design-genius

A universal design language engine and bespoke design-system generator for ANY medium in existence: physical architecture, industrial hardware, operating system GUI/TUI, wearables, editorial print, luxury packaging, spatial/AR, posters, canvas shaders, instruments, apps, dashboards, and digital products. It reads a local design library, fuses two or three structurally distinct real systems, and emits one bespoke spec with real tokens, pigment formulations, and geometry. It never designs from the model's memory of what "good design" looks like.

The design language engine that never repeats itself.

Downstream Implementation: Directly implementable across modern production stacks: Tailwind CSS v4, W3C DTCG design tokens, CSS variables, React, and Android Jetpack Compose.

---

The problem it exists to kill

Every "claude design prompt" and "design skill" on the market ships the same thing: a centered hero, three feature cards, a purple gradient, a generic shadow, a headline that says "Modern SaaS." The reason is structural, not a lack of effort. Those outputs come from the model's trained average of every site it has seen, and the average of everything is nothing. It is forgettable on arrival.

Design-genius refuses the average by refusing to work from memory at all. It treats the model's remembered "good design" as exactly what it is: a source of clichés. Proof lives in files, not vibes.

How it works

A four-stage pipeline, always:

1. Harvest intent. One sharp question, never a form. It pulls audience, mood, brand voice, platform, and the single constraint that matters.
2. Reason over the library. It lists the design systems on disk, picks 2-3 structurally different ones, and reads the actual token blocks. Real hex values, real type scales. Then it fuses two into a skeleton and texture, and borrows single tokens or levers from one or two more. Every borrow is named, so the spec carries its own receipts.
3. Force uniqueness. Nine levers, each one aimed at beating the AI default: layout archetype, color strategy, type voice, motion baseline, craft and micro-polish, a signature detail, copy discipline, a slop-rejection check, and a pre-mortem.
4. Emit a buildable DESIGN.md. Real token values, a full type scale, spacing, radius, component patterns, motion, and the signature detail. No --accent: pick me. A build tool can pick it up and ship.

A hard read-gate sits at the front: if it has not opened the actual library files, it stops and reads. Designing without reading is vibecode, and the skill is built to refuse it.

The pipeline ends with a cold self-audit against ten dimensions, plus, for large or high-stakes tasks, one independent critic agent in a fresh context whose fail list gets treated as ground truth.

What makes it different

Approach	What you get
shadcn and UI libraries	Components, not identity
VoltAgent DESIGN.md systems	A static system you copy whole
Generic "design skills" and prompts	The average of everything, which is nothing
design-genius	A combined, bespoke design, one per request

Where a component library gives you buttons, design-genius gives you the identity those buttons live inside. Where a design-system corpus hands you one person's whole look to adopt, it fuses the parts that fit your product and cites what it took from each.

Sounds like you

You reach for this when a phrase from the list triggers:

• "Design a landing page for X, not generic."
• "Rebuild my site. I don't want it to look like every other AI portfolio."
• "Make a unique look for Z."
• "I want a design system, not a theme."
• "No AI slop."
• "Give me a style, not a template."

It works from a text description, a PRD, a reference URL, or a screenshot. It generalizes: it is not tuned to any one project or portfolio type, so the same skill that ships a developer portfolio also ships a SaaS landing page or a dashboard shell.

The design library it reasons over

The skill reads a local design library. You control it, you grow it, and the whole thing adapts to whatever is present. It lses before it reads and adjusts if a repo is missing.

Three repos are required:

• VoltAgent/awesome-design-md — 74 real design systems with hex tokens and type scales. The corpus and the ground truth.
• xiaopu-ai web-design — the engine room: scene defaults, interaction patterns, a motion library, style seeds, and the DESIGN.md template.
• shadcn-ui/ui — accessible component source for buildable output.

Optional repos add surface area across 32 active local repositories and skills: anti-slop kits (taste-skill, hallmark, .impeccable), design systems (astryx, coss), specialized primitives (assistant-ui, mapcn, liquid-glass-js, styleseed, tui-studio), interactive components (inspira-ui, react-bits), layout engines (pretext, scroll-craft), diagrams (diagram-design), and 3D (threeui, img2threejs). The full inventory, skip list, and catalog live in LIBRARY.md and ../README.md. The path resolves from the skill's own directory by default, or from $DESIGN_LIB when you set it.

The payoff is compounding. Drop any real design system into awesome-design-md/design-md/<name>/DESIGN.md and the skill designs in that direction too. Your taste becomes the library, and the library is what the skill reasons over.

Install (Claude Code)

ln -s "$PWD" ~/.claude/skills/design-genius   # macOS / Linux
# Windows (PowerShell), from this folder:
#   New-Item -ItemType Junction -Path $HOME\.claude\skills\design-genius -Target (Get-Location)

Clone the library, then say:

"Use design-genius, build me a DESIGN.md for <project>."

It outputs the spec plus two closing lines: which systems got fused, and the signature detail. Say "design X" for the look, "build X" once the spec is ready. There is no menu of variants; one bespoke design per run, with variants available as a follow-up.

Documentation

• USAGE.md — one-time setup, the two run modes, the critic, and how the skill gets better over time.
• LIBRARY.md — what to clone, what to skip, and how the library grows.
• CONTRIBUTING.md — the corpus submission bar and the skill-diff rules for maintainers.
• CLAUDE.md — agent instructions for anyone editing the skill itself.

Build by

Design-genius is made by Akash Priyadarshi, an AI-augmented engineering generalist who ships complete systems alone: Rust, Zig, Go, Kotlin, Dart, TypeScript, and the agent workflows that multiply him. He builds everything he uses and open-sources almost all of it.

More of his work: akashpriyadarshi.vercel.app · GitHub @AkashPriyadarshii

Design-genius grew out of a simple frustration: the same slop in every "design prompt" he tried, and no way to teach a model to design instead of average. So he built the thing that reads real design systems and fuses them, and he open-sourced it because the more design systems people add, the better it gets for everyone. If it saves you from one purple-gradient hero, the hours it took were worth it.

GitHub topics

Set these on the repository when you push (GitHub's topic limit is 20):

claude-code  claude-code-skill  design-system  design-tokens  ui-design
design  design-engineering  design-spec  non-slop  ai-design
design-skill  design-md  landing-page  ui-ux  brand-identity
design-engineer  claude  agent-skill  design-tools  frontend

License

MIT. FOSS, on purpose. The value lives in the library-routing mechanism, so the more design systems you or the community add, the more it compounds.