AI-native design, told through one session

I used to design in static comps and hand them off. Now I design in a live loop with an AI agent that prototypes, tests in a real browser, and ships to the codebase, all inside one conversation. The unit of design stopped being a file and became a conversation that emits working artifacts.

Die Cut Studio is a parametric dieline tool for crafters: type a box's dimensions, get a print-ready cut file at exact 1:1 scale. Below is a real session: the explorations, the dead ends, the decisions, and the shipped result. Nothing here is a mockup of a mockup; these are the actual interactive prototypes, embedded live.

Prototype in the real medium, not a picture of it

The first question: where should the 3D preview live in the editor? Instead of three static frames, I got three interactive mockups: real fonts, real colors, a box that actually rotates and folds. A comp describes an idea; a prototype argues for it. I dragged the corner dock, it snapped shut into a tidy pill, and I knew in half a second.

Diverge wide, converge by reaction

The cheapest thing in this way of working is volume, so I spend it freely, then I converge not by writing a spec, but by reacting.

The agent owns the mechanics: flexbox, media queries, React state. I own the taste.

Twenty wordmarks → one mark

For the brand I asked for twenty wordmark explorations, each in a different open-source font, each playing with the product's own language: cut lines, kerfs, crop marks, the cut/crease/bleed layers as a little dot triad.

Then: how does it read in the nav, and in colour?

I narrowed to the type-forward lockup ("Die Cut Studio" in Bricolage Grotesque with the triad as the "o") and pressure-tested it in context and across grounds.

The finalist: "Blush"

One light ground, one dark ground, the triad in the brand's three layer colours, on white where it has to live. Decided, then wired into the real Logo.tsx and favicon, the same session.

Four control layouts → floating cards

The editor's controls needed a new home. Four interactive layouts: a vertical accordion, an icon-tab inspector, dock cards, and free-floating cards over a full-bleed canvas. I picked the floating cards, then refined them through reactions: the title moved to the toolbar, "collapse all" went away, the collapsed pills turned white, and a clipped shadow got fixed.

Keep the irreversible decisions human

The agent generates relentlessly, but it doesn't decide the things that are mine. Solid rings or dashed? Tab bar or bottom sheet? How much blush in the UI? It stopped and asked, with previews, every time. AI-native isn't "the machine has taste." It's "the machine clears everything off your plate except taste."

Verification is part of the design

The agent tested its own work in a real browser and caught what comps hide: pill shadows clipped by an overflow container; an Export panel that broke when collapsed (an inline display:flex beating the CSS that hides collapsed content); emoji icons that looked inconsistent. Design QA moved inside the loop. "Looks done" and "is done" stopped being different states.

From "how does this look on mobile?" to a shipped phone app

On a phone the floating cards can't sit over the canvas, so the controls move to a FAB-driven surface. Three patterns: bottom sheet, speed-dial, and a native-feeling tab bar.

Exploration and production share one thread

The same conversation that explored all of the above ended with the chosen work in the live codebase: the new brand, the floating-card editor, a phone-first mobile editor with a Lucide tab bar and the full-sheet export preview. 273 tests still green, committed and pushed. No handoff, no translation loss.

The bottleneck moved from production to taste

I push fewer pixels and make more judgments. The skills that matter now: knowing when to diverge and when to close; feedback as crisp reaction; reading a screenshot like a hawk; and saying no. Every "this is the one" in this session was mine.

The honest caveats: the agent over-produces (curation is on me), needs hard constraints up front (open-source fonts, IP-safe, the exact palette), will confidently build the wrong scope if I don't confirm it, and has zero taste of its own. It has range, fluency, and tirelessness, but no opinion about what's good.

AI-native design isn't the automation of design. It's a new instrument: one that prototypes in the real medium, tests itself, and ships, fast enough to keep up with a conversation.