I'm currently building a side project that integrates an LLM to help users generate structured content from messy input. While experimenting, I’ve noticed that even small changes in phrasing can drastically affect output quality — and it's made me wonder how much of this is due to prompt design vs. the model’s own quirks.

Curious to hear from others:

What are some common mistakes you've seen (or made) when prompting LLMs?

Any patterns in user behavior that lead to unreliable or unexpected results?

Are there prompt-writing techniques that work across models (ChatGPT, Claude, Llama, etc.)?

Would love to collect insights and even horror stories from folks deploying or experimenting with LLMs.