The first prompt that actually worked for me

For weeks I was writing prompts like how do I fix this bug, and getting back answers that were technically correct but not useful.

The prompt that broke through was ugly and long. I pasted the file I was editing, the error I was seeing, what I had already tried, what I thought the cause was, and what a correct fix would look like in plain English.

The answer I got back was not magic. It was just a lot better because the model had enough context to reason. I realized I had been treating the model like a search engine instead of like a colleague who had just walked into the room.

Since then, my prompts have gotten longer, not shorter. I include the constraint, the past attempt, and the shape of the answer I want. I get better results and I spend less time arguing with the model.