LLMs don't really think, they emulate their training data. Which has a lot of examples of humans walking through problems to arrive at an answer. So naturally, if we prompt an LLM to do the same, it will emulate those examples (which tend to be more correct).

LLMs are BAD at evaluating earlier thinking errors, precisely because there's not copious examples of text where humans thinking through a problem, screwing up, going back, correcting their earlier statement, and continuing. (a good example catches these and corrects them)

Claude code is actually great at that