OpenAI published a prompting guide for GPT-5.6 Sol that reverses earlier advice on prompt engineering. Internal coding-agent tests showed lean system prompts improved evaluation scores by roughly 10-15% while cutting total tokens by 41-66% and costs by 33-67%. The guide introduces outcome-first prompting, advising developers to define success criteria and stopping conditions rather than writing detailed how-to instructions, and includes new sections on Programmatic Tool Calling and the text.verbosity API parameter.
OpenAI Internal Tests Show 10-15% Performance Gains with Lean Prompts
OpenAI backed the new approach with internal testing data. In coding agent tests, leaner system prompts improved evaluation scores by roughly 10-15%. The tests also showed token usage dropped by 41-66% and costs decreased by 33-67% compared to longer prompts. The core recommendation is outcome-first prompting: define what good looks like, set stopping conditions, and remove detailed how-to instructions, repeated style rules, and examples that don't change behavior.
OpenAI Shifts from Scaffolding to Outcome-First Prompting
The GPT-5 prompting guide, published at launch in August 2025, focused on adding scaffolding. It included XML persistence blocks, detailed context-gathering templates for parallelizing searches, and tool preamble scripts narrating every step. The philosophy was calibrating eagerness by building explicit rails for when to intensify effort or stand down. The GPT-5.6 guide tells developers to trim repeated rules, style instructions that don't change behavior, examples that do nothing, and process steps the model already handles reliably. What remains is simpler: the user-visible outcome, success criteria, stopping conditions, and hard constraints. The guide warns that GPT-5.6 follows prompt contracts closely and that conflicting rules can create more instability than missing detail. When the model hits conflicting instructions, it burns reasoning tokens trying to reconcile both, which is slower, more expensive, and often wrong. OpenAI advises against using absolutes like "always do this" or "never do that" to steer behavior.
OpenAI Introduces text.verbosity Parameter and Programmatic Tool Calling
The guide includes two concrete additions. The first is the text.verbosity parameter: Because GPT-5.6 is already more concise by default than GPT-5.5, old "be brief" instructions now over-correct and make responses too short. Developers can set a global default via the parameter, then override per task in the prompt. The second is a section on Programmatic Tool Calling for bounded workflows where code handles filtering, batching, or aggregating large intermediate outputs and returns a compact result, offloading that work from the model's judgment entirely. Both features were absent from the GPT-5 playbook.
Developers Test New Guidelines on TYPE OR DIE Game Build
Developers used the guide to optimize prompts for TYPE OR DIE, a first-person typing survival horror game built to benchmark coding abilities. The result was more polished: GPT-5.6 Sol tackled auto-aim logic more efficiently than on previous runs, visuals had more coherence, and the overall feel of the game was cleaner. The model didn't jump straight to code—it mapped the entire problem first, planned each system before writing a line. The optimized prompt is available on Github, with both the original GPT-5.6 game version and the version created under the newer prompt accessible for comparison.
FAQ
What did OpenAI change in the GPT-5.6 Sol prompting guide?
OpenAI published a prompting guide for GPT-5.6 Sol that reverses earlier advice by recommending lean prompts focused on outcomes rather than detailed how-to instructions. The guide introduces outcome-first prompting, advising developers to define success criteria and stopping conditions while removing repeated rules, style instructions, and examples that don't change behavior.
How much did lean prompts improve performance in OpenAI's internal tests?
In OpenAI's internal coding-agent tests, leaner system prompts improved evaluation scores by roughly 10-15% while cutting total tokens by 41-66% and costs by 33-67% compared to longer prompts.
What new features does the GPT-5.6 prompting guide introduce?
The GPT-5.6 guide introduces the text.verbosity API parameter to control response length globally and per-task, and a section on Programmatic Tool Calling for workflows where code handles filtering, batching, or aggregating outputs. Both features were absent from the GPT-5 playbook published in August 2025.
