Ask HN: How do you distinguish prompt issues from model failures in LLMs today