Summary

  • OpenAI has launched a new prompting guide for GPT-5.6 Sol, revising previous methods.
  • Tests showed that streamlined prompts improved evaluation scores by 10-15%.
  • The guide introduces Programmatic Tool Calling and emphasizes the text.verbosity API, absent in the previous version.

OpenAI has introduced a new prompting guide for its latest model, GPT-5.6 Sol, advocating a more streamlined approach to prompt creation. The key message for users is that brevity is beneficial; focusing on the desired outcomes rather than lengthy instructions can yield better results.

This shift encourages users to specify the intended results and set clear stopping conditions, minimizing the need for detailed instructions and repetitive style guidelines that previously cluttered prompts.

Evidence from internal tests with coding agents suggests that concise system prompts not only improved evaluation scores by 10-15% but also reduced the number of tokens used by 41-66%, leading to cost savings of 33-67%.

Key Differences Between GPT-5 and GPT-5.6

The initial GPT-5 prompting guide, released in August 2025, emphasized extensive scaffolding, including detailed templates and instructions designed to guide the model through problem-solving processes. This method aimed to control the model's output by providing structured guidelines.

In contrast, the GPT-5.6 guide advocates for simplicity, advising users to eliminate unnecessary rules and ineffective style instructions. The focus is now on defining the user's visible outcome, success criteria, and clear constraints. For example, prompts should start with directives like "Resolve the customer's issue end to end," followed by precise definitions of completion and required actions.

The new guidelines also address risk management, noting that GPT-5.6 adheres closely to prompt contracts. The guide warns against using conflicting instructions, which can create instability and slow down processing. Users are discouraged from employing absolute phrases that could limit the model's adaptability.

Two notable additions to the guide include the text.verbosity parameter, which allows for customizable response lengths, and a section on Programmatic Tool Calling, aimed at managing large outputs more efficiently through structured workflows.

Performance Insights

Utilizing the new guidelines, a prompt was optimized for the game TYPE OR DIE, resulting in significant improvements in the model's performance, especially in handling game mechanics. However, the process took longer as the model prioritized planning before executing code, illustrating the guide's focus on setting clear destinations for optimal problem-solving.

Users can access the updated prompt on GitHub and explore the enhanced gameplay experience created under these new guidelines. Additionally, developers can create custom GPT models that incorporate the entire prompting guide, enabling them to analyze and refine prompts further.