It’s interesting that LLMs struggle more to give excel feedback, I think that it’s likely due to the lower amount of training data available when compared to questions re code issues (e.g, on stackoverflow which often have an explanation accompanied by a code snippet).