From my understanding I would say. Continuously iterate and refine your prompts based on the observed behavior of the LLM. Analyze the cases where function calls are missed and identify patterns or anomalies in the prompts that may contribute to the miss rate.