OpenAI releases new GPT-5.6 prompt engineering guidelines, recommending simplifying system prompts

Deep Tide TechFlow news flash: On July 14, according to Decrypt, OpenAI released a new prompt engineering guide for its flagship model GPT-5.6 Sol. The core idea is “result-first prompting”—clearly define goals and termination conditions, and avoid verbose process-oriented instructions. Internal testing by programming agents shows that the streamlined system prompt increases evaluation scores by about 10–15% while reducing total token usage by 41–66%, lowering costs by 33–67%.

The guide also added a text.verbosity API parameter to globally control output detail, as well as a new “programmatic tool calling” section that supports handing intermediate tasks such as filtering and batch processing to code, reducing the model’s decision-making burden. In addition, the guide warns that GPT-5.6 strictly follows prompt rules—conflicting instructions within a prompt will consume large amounts of reasoning tokens and may lead to incorrect outputs.