What they did
Shimomoto built a personal site with Claude Code and then spent half a day on details outside his own expertise: hanging a Japanese corner bracket past the text edge, line-start kinsoku, lone words on a last line, link underline position and dark-scheme contrast. None of it was covered by his check script, so any later CSS change could break it unnoticed. He had the agent split each detail into a unit recording what the rule is, what moves it, how it is measured and why it matters, with criteria taken from JLReq and WCAG rather than from his own head. Four reviewer agents, for Japanese typesetting, English typesetting, accessibility and art direction, are spun up fresh for every review and measure the built pages in a real browser, without seeing past findings. Once a unit is fixed it becomes a frozen check that all run from one command.
How it went
Checks caught real drift: on the published site the character after the bracket sat 0.84px off the edge because heading letter spacing applied to the bracket too. He verifies each check by breaking its cause on purpose, which exposed two checks that stayed green, both because of how he had broken them. Over three rounds per session he never reached a round with zero new findings, taking seven or more each time. Two fixes broke other details, one doubling English word gaps through non-breaking spaces, so he now limits each fix to the language and page a finding names. The checks run on his Mac, not in CI.
Worth knowing
A unit only means something measured against the published version; one run flagged a bug already fixed in production because it ran on an older branch. Judgments of taste cannot be delegated and checked the same way: he handed over several and later overturned two after opening the page in Safari.