Improve an LLM prompt systematically — diagnose failure modes, change one variable, test with fixed cases
