LLM becomes an Award-Winning App & Site Designer

You can spot an AI built site in about two seconds. Same layout, same soft blue, same tidy little uppercase label stamped above every section. Apps are worse: the same stat cards, the same purple gradient, a dashboard where a product should be. Your agent designs from memory instead of looking at what good work in that category looks like right now.

This changes the order of operations. Before it writes a line of code, your agent pulls the live Google Fonts catalog, searches what the leaders in your category are doing this month, and lists the colors that physically exist in your subject's world. A letterpress shop has ink, lead type and oiled steel. A dive bar has neon, vinyl and amber whiskey. The palette comes from there, not from a template.

Then it writes three competing directions and has to justify the winner against the other two. The first idea is always the one every other agent would have had.

Now for native apps

V4 adds a full iOS and Android module. Your agent designs with real platform rules: system navigation, Liquid Glass on iOS 26, Material 3 Expressive on Android, safe areas, 44pt and 48dp targets, Dynamic Type, and the four states every screen needs. It builds each app around one object that is the interface, the way Napper's night clock or Copilot's chart is, instead of a grid of stat tiles. Then it runs a subtraction pass: remove an element, look, repeat until the screen gets worse. Works for SwiftUI, Jetpack Compose, React Native, Expo and Flutter.

The proof

I gave the same brief to agents that knew nothing about each other. Without the skill they all came back with the same look. With it, each one came back with a different design, and every one of them was pulled from something real about the business instead of from a template.

Same test on a mobile app brief for v4. Before the fix, two agents built the same screen. After, two different apps, as if two different designers had taken the brief. Same prompt, same model.

New in v4

A native mobile module for iOS and Android, with its own list of app specific AI tells. Three directions before one, so independent runs stop converging on the same answer. A copy rulebook: sentence case headlines, one concrete claim, no slogan formulas, no invented ratings, user counts or testimonials. A tell check script with 45 known AI design tells that your agent runs on its own output before you see it. The audit is the ceiling, the script is the floor. The skill was audited for its own AI defaults. The fixed component kit, the dated technique lists and every hard coded color, shadow and easing value are gone. Component form is derived from research now. Platform facts current to October 2026: scroll driven animations, View Transitions, @starting-style, linear() easing, anchor positioning, GSAP free, Motion, WebGPU.

What you get

The design engine, the native mobile module, the tell check script, and six specialist modules that load only when the job needs them: usability audits built on Nielsen's ten heuristics, journey maps, UX metrics, state modeling, error and empty states, and written design rationale. Plus a copy guardian that keeps page copy sounding human.

Works with Claude Code, Cursor, Codex CLI, Copilot and anything else that reads SKILL.md. Installs in about thirty seconds and it is yours to keep, updates included.