Give uxskill your brand color and a short brief. It builds a design system in nine foundations, checks every color pairing against WCAG in light, dark and high contrast, and writes it only when they all pass. Arabic and right to left built in.
From one color and a short brief it works out the nine foundations below, measures every color pairing, and hands you the files.

tokens.csstokens.jsonfonts.cssart/system-report.mdText on its background must reach 4.5:1 (WCAG 1.4.3), control edges and focus rings 3:1 (1.4.11), and text in high contrast 7:1 (1.4.6). If one pairing falls short in light, dark or high contrast, nothing is written and the report names it.
Every system carries an Arabic face and its own type scale. Set dir="rtl" and the text switches to them. There is no second system to keep in step.
171 rules flag what reads as generated and judge values against your own tokens. With --render it measures the real page in headless Chromium.
Already on 3.x? Nothing is rewritten for you. Read the guide to moving from 3.x to 4.0: build a new system and move the page onto its roles, or keep yours and extend it.
Ask any model to "build a landing." You get the average of every page it ever ingested, in ten seconds, every time. Here is what the centroid looks like.
Not a bigger model. A deterministic layer that compiles a design language (warmth, contrast, geometry, type) instead of reaching for the nearest template.
Five deterministic steps, start to ship.
Industry, audience, mood, parsed into a structured intake, no guesswork.
Warmth, contrast, density, geometry, formality, motion, type, tuned, then emitted as tokens.
Primary pulled from your logo’s pixels, not the nearest CSS named colour.
Gradient headers, three equal cards, filler verbs, fails the build here.
Rendered at true 360/390px in headless Chrome. Horizontal scroll fails, not your users.
The catalogue used to be a menu: pick Stripe, get Stripe-flavoured tokens. v3 turned that on its head, Stripe, Linear, Vercel, Ferrari, Anthropic become the vocabulary the engine distils from, so output is novel every call. Deterministic, offline, no model in the loop.
1,262 structured entries, queryable and cross-referenced
Claude Code, Cursor, Windsurf, Copilot, Gemini, Codex, Zed, JetBrains, Cline, Continue, Aider, Tabnine, one install, no per-tool config. Your team keeps its editor; the floor comes with it.
Same brief, same system. No model in the engine, no telemetry, runs on a plane.
Primary from your logo's pixels; a fidelity hard floor fails any off-brand output.
Rendered at real 360 / 390px viewports, horizontal scroll fails the build, not your users.
$ pip install --upgrade uxskill$ pip install --upgrade 'uxskill[mcp]'$ npx uxskill@latest