I've noticed a strange phenomenon. No matter how well-written the prompt is, the quality of the first generation will never reach the upper limit of AI. Allowing AI to reflect and identify its shortcomings will always lead to quality improvements. Perhaps it's because there's context in the conversation? However, the number of iterations should not be too many, otherwise it will start to produce useless information.
