Would be curious to run this through DSPy/GEPA and see if it can squeeze even further performance by optimizing the prompt