My early prompts all had some version of the same shape: “give me some creative directions for X.” The outputs were fluent, confident, and almost never usable. It took longer than I’d like to admit to notice the pattern wasn’t the model. It was the question.

What ‘be creative’ actually asks for

Asking an open-ended question invites an open-ended answer, and an open-ended answer, from a system trained to sound plausible across every register at once, tends toward the statistically safe center of whatever category you asked about. “Creative tagline ideas” produces the taglines that sound most like every other tagline, because that’s what’s most represented in the space of plausible answers. I was asking for the thing least likely to actually be novel, using the word most associated with novelty.

The shift started by accident, on a day I was in a hurry and wrote an oddly specific prompt — a length constraint, a banned word list, a required reference to a real detail from the project — because I didn’t have time to explain the vibe I wanted. The output was sharper than anything I’d gotten from a week of open-ended requests. Not because it was more “creative” in any way I could define. Because the constraints closed off the safe, generic answers and forced something more specific into the gap.

Constraints as the actual prompt

What I do differently now: instead of describing the quality I want (clever, bold, unexpected), I describe the shape of the output and the things it’s not allowed to do. Not “a punchy headline” but “a headline under six words that doesn’t use the word ‘seamless’ and references the actual number from the case study.” The second version produces worse results maybe a third of the time. It produces genuinely surprising results far more often than the first version ever did, because there’s less room for the model to retreat to its safest average answer.

This generalizes past writing. Asking for “an interesting layout” gets you the layout most represented in the training data’s idea of interesting, which is to say, not interesting. Asking for a layout that fits a specific, unusual constraint — a fixed number of elements, an aspect ratio nobody designs for, a rule that seems arbitrary — tends to produce something further from the center of the distribution, purely because the center doesn’t fit the constraint.

The uncomfortable part

This means the actual creative work moved. It’s not in the request for ideas anymore — it’s in choosing the constraint. A good constraint is doing the job “be creative” used to pretend to do, and coming up with a good one is a real skill, not a shortcut around having one. I spend more time now thinking about what to forbid than what to ask for, which is a strange sentence to have become true, but it’s the difference between output I use and output I regenerate five more times hoping for something better.

Rhys O’Sullivan is a product designer at Fieldstone, previously Loop. This is one in an occasional series of notes on research, systems, and the process of making things.