Use few-shot learning. Build a simple prompt with basic examples of how to use the API and it will do significantly better.

LLMs just guess, so you have to give it a cheatsheet to help it guess closer to what you want.