Building a unique AI-native design workflow
How an Art Director and Claude built two production-deployed properties using a human-AI design system — and what that process actually looks like.
Most of what gets called an "AI workflow" is a screenshot of a chat. This one isn't. Over the last eighteen months I've shipped two production properties — this site and Sidekick — using a real, repeatable loop between a senior art director (me) and Claude. The loop is not magic. It's a sequence of moves that turn good taste into shipped product, faster. Here's what it actually looks like.
Start from the brief, not the canvas
Every piece of work starts the same way: a brief written in plain English, in a single Claude conversation. What's the goal, who's the audience, what's the constraint, what does success look like. The brief gets refined back and forth until the language is sharp enough to direct any decision downstream. The canvas only opens once the brief is so clear that the next move is obvious.
“If you can't write the brief in plain English, you can't design the answer.”
Direction before pixels
Before any layout, I write a tight art direction note: what does this page sound like, what does it feel like, what does it borrow from. Tone, restraint, references, the one thing it absolutely can't look like. Claude helps me pressure-test the direction against the brief — does this serve the goal, or am I drifting into taste-for-taste's-sake. The direction note is the spec; everything below it has to defend itself against the note.
Ship in code, not in mocks
Mocks lie. They look right in Figma and break in production. So the moment the direction is set, I move into Framer or React and build the real thing — type sized at production weight, motion at real frame rate, content at real length. Claude is the implementation partner here: I describe the move I want, it scaffolds the component, I edit it down to the version that's actually mine. The output is a live page, not a screenshot.
“The mockup is a hypothesis. The deployed page is the answer.”
Iteration is a conversation
Once the first version is live, the iteration loop runs in the same window as the work: open the page, take a screenshot, paste it into Claude, talk through what's not landing. The critique is fast because the artifact is real — there's no translation cost from mock to reality. Most rounds are minutes, not days. The work converges on the brief instead of drifting away from it.
Taste stays human
What AI does well: scaffolding, repetition, plumbing, the fifty small CSS decisions that aren't load-bearing. What it doesn't do: choose the typeface, decide what gets cut, hold the room. Every page that ships gets its final pass from one set of eyes, not a model. The workflow doesn't make the design — it removes the friction between the idea and the page so the design can be made faster, sharper, and at a higher altitude.
“The fastest tool in the world is still useless without the right brief.”
AI handles the plumbing. Craft tools handle the polish. Every part of the workflow is opinionated about which tool runs which move.
Two production properties built with the same loop.
This site. Designed, built, and rewritten end-to-end inside the workflow — every page is a Framer code component, every word a Claude conversation.
A consumer fintech AI tool built with Claude + Plaid. Live product, real users, all designed and shipped using the same human-AI loop.