Every few weeks someone shares a "magic prompt" that supposedly turns any AI into a world-class designer. A month later the model updates and the magic quietly stops working.

Meanwhile, the thing that consistently gets me good output hasn't changed in two years. It's not a prompt. It's a spec.

When I moved from design into product management, writing specs was the skill I had to learn fastest. A developer can't read your mind, and a vague ticket comes back as exactly what you said rather than what you meant. Turns out AI is the most literal developer you'll ever work with. Same rules apply, only more so.

Why "make it look modern" fails

Ask a human designer to "make it look modern" and they'll fill the gaps with taste, context and a few questions over coffee.

Ask an AI and it fills the gaps with the average of everything it's seen. Rounded corners, a gradient, a hero with a stock-looking illustration. Technically modern. Utterly forgettable.

The AI didn't misunderstand you. You just didn't say anything. Every gap in a brief gets filled with the most statistically likely answer, and the most likely answer is, by definition, generic.

So the job isn't to find clever words. It's to leave fewer gaps.

The five parts of a brief that survives contact

When I hand work to an AI now, whether that's a page, a component or a whole site, the brief has five parts. They'll look familiar to anyone who's written a product spec.

1. The outcome, not the output. Not "a landing page". Instead: "a landing page that makes a sceptical visitor curious enough to click through to the book." The outcome tells the model what to optimise for when it hits a decision you didn't anticipate.

2. The audience, specifically. "Users" means nothing. "People who've never heard of this product and are mildly suspicious of AI" changes the copy, the pacing and the tone.

3. The constraints you'd normally carry in your head. Brand colours, typefaces, accessibility needs, what tech it'll run on, what must not change. Designers carry a huge amount of this implicitly. AI has none of it. Write it down.

4. Concrete references and anti-references. "Feels like X, not like Y" does more work than a paragraph of adjectives. Anti-references are underrated: telling a model what you don't want rules out the generic default fastest.

5. What "done" looks like. Acceptance criteria. Works at phone width. No horizontal scroll. Loads without the hero image. If you can't say when it's finished, the AI certainly can't.

Write it as a document, not a chat

The biggest shift in my own workflow was moving briefs out of the chat box and into a proper document. A markdown file I can version, reuse and hand over again.

That sounds like admin. It's actually leverage. When the first result is 80% right, I don't argue with the model across fifteen messages. I fix the brief and run it again. The spec becomes the source of truth, and the output becomes disposable.

That's a very product-manager way of thinking, and it's the bit I most want designers to steal. Your artefact used to be the design. Increasingly, your artefact is the intent behind the design, written clearly enough that something else can execute it.

"Isn't this just prompt engineering?"

Sort of. But "prompt engineering" suggests tricks: magic phrases, role-play, secret formats. Tricks are tied to a model and they decay.

Briefing is tied to thinking. A clear outcome, a specific audience, explicit constraints and a definition of done will work on this year's model, next year's model, and the junior designer who joins your team in January. It's the most transferable skill on the board.

Try this

Next time you reach for an AI design tool, stop before typing. Open a blank doc and answer five questions:

  • What should this make someone do or feel?
  • Who exactly is it for?
  • What can't change?
  • What should it feel like, and what should it definitely not feel like?
  • How will I know it's finished?

Then paste that in. I'd bet the first result is better than your last three attempts combined.

The designers who thrive with AI won't be the ones who learned the most prompts. They'll be the ones who learned to say what they mean.