Comment by ambicapter

Maybe the training data that included the words like "skim" also provided shallower analysis than training that was close to the words "in great detail", so the LLM is just reproducing those respective words distribution when prompted with directions to do either.