Prompting Isn't a Dev Skill Anymore — It's a Design Skill

A few months ago, one of our designers pulled up a screen she’d generated with an AI tool in about ninety seconds and said something that stuck with me: “This looks like a template someone gave up on halfway through.” That’s the honest state of most AI-generated design work right now. Not bad, exactly. Just generic — the visual equivalent of small talk.

The tools aren’t the problem. The prompts are.

For the last year, “prompting” has mostly been treated as a developer’s problem — something you learn to get better code out of a chat window. But at an agency, we’ve found the opposite is true. The people getting the most out of these tools aren’t the ones with the cleverest technical setup. They’re the designers who’ve quietly gotten very good at describing what they want, the same way they’d brief a junior designer or write a note for a developer picking up their Figma file.

That’s the skill nobody’s teaching yet. So here’s what we’ve actually learned by doing it wrong first.

Constraints do more work than adjectives

The instinct when prompting an AI tool is to describe the feeling you want — “modern,” “clean,” “premium.” Those words mean almost nothing to a model, because they mean almost nothing concrete. Compare it to briefing a new designer on your team: you wouldn’t just say “make it feel premium.” You’d say stick to the existing grid, use these two typefaces, keep the palette to three colors, and match the spacing scale we already have.

The same brief works better on a model than any adjective does. Give it the actual constraints — grid, spacing tokens, type scale, existing component names — and the output stops looking like a stock template and starts looking like it belongs on your site.

Feed it your system, not a blank page

Most people open one of these tools cold and describe a screen from scratch, the way you’d describe a room to someone who’s never seen your house. It’s slow and the result rarely fits anything you already have.

It works much better in reverse. Paste in your existing component names, your token names, a short description of your design system’s rules. Then ask for the new screen within that system, not next to it. The tool isn’t inventing a UI anymore — it’s assembling one out of parts you’ve already approved. That single change is probably the biggest quality jump we’ve seen, more than any prompt phrasing trick.

Build in layers, the way you’d actually design

The other mistake is trying to get the whole screen right in one prompt — layout, hierarchy, copy, color, and spacing, all at once. That’s not how design works when a person does it, and it’s not how it works with a model either.

Ask for structure first: what sections exist, in what order, with what relative weight. Once that’s right, ask for hierarchy — what draws the eye first, second, third. Only then get into detail: color, type, spacing, micro-copy. Each pass is a smaller, more specific question, and smaller questions get better answers. It’s slower than a single mega-prompt by about two minutes and faster than starting over by about an hour.

Treat the output like a draft from a junior, not a decision

This is the part that matters most and gets skipped most often. AI output is a first pass. It’s meant to be critiqued, not shipped. The teams getting burned by these tools are usually the ones treating generation as the finish line instead of the starting point — publishing the first thing the tool hands back because it looks finished at a glance.

Look at it the way you’d look at a junior designer’s first draft. Where does the hierarchy break down under real content instead of placeholder text? Where does it ignore an edge case — empty states, long strings, error messages? Where does it quietly drift from the design system because the model filled a gap with its best guess instead of your actual pattern? That review pass is still design work. Arguably it’s the design work that matters most now, because the blank-page part got faster and the judgment part didn’t.

What this actually changes

None of this makes the tools smarter. It makes the person using them clearer — about constraints, about systems, about what a first draft is for. That’s not a technical skill. It’s the same skill that made someone a good design lead before any of this existed: knowing exactly what you want before you ask for it, and knowing what’s wrong with what you get back.

The tools will keep changing. That part won’t.