Program with the LLM the way you should program. Have it always build the most naive implementation, then ask it to clean it up, then ask it to add one feature at a time, then ask for optimizations, then ask for alternatives, then ask if you are going down the right path or not. Keep working incrementally.

LLM outputs are like clay, you have to sculpt it.