ux-ui-agent-skills

Screens an agent built.

A design-system kit for Claude Code. Whole product screens, not component sheets - every one rendered from a single token theme and signed off by a real gate, in light and in dark.

20product screens
51objective checks
2144design systems
0chart libraries
Meridian Terminalgated, light and dark
Meridian Terminal: a candlestick chart with volume and a moving average, an order book with depth bars, a market depth curve, a time and sales tape and a watchlist
Ten chart types on one screen, every one inline SVG

The kit ships 2,144 named design systems, each as a full written specification.

The difference is not an opinion

Drag it.

Left is the same dashboard written the way a model writes it when nothing stops it. Right is the reference app in this repo. Both are real files you can open; the left one is tests/fixtures/bad/slop-screen.html, and ten of the gates reject it.

The same dashboard built to the kit's rules: one hero revenue card carrying the revenue figure at display size, then three quieter stat cards for active users, churn and NPS The same dashboard as a model writes it unprompted: a purple gradient header with emoji, four identical stat cards with no one leading, grey low-contrast labels, and a blue Delete Account button

Measured, not claimed

51objective gates, all green
56pages rendered here
26runnable skills
2144design systems
0chart libraries

Five websites

Not only dashboards.

A product launch, a furniture shop, a long read in Thai, a neighbourhood restaurant and a public service page. Each takes its visual language from one design system in the library, and each passed the same gates as every screen below. The brands, people and figures are fictional, and every page says so.

Twenty screens

Whole screens, not component sheets.

A checkout, a patient chart, a threat console, a factory line, a railway timetable and a trial readout. Same tokens, same render checks, different job.

Everything else

23 harnesses and six more screens

Every variant and state of each component, rendered so the gates can measure them, plus the reference app, two aesthetic studies, the two dense consoles and an iPhone Duo simulator. Twenty-nine links, twenty-nine files in this repo.

A passing gate is not taste. It is the floor.

Fifty-one checks prove contrast, keyboard, target size, overflow and intent. They cannot prove any of this is good, so the kit also runs a critic that argues for rejection, and says so on the page.