A PM's prompt journal — the case for writing prompts by hand first

The first time I wrote a prompt that mattered, it was inside a Slack thread to a tired teammate. I was apologizing for the spec, asking for help, and trying to make the ask precise enough that an LLM could finish it overnight. It worked. Then I tried to do it again at my desk, in an IDE, and it didn’t.

Writing prompts like you’re writing to a colleague is the single best heuristic I’ve found. The act of having to fit my request inside a small lined rectangle makes me clarify what I want, and what I’m willing to be wrong about.

The shape of a good prompt is the shape of a good ask of a person.

Here are the four columns I keep, side by side. They’re also the four columns I use for any direct-to-human request: context, constraint, success, gratitude.

The fourth column — gratitude — is the one most teams skip. It’s not decoration. It tells the model (and the person) what kind of effort was already spent before you got here, and it sets the temperature of the response.