- OpenAI published a dedicated prompting guide for GPT-5.6 Sol that changes earlier advice.
- Internal coding-agent tests showed lean system prompts improved eval scores by roughly 10–15%.
- The guide introduces a first-ever section on Programmatic Tool Calling and highlights the text.verbosity API parameter.
OpenAI published a new prompting guide for GPT-5.6 Sol, its flagship model, and the main message contradicts the verbose, multi-page system prompts common last year. The core idea is outcome-first prompting: define what good looks like, set stopping conditions, and get out of the way.
Detailed how-to instructions, repeated style rules, and ineffective examples are now considered noise. OpenAI backs this with numbers: internal coding-agent tests showed leaner prompts improved evaluation scores by roughly 10–15% while cutting total tokens by 41–66% and costs by 33–67%.
The GPT-5 prompting guide, published at launch in August 2025, focused on adding scaffolding like XML persistence blocks and tool preamble scripts. GPT-5.6 mostly does not need those rails, and the new guide tells users to trim them.
The guide warns that conflicting rules can create more instability than missing detail, as GPT-5.6 burns reasoning tokens trying to reconcile conflicting instructions. OpenAI also advises against resorting to absolutes like “always do this” or “never do that.”
Two concrete additions are the text.verbosity parameter to avoid over-correcting for conciseness and a section on Programmatic Tool Calling for bounded workflows. The guide’s model of a good prompt starts with “Resolve the customer’s issue end to end” and specifies exactly what done looks like.
The new prompt is available on GitHub so users can check it out. The original GPT-5.6 game can be played by clicking this link, and the game created under the newer prompt is available here.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
