Prompting for design: what works, what doesn't, and what's just hype

The framing I see repeated everywhere is: AI is going to change design. That's probably true, eventually. What's also true is that most advice about how to use it now is too generic to be useful.

"Write detailed prompts." "Give it context about your users." "Iterate on the output." These describe using any AI tool for any task. They're not design-specific, and they don't tell you what to actually work on.

Where the value actually is

For design work specifically, the prompt patterns that return consistent value are narrow.

Writing content for UI states is one. Error messages, empty states, confirmation dialogs, button labels for actions that need to be specific. These have short, structured outputs where you can validate accuracy immediately. "Write five variations of an error message for a failed payment that tells the user exactly what to do next" works because the answer space is defined. You evaluate and pick.

Generating alternative component API options works for the same reason. You have specific constraints, you want to see options you might not have considered, and you can evaluate the output against your requirements. You're not generating the decision. You're expanding the option set before you make it.

Research synthesis, when you have transcripts or notes. Not as a substitute for reading them yourself, but as a way to surface patterns before you form your own interpretation. It catches the thing mentioned once across fifteen interviews before your read-through buries it.

Where it doesn't work

Open-ended concept generation returns generic output because the output space is unbounded. "Generate UI concepts for a dashboard" produces the dashboard everyone would produce. There's nothing to push against, so the model produces the middle of the distribution.

Anything requiring product context it doesn't have. Naming, prioritization, decisions that depend on knowing your specific user base, your existing patterns, or your organization's constraints. These aren't limitations you can prompt your way around. They're structural.

Visual generation is genuinely limited for high-fidelity UI work right now. The outputs are plausible enough to use as inspiration and not reliable enough to use as design. That line moves, but that's where it is.

The hype part

The hype version of AI in design is that it changes the core job. I don't think that's where we are. The core job is still identifying the right problem, making judgment calls about tradeoffs, understanding user behavior, and being in the room when hard decisions get made.

What's changed is the cost of producing certain kinds of output. Documentation, alternative options, synthesis, coverage checking. The work that's tedious and time-consuming but doesn't require invention. Getting faster at that work frees up time for the parts of the job that AI isn't replacing anytime soon.

That's worth knowing about. It's just a different claim than transformation.