CCDS: a design system that ships itself

The problem

A drawing of a screen answers questions about layout. It cannot tell you what the keyboard does, what a screen reader says, or how the page behaves on a phone. So every approved drawing gets rebuilt in code, and the rebuild is where the time, the money and the details go missing.

Prototypes made of the product

CCDS is a sandbox that a designer or product manager runs on their own laptop. It uses the real component package and the real ClearCo tokens. So anything built in it is made of the same material as the product and does not have to be rebuilt.

The rules are enforced rather than written down. Components use named tokens, never raw colors, and a build fails if a token is missing, a value drifts or the docs name something that does not exist. A rule that fails a build survives a deadline. A rule in a wiki does not.

The design language, accessibility standard, voice and component specs live in the repository as plain markdown. People read them, and so do coding agents, before they generate anything. One specification serves both.

Adoption

I built it in two weeks, working with an agent. Two days after I showed it, someone forked it into their own repository without being asked. Nobody nodding in a meeting has ever told me as much as that did.

AI built in

My team built an assistant into the sandbox, and it answers questions about the system itself. The model first picks the reference pages most likely to hold the answer, then answers only from those pages, with citations that link back to them. If the answer is not there, it says so instead of inventing a component.

Some people in the test data asked to be erased. A build gate checks every piece of text a model could be handed, and fails if one of those names appears anywhere. A privacy promise that nothing tests is only a promise.

Where this is going

I am still at ClearCo, leading the team that builds on this. The aim has not changed: stop paying for design twice. A static mockup is paid for once to draw and again to rebuild, and then it is thrown away. The first version of anything we design should already be working code. Customers can use it, engineers can ship it, and nobody has to translate it.

Every screen we prototype this way is one fewer handoff, one fewer rebuild and one fewer week spent waiting to learn what a customer thinks. That is where I am taking the team.