What they did
Metcalf replaced a PHP portfolio from 2014 to 2023 with a design system and a site on top, built by directing Claude Code over two weeks. The agent got a private folder of 152 files: a resume, the old site, project notes and 104 images of past presentations, plus a voice guide it turned into writing rules. It asked questions four at a time with a recommended answer each, and showed three candidate looks as screenshots before any component existed. Colours, type and motion were given names by job, with the build failing on a raw colour value. Every component lives in Storybook with light and dark examples, which agents can query through Storybook's MCP server. A decision log of taste calls is read at the start of every session.
How it went
The site has case studies, a presentation, private pages, a small phone app and a public Storybook. Screenshot tests in both themes fail if more than 1% of pixels move; that caught a list that stopped folding after one link was removed. All 554 checks take about four minutes, so a script runs only the affected ones, 32 checks in 31 seconds for that change. The automated drift check passed, but three agents comparing about 255 doc claims to the code found 46 that did not match, including pages never pictured in dark mode and 36 pixel buttons under the 44 pixel tap rule.
Worth knowing
Not every rule is enforced: nothing checks that components use the named springs for motion, so that holds only through review. Reading the docs against the code with separate agents found problems the automated checks missed entirely. A UX designer and a content designer also reviewed the site by hand.