Summary
- OpenAI has released a new prompting guide for GPT-5.6 Sol, which revises previous recommendations.
- Tests with internal coding agents indicated that streamlined system prompts enhanced evaluation scores by about 10–15%.
- This guide introduces a novel section on Programmatic Tool Calling and emphasizes the text.verbosity API parameter—both of which were not included in the GPT-5 guide.
OpenAI has unveiled a new prompting guide for its recently launched flagship model, GPT-5.6 Sol. The primary takeaway is likely to surprise those who have spent the past year crafting extensive system prompts: less is more. The emphasis is now on outcome-first prompting, where one should define the desired results, establish stopping conditions, and then step back.
Instructions that are overly detailed, repeated style guidelines, and examples that fail to influence behavior are now deemed unnecessary.
OpenAI supports this shift with data: in tests with coding agents, more concise system prompts led to evaluation scores rising by approximately 10–15%, while reducing total tokens by 41–66% and costs by 33–67%.
Comparing GPT-5 and GPT-5.6: Key Changes
The GPT-5 prompting guide, launched in August 2025, focused on providing extensive scaffolding. It included XML persistence blocks instructing the model to continue working until the issue was fully resolved, detailed templates for context gathering, and scripts that explained each step aloud.
This approach aimed to manage the model's eagerness by creating specific guidelines on when to exert more effort or to ease off.
However, GPT-5.6 largely eliminates the need for such scaffolding. The new guide advises users to cut back on repeated rules, ineffective style instructions, and unnecessary examples. Essentially, the previous blocks intended to assist the model are now seen as additional complexity that it must work around.
The focus now is on the essential elements: the visible outcome for the user, the criteria for success, stopping conditions, and strict constraints. A model prompt should begin with "Resolve the customer's issue end to end," followed by a clear definition of what completion looks like, the necessary actions to take, and instructions for handling missing evidence. There's no room for vague directives like "be thorough" or "keep going"—just a straightforward destination.
The guide also highlights a change in risk assessment. It cautions that GPT-5.6 adheres closely to prompt contracts, and that "conflicting rules can lead to greater instability than missing details."
Unlike earlier models that would choose one instruction when faced with conflicting guidance, GPT-5.6 will expend reasoning tokens to try to resolve both, which can slow down processing, increase costs, and often lead to errors. If your system prompt contains overlapping rules—which is common in production prompts—this is the first issue to address.
Moreover, OpenAI strongly discourages the use of absolute phrases like "always do this" or "never do that" to influence the AI's behavior.
Two significant additions characterize the differences. The first is the text.verbosity parameter: Since GPT-5.6 is inherently more concise than GPT-5.5, previous instructions to "be brief" now risk making responses excessively short. The parameter allows for a global default to be set, which can then be adjusted for specific tasks in the prompt. The second addition is a section on Programmatic Tool Calling, designed for structured workflows where code manages filtering, batching, or consolidating large outputs, thus relieving the model of that burden.
Does It Deliver Results?
Utilizing the new guide, we optimized our prompt for TYPE OR DIE, a first-person typing survival horror game designed to evaluate a model's coding capabilities. The outcome was notably improved: GPT-5.6 Sol handled the auto-aim logic more effectively than before, the visuals were more coherent, and the overall game experience felt refined.
However, this process took longer to complete. The model did not immediately generate code; instead, it first mapped out the problem and planned each system before writing any code. This demonstrates the guide's intended function—setting the destination allows the model to determine the best path.
The updated prompt is accessible on our Github, where you can view it.
You can experience the original GPT 5.6 game by clicking on this link.
The game created under the new prompt can be found here.
If you wish to explore further or prefer not to memorize all these new guidelines, you can develop your own custom GPT and provide it with the complete guide as its knowledge base. This setup allows it to analyze any prompt you provide, comprehend the underlying logic, and rewrite it in the style of GPT-5.6. Essentially, you can use prompt engineering to create better prompts.
Promptception. You're welcome.