OpenAI’s official release of the latest GPT-5.6 prompt writing guidelines strongly recommends reading the original text yourself.

Many people say GPT-5.6 isn’t very usable—it always gets stuck in infinite loops, and a single task takes a very long time. A large part of the reason is that the prompt hasn’t been written well. Now the official team defines the best practices for prompt writing here. I tried it myself, and the results are indeed significant.
Document link:
A brief summary of the full text:
1. GPT-5.6 has stronger task planning abilities, so when writing prompts you don’t need to tightly control the model’s thinking process with lots of precise steps. Instead, you should give it enough autonomy. The model will plan the optimal path to solve the problem on its own.
2. Describe the prompt in the most concise and efficient language—directly define what counts as “acceptance completed,” what the constraints are, and what context the model needs to understand.
3. Don’t force the model to show its reasoning process. Tell the model when it should use tools (for example, if it needs the latest data, it can search online) rather than making tool calls mandatory.
4. Have the model verify the accuracy of its answer after completing it, and allow it to correct itself.
5. Treat GPT-5.6 like an employee—not an answer generator. Just describe the final goal, the context, what resources it has, and then let it check itself after it’s done.
At the end, the official also provides prompt templates—recommended to try them.