OpenAI tells developers to shorten GPT-5.6 prompts

OpenAI published a prompting guide for GPT-5.6 Sol that tells developers to write shorter, outcome-first system prompts. The company said internal coding-agent tests found leaner prompts raised evaluation scores about 10–15% while cutting total tokens 41–66% and lowering costs 33–67%.

The guide instructs prompt authors to state the desired outcome, list clear success criteria and stopping conditions, and remove repeated rules, redundant examples and process steps the model already handles. It recommends describing what done looks like, the actions that must be completed before responding, and what to do when evidence is missing, rather than including long scaffolding that explains how the model should work.

OpenAI contrasted the guidance with the GPT-5 playbook published in August 2025. The earlier playbook included XML persistence blocks, explicit parallel search templates and verbose tool preambles intended to calibrate when a model should escalate or halt. The GPT-5.6 guide advises trimming overlapping or conflicting rules because the new model will try to reconcile contradictions, which consumes reasoning tokens and can produce slower or incorrect outputs.

Two features in the guide support the shorter-prompt approach. A text.verbosity API parameter lets developers set a global response-length preference and override it per task; OpenAI said GPT-5.6 is more concise by default, and old briefness instructions can make outputs too short. A section on Programmatic Tool Calling explains when to offload bounded, high-volume work such as filtering, batching or aggregating intermediate outputs to external code so the model returns a compact final result instead of handling all intermediate steps itself.

OpenAI cited internal coding-agent experiments showing the benefits of leaner system prompts. The company also tested the approach on a project called TYPE OR DIE, a first-person typing survival game used to benchmark coding ability. Under the revised prompt style, GPT-5.6 Sol mapped the full problem and planned systems before producing code. Developers reported cleaner visuals, more coherent auto-aim logic and a more polished feel, although the build process took longer because the model spent more time planning.

The guide warns against absolute directives such as always do this or never do that, and notes that conflicting instructions can be worse than missing details. It encourages teams to specify explicit stopping conditions and hard constraints so the model can choose how to reach the stated outcome without parsing redundant rules.

OpenAI uploaded prompt examples to GitHub and suggested teams can build custom GPTs that use the guide as a knowledge base to analyze and rewrite prompts into the GPT-5.6 style. The guidance follows the earlier GPT-5 manual but shifts emphasis from adding scaffolding to pruning system prompts and focusing on explicit outcomes.