When GPT starts to auto-incorporate the best yet known prompts, then we'll have a better idea of its potential. You must, must use the best prompts, of which many are not widely known, and some have not (of course) been discovered ... yet. Even with human experts, you must provide sufficient detail, and the expert must ask clarifying questions for differential diagnosis.
This is nonsense - a LLMs output is a pseudo random mishmash of its training data. How can there be a "best prompt" when the same prompt gives you different output each time, but there is only one correct output?
This is true in trivial ways: ask it to reply only in json.
This is also true in slightly less trivial ways: ask for a “description of X”, a “short description of X”, and a “one-sentence description of X”.
This continues to be true in increasingly more complex ways: Prompt it with a quote, a rating of that quote as 5/10 on complexity, and request it to give you two new quotes with ratings of 2/10 and 8/10 on complexity. Follow it up with a request for two more quotes at -5/10 and 20/10 ratings. Then try the whole process again but with metrics other than “complexity”: information content, eloquence, humour.
In this way, if there is a single correct output, there is a corresponding single best input prompt that most tightly constrains the output space to the smallest space that still contains the correct output.