Skip to content
← Work log

The rest of the components, and the check getting it wrong

This was meant to start with a GitHub issue. I'd planned out a work log for the site, a place for short entries at /log about the small moments in a session that don't deserve a whole post, and a skill that drafts them from whatever I'm working on. I asked Claude to implement the plan, and the plan's first line said its destination was an issue, but Claude built the whole thing instead. It was done and working by the time I noticed, so it stayed. Two things went sideways along the way. A bad entry was supposed to fail the build, and it did print the error, but TanStack's prerender then rendered zero pages and exited as if everything was grand. There's a small check after the build now that fails it if the key pages aren't there. And the /log page crashed in the browser at first, because it was pulling a server-only file (and the frontmatter parser that came with it) into the client code.

Then I pulled a fresh copy of the design system, because Claude Design had added two new components overnight: a terminal that types to itself, and a depth-of-field effect that blurs HUD layers behind a sharp front one. That left 41 components still to port, so I decided to stop doing them in groups and give each one its own Sonnet agent, each with its own copy of the repo, branch and dev server, six at a time so the laptop didn't fall over. Claude wrote a single brief that all of them read, and merged each branch as it came back.

The order mattered more than I expected. Some components are built out of others, which is easy to see from the imports, but the previews the check compares against also render other components, and those don't show up in the imports. The grid screen's demo sits inside the scrolling chapters component, the iso floor's demo needs the grid screen, and four of the HUD panels all appear in each other's demos, so those four went to one agent together. Most of the agents came back with exact matches on the first try. The interesting ones were the ones that didn't.

Four times the problem turned out to be the check rather than the port. The text that sizes itself to fill its box gave different sizes on every run, on both sides, because the check was taking the screenshot before the component had finished measuring itself, so now it waits until nothing on the page is changing. Then one of the display components came out wider in the design system's own preview than in ours. Ours was the correct one — the design system's preview loads its fonts lazily, after that component has already measured itself with a fallback font. Loading every font before measuring fixed that. Then a full run crashed halfway because freezing the clock raced real time once the machine was busy, and before all that I'd broken the type check myself with a function that turned out not to exist in our version of Playwright.

Twice it was a real bug on the site. A small ▲ in one of the stat components was a fraction of a pixel off, and it turned out Geist doesn't have that glyph, so the browser was drawing it with the size-adjusted Courier New we only keep around for while the real font loads. Those stand-in fonts are now limited to the characters Geist actually covers. And the topographic background, which draws to a canvas, came back about 70,000 pixels off. The agent called it canvas noise and stopped, which is what the brief asks for, but it wasn't noise. The component reads its colours as text, and our colours are written in a newer format than the design system's, so the canvas was getting colour strings it couldn't use. It works out the actual colours first now, and that went into the brief before the next couple of canvas components went out.

The rest were small. A handful of components are off by a pixel or two by 1/255 on glass or on hard-edged gradients, and each of those got a small named allowance like the earlier ones. The depth-of-field one is the exception: its WebGL grain layer is off across most of the field by up to 5/255, which is over the limit I'd set myself, so it's marked as an exception in the script rather than quietly raising the limit for everyone.

The last surprise came from a side note while it was all running. The display components use the full range of Barlow Condensed and Zilla Slab weights, and we were only hosting three of each, so they'd have been getting browser-faked weights. The check couldn't catch it either, because it gives both pages our fonts, so the faked weights looked identical on both. Every weight and italic is self-hosted now. A browser only downloads the ones a page actually uses.

So every component in the design system is now in the site, and all 95 checks match in both themes. Nothing on the site uses most of them yet, and the pages still look broken. Next up is planning what the pages actually look like.

rebuild
AI-OMETER— DISCLOSURE INDEX
MACHINE-LED80%