you need to offload a lot of this to tools & trusting the model to dynamically handle situations on its own. We have a model that diagnoses and troubleshoots hvac systems with a 90+% accuracy for non mechanical failures and it's prompt is 400 tokens