This case study tests an AI-assisted answer to a gap still open in Carbon's Button component today: no pressed or toggle variant, confirmed against Carbon's live Storybook. One token-and-spec source of truth, translated into React, Angular, Vue, and vanilla JS, then checked against the live design file twice: once by Claude, once by Claude Code, a day apart, rather than trusted on sight. Built solo as a proof-of-concept against a hypothetical organization grounded in Carbon's own public documentation, not a commissioned engagement.
The Cost of System Gaps
Carbon's own component list, no toggle or pressed variant exists.
The Human Element
Why automated pipelines still require sharp human judgment and rigorous governance
While this proof-of-concept proves that automated cross-framework translation holds on a complex behavioral component, scaling the pipeline requires distinct organizational governance.
Directing an AI translation pipeline successfully demands a precise specification and sharp design judgment. The designer caught a critical visual mismatch (square buttons against circular source components) by eye, proving that while tools like Claude Code can run deep repository integrations (git, build tooling, CI), a designer's eye is still required to detect error. The automated mechanism translates reliably, but it still requires human oversight to govern usage and negotiate scale across an organization.