One unexpected discovery that I have made while building an AI-based system: the LLM's are bad at designing prompts. We tend to think that the AI has some sort

One unexpected discovery that I have made while building an AI-based system: the LLM's are bad at designing prompts. We tend to think that the AI has some sort of self-knowledge and should be good at designing prompts for itself but it's really not. Been struggling with a task that heavily depended on prompts, ended up rewriting all my prompts from scratch in my own words, and it finally worked. Then every time I ask Claude to fix something in the prompts, it invariably makes it worse. A very strange phenomenon that can probably be explained by the quality of prompt design advice that made it to the training dataset. Bottomline, all the prompt design advice that you can find on the internet is really not great.