If performance is a concern, just re-prompt GPT if there are any performance concerns, and to propose fixes for them if there are.

This is called reflexion and has been found to be an effective way to improve response accuracy and usefulness (also reduces hallucinations).

https://arxiv.org/abs/2303.11366