The case for small, specific prompts

We've reviewed a lot of prompts this year. The good ones are almost always short. Not because short is a virtue — but because the kind of person who can say what they want in three lines has already done the work that a three-page prompt is avoiding.

What long prompts are hiding

When a prompt grows past a page, it's usually carrying one of these:

• Uncertainty about what "good output" looks like, compensated with examples.
• Residue from fixing old bugs the model has since grown out of.
• Fear that skipping an instruction means the model won't do it.

The model doesn't need your fear. It needs the task.

An exercise

Take your longest prompt. Delete everything except the single sentence that describes what you want. Run it. Add things back only when an output fails, and only the specific instruction that would have caught that failure.

You'll usually end up at something shorter than the original, stronger than the original, and easier to maintain.