Low temperature, heavy prompting to answer in a structured way. Sadly can't fine train models since this is API based but the approach does work!