"this should return X, it returns Y, find out why"
With enough tooling LLMs can pretty easily figure out the reason eventually.