I tested around it and doing two passes (generate code and "make it more efficient") works best, with system prompt engineering to result in less code output: https://github.com/minimaxir/simpleaichat/blob/main/examples...