[DigitalToday reporter Jinju Hong (홍진주)] OpenAI released a new prompt guide for GPT-5.6 and suggested that concise, clear instructions are more effective than long system prompts. The approach is to clearly set only the desired outcome, completion criteria and stop conditions, while leaving detailed execution to the model.

On July 13 local time, blockchain outlet Decrypt said the core of the guide is "outcome-first prompting". It said users should first state the final result they want and the conditions for success, while minimising process descriptions and repetitive rules.

OpenAI also released results from internal tests of a coding agent. After shortening system prompts, evaluation scores improved by about 10 to 15 percent. Total token use fell 41 to 66 percent. Costs also fell 33 to 67 percent. This means the previous approach of using system prompts running several pages could hinder performance rather than improve it in the latest models.

The change contrasts with the GPT-5 prompt guide released in August 2025. At the time, it recommended instructions such as continuing to work until the problem is solved, parallel search, step-by-step context gathering, and adding explanations of procedures before using tools. OpenAI explained that GPT-5.6 can handle much of this on its own, so it is better to reduce redundant rules, unnecessary behaviour examples and excessive style directions.

The new guide says a good prompt does not start with an abstract command. OpenAI advised stating the goal first, such as "solve the customer’s problem from start to finish", rather than phrases like "be thorough" or "keep going". It also said it is effective to clearly specify how far to go for completion, what tasks must be finished before responding, and how to handle situations with insufficient grounds.

Its approach to risk management also changed. OpenAI warned that GPT-5.6 follows the prompt contract very strictly and that conflicting rules can create greater instability than a lack of detailed instructions. It said earlier models tended to choose one of two conflicting instructions, but GPT-5.6 may try to satisfy both, using more reasoning tokens and becoming more inefficient in terms of speed and cost.

It stressed that in real service operations, system prompts often contain overlapping rules, so conflicts should be checked first. It also recommended reducing absolute expressions such as "always do this" or "never do this". Repeating such phrases to tightly control model behaviour can instead cause instruction conflicts and excessive correction, it said.

The guide also adds settings that were not in the existing GPT-5 guidance. One example is the 'text.verbosity' parameter. GPT-5.6 tends to answer more concisely than GPT-5.5 by default, so repeating phrases like "answer briefly" as before could make responses overly short.

OpenAI advised setting overall response length with text.verbosity and handling only task-specific fine-tuning in the prompt.

A "Programmatic Tool Calling" section was also added. It recommended that clearly scoped tasks such as filtering, batch processing and aggregating large intermediate results be handled by code, with only compressed results passed to the model. The approach aims to boost efficiency by separating tool calls and data processing programmatically rather than having the model judge every intermediate step itself.

Decrypt also introduced a case applying the new guide to optimising prompts for the benchmark typing survival horror game 'TYPE OR DIE'. GPT-5.6 Sol handled auto-aim logic more efficiently than in the previous run and produced a more organised result in visual completeness and overall composition. But the outlet said development took longer because it structured the whole problem, planned each system first and then began implementation rather than moving straight to writing code.

The guide shows that the benchmark for prompt engineering is changing. In the past, it was common to tightly control model behaviour with long instructions, but the emphasis is now shifting to clearly presenting success conditions and constraints while leaving the execution path to the model.

As a result, companies and developers adopting GPT-5.6 are expected to prioritise removing conflicting instructions rather than continually adding rules to prompts, and to redesign result criteria, stop conditions and tool-separation structures.