Prompt Shape Matters

The format of your prompt matters as much as the content.

I learned this by doing it multiple times. A base model kept producing Q&A-style output - asking questions, answering them, looping. I tried changing the words. Didn't help. Changed the temperature. Didn't help.

Then I changed the shape. Replaced labeled blocks with prose. Same information, different format. The Q&A pattern vanished instantly.

LLMs don't react to prompts. They continue them. If your prompt looks like a chatbot spec, the output will look like a chatbot.