The cliché is that AI design output looks like AI: the same purple gradients, the same centred hero, the same rounded everything. That's true of lazy prompting. It is not a property of the model — it's a property of the brief. Give Claude a real design language and constraints, and it produces interfaces with a point of view.
From brief to screen, fast
The slowest part of design has always been the first draft — getting something concrete enough to react to. Claude collapses that. Describe the brand, the audience, the constraints and the content, and it returns a working layout you can critique in minutes rather than days. Critique is easy; the blank canvas is hard. Removing the blank canvas is most of the value.
Design tokens are the steering wheel
The single biggest lever on quality is giving the model your design system, not a vibe. Concrete tokens — a palette, a type scale, spacing, radius, motion rules — turn "make it look good" into "compose within these rules." That's the difference between generic and on-brand.
AI design looks generic when it's asked for "a nice landing page." It looks like you when it's asked to build within your tokens.
This site is the example. Its "Orbital" system — a fixed palette, Syne + IBM Plex type, an aurora-and-grid backdrop, glass cards with a specific hover glow — was defined once, then held consistently across every page. The model isn't inventing a look each time; it's applying one.
Where it shines
- Exploration — five takes on a section in the time one used to take, so you choose rather than guess.
- Consistency — applying an established pattern across many screens without drift.
- The unloved screens — empty states, error pages, settings — the ones that ship half-finished because nobody had time.
- Production-grade markup — not a mockup to rebuild, but real, accessible HTML/CSS you can ship.
Where the human stays in charge
Taste, hierarchy, knowing what to cut — that's still you. The model will happily make a beautiful version of the wrong idea. Direction, brand judgment, and the decision about what not to include don't come from a prompt. The workflow that works: a human sets the system and the intent, the model produces, the human edits. Same as with code.
Designing and building distinctive, fast, accessible interfaces — with or without AI in the loop — is what we do. If you've got a product that deserves better than a template, start a project.