The problem
Every project I started was rebuilding the same buttons, dialogs, and form fields — and every rebuild drifted a little. Different focus rings, different dark-mode handling, different accessibility gaps. I wanted one place where a component gets designed once, tested completely, and then consumed everywhere, so the design conversation happens a single time instead of once per repo.
The approach
The library is built on Radix UI primitives with Tailwind and CVA for variants, published to npm as @crawfordyoung/ui. Radix handles the hard accessibility work — focus trapping, keyboard navigation, ARIA wiring — and I own everything visual on top of it. The design is dark-mode-first: every component is styled and reviewed in dark mode before light mode gets a pass, because dark is the default experience on my sites and designing light-first always left dark looking like an afterthought.
Discipline is the actual product here. Every component ships with 100% test coverage, a Storybook story, and an axe accessibility pass before it merges. Storybook is the dev environment — I design in it, not in a consuming app. Releases go through Changesets so version bumps are deliberate, and the build outputs ESM, CJS, and type declarations so any consumer setup works.
What I'd do differently
I'd design the extension model first. For a long stretch, whenever an app needed something extra on a component — a streak badge on a card, a lock icon on a calendar chip — I added a typed prop for it. Six waves of calendar work went into props encoding app concerns while a perfectly good render slot sat unused. The library now has a firm rule sorting every component into a class: primitives extend through asChild and className, composed shells expose named slots, and stateful interiors expose headless hooks with render-prop context. Apps own their domain; the library owns behavior. Getting there meant deleting shipped API, which is the expensive way to learn it.
What broke
Time handling. Calendar components originally read hours off ISO strings directly, which silently reports UTC digits instead of the viewer's wall-clock time. Three separate features had independently grown the same bug — chip labels, time inputs, and overnight event splitting all showed the wrong hours for anyone not in UTC. The fix became a library-wide invariant: displayed times always come from local date getters on a parsed instant, never from slicing the string. One bug, three homes, one rule.