A chat that lets a design system build for itself: ask for a component and get the component, sketch a screen from a sentence, make it a real prototype, and have something that did not build it check the result. Every choice is traced to a contract, every job touches only what it should, and a person decides.
A chat over a design system where Claude Code builds and the system decides what it builds from: answers from the contracts, sketches, prototypes, revisions and an independent check.
An AI that can build UI is not the same as one that builds from a design system. Asked for a screen, it makes something that looks right: a lookalike button, a new card no one approved, a number it invented. Nothing says where a choice came from, and the tool that built the screen is also the one that says it is done.
Every layer that was missing had already been built here, by one person: Agnostic, the design system, 174 components from atoms to organisms (41 atoms, 90 molecules, 42 organisms), each with a contract and usage rules, 47 of them conversational; 19 pattern contracts for how components combine into screens; the Alfred CLI, a command line that answers from those contracts; the producer skill, which builds screens from them; and Concierge, a separate command line that inspects work it did not make. Using them meant a terminal, file paths and knowing which command to run. Ask the system is the layer that puts them in one conversation.
The model is not the hard part. A design system needs AI to build from its rules, leave evidence, stay inside its job and be checked by something else.
The value is not the chat. It is the layers under it, each built by one person: the components, the contracts, the patterns and their usage rules, the conversational rules, the Alfred CLI, the producer skill and Concierge, which Claude Code can use directly too. The chat is the way in for people who do not live in a terminal.
For one-off engineering, Claude Code alone is better. This earns its keep on repeated design-system work, where consistency and evidence matter more than speed.
The first full check after the work found six problems, one of them a pin to a commit a squash had removed. Checking only before a deploy is fast to build with and finds things late; both were true here.
Home About Using AI Ask the system Raptive Onboarding Nextdoor Faves Nextdoor Business Design system Trulia Synthesizer Alfred CLI Resume