currently effective prompt, remove sections one by one that look suspicious, and watch eval scores—meaning, in simple terms, run the model on a fixed set of tasks and use scores to judge whether the answers are good and whether performance drops.

What truly should be kept is what the user can see: results, success and stopping criteria, safety and business constraints, tool selection rules that vary by context, and the required output format. The rest is mostly extra text engineers add to feel more comfortable.

The guidelines also point out a frequently ignored aspect: tool descriptions themselves are also part of the prompt. The more tools you give the model—and the more vague the descriptions—the higher the cost for the model to decide which tool to use. OpenAI recommends only providing task-relevant tools. Each tool description should explain what it does, when it should be used, and how it should behave when errors occur. In other words, trimming isn’t only about “instructions”—even the “toolbox” itself should be slimmed down.

Stop writing steps—just write the “end point”

The most core line in the guidelines is: “Define the result, important constraints, usable evidence, and completion standards, then leave room for the model to choose an efficient path itself.” In simple terms: tell the model where to end up and what red lines it can’t cross—don’t dictate how it should take every step.

The example OpenAI gives is: “Solve the request with the fewest useful tool loops, but don’t let reducing the number of loops sacrifice correctness, necessary evidence, or citations.” This is an “if X then Y” decision rule, not a rigid command.

Another detail that’s easy to overlook is that GPT-5.6 answers are already more concise by default. Previously, engineers wrote reminders in prompts like “Please answer briefly,” but now they may be redundant or even harmful. The guidelines recommend using the parameter text.verbosity—simply put, control how long the model’s response should be with a single parameter, separate from tone. Split into three tiers: low, medium, high dedicated to length; personality such as tone and formality is described separately, and both are kept short.

If you truly need an even shorter answer, the guidelines recommend stating explicitly what to keep and what to cut, rather than vaguely saying “make it shorter.” As for how much effort the model should spend thinking—that is, reasoning effort (how much “thinking effort” the model should expend before responding)—the guidelines split it into low, medium, high, xhigh, and max. Before increasing it, OpenAI suggests first checking whether the prompt itself clearly states success criteria and the validation loop. Often, “making it clear” is more effective than “making the model think more.”

From “stacking instructions” to “measurement-based adjustments”

The most practical impact on developers is a prompt migration workflow attached to the guidelines.

OpenAI puts it plainly: when switching models, don’t rewrite your entire prompt in one go. The reason is that if you change the model, reasoning settings, prompt, and toolset at the same time, you won’t be able to tell which change caused behavior differences. The correct order is: switch models first, keep the original reasoning strength settings, run one representative eval as the baseline, then remove outdated scaffolding and duplicate instructions. Only make minimal adjustments to areas where the eval truly shows a drop, and then test again. After changing each variable, measure again—never touch things randomly at the same time.

The guidelines also require developers to spell out how many decisions the model can make on its own, rather than being vague. OpenAI’s example policy is: for requests like answering, explaining, and reviewing, the model may only check and report—it can’t make changes. For requests that involve changes or fixes, it can make local changes within scope and run non-destructive validations. For requests that truly involve external writing, destructive actions, or expanding the task scope, it must stop and ask first.

Writing more and more instructions used to be how engineers found reassurance. Now it seems this may slow the model down and also increase the bill. When the model gets smarter, people should learn to speak less—save their effort for measurement and validation.