robzepdev

Verdict now reads in five languages. The thing it shows cannot be translated, so the page says so out loud. Release notes are kept verbatim, summaries are generated in English, and an Italian page carries a one-line notice about it. 🧵

Translating the layout is the other half. Four verdicts say less in Italian, because the line holds 26 characters including the label. German needed two verbs for follow: folgen takes the dative, but it forms no attributive participle.

Verdict now reads in five languages. The thing it shows cannot be translated, so the page says so out loud. Release notes are kept verbatim, summaries are generated in English, and an Italian page carries a one-line notice about it. 🧵

One rule that saved the rest: next/link is imported in exactly one file, inside a LocaleLink. A link that forgets the prefix drops the reader back into English halfway through a path, and a typecheck never sees it.

Verdict now reads in five languages. The thing it shows cannot be translated, so the page says so out loud. Release notes are kept verbatim, summaries are generated in English, and an Italian page carries a one-line notice about it. 🧵

Why the summaries stay English: each extra language is one more model call per release. Cheap per call, times five languages, times every release of 37 technologies. So the notice shows up only where a second language is actually on screen.

Verdict now reads in five languages. The thing it shows cannot be translated, so the page says so out loud. Release notes are kept verbatim, summaries are generated in English, and an Italian page carries a one-line notice about it. 🧵

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

My design doc had been caught out of date four sessions running, so the typography table is now held by a test in both directions: nothing in the doc the CSS does not render, nothing in the CSS the doc does not state. The prose is still held by nothing.

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

Then the ones the eye cannot catch. A clamp in vw inside a fixed column measures the wrong thing: above ~1340px 'ok: nothing needs action' asked for 346px inside a 326px panel. The copy never grew. The clamp grew with the viewport while the column stayed 360px.

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

So under reduce the shimmer becomes a fade between 0.6 and 1 opacity. No direction, no translation, no parallax. That is the substitution WCAG 2.2 asks for in 2.3.3, rather than nothing at all.

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

Second one: I built loading skeletons and then could not see my own loading state. Fill #161a20 on a #0f1115 page is three points apart. What arrived was the 1px border alone, and an empty border reads as a broken control, not as waiting.

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

The cause is one line of my own CSS. globals.css turns off Tailwind 4's default spacing scale with --spacing: initial. In Tailwind 4 the inset utilities are built from that scale. No --spacing-0, no top-0 class. It is not an override losing. The rule is never written.

I spent this week redoing the look of every screen in Verdict. The design work is the boring part of this thread. The interesting part is four things the browser was doing that I could not see, and one of them had been live and broken in production the whole time. 🧵

The AI summaries are live on the tracker now. Not a local build, not a mock: this is production, grounded on the official release notes only, with a line pointer next to every claim. 126 of them went out in a single batch. Here is what that actually cost me.

Unrelated bug, same lesson. My monospace column budget assumed 0.6em per character, the number everyone knows. Fragment Mono advances 9px at 14px. 6 overflows out of 13,496 checks, all at 768px, on lines of exactly 79 characters. Remeasured, then 0 of 13,496.

The AI summaries are live on the tracker now. Not a local build, not a mock: this is production, grounded on the official release notes only, with a line pointer next to every claim. 126 of them went out in a single batch. Here is what that actually cost me.

The fix is a gap, not a bigger number. The schema now allows 140. The prompt still asks for 120. The model keeps aiming at 120, and the few characters it overshoots land in the margin instead of throwing away a paid summary. Asking for 140 would just move the wall to 141.

The AI summaries are live on the tracker now. Not a local build, not a mock: this is production, grounded on the official release notes only, with a line pointer next to every claim. 126 of them went out in a single batch. Here is what that actually cost me.

4 of the 126 came back failed. Not hallucinations, not malformed JSON. Titles of 121 to 126 characters against a limit of 120. Everything else about those summaries was correct.

The AI summaries are live on the tracker now. Not a local build, not a mock: this is production, grounded on the official release notes only, with a line pointer next to every claim. 126 of them went out in a single batch. Here is what that actually cost me.

Real cost: $4.16. I had estimated $2.94. Off by a third. The miss is the interesting part. My median was almost exactly right, $0.0245. My mean was not, $0.033. I had estimated on a tail I had never seen.

The AI summaries are live on the tracker now. Not a local build, not a mock: this is production, grounded on the official release notes only, with a line pointer next to every claim. 126 of them went out in a single batch. Here is what that actually cost me.

The release tracker has accounts now. You follow the tools in your stack, and one page tells you what shipped in the last 30 days and whether upgrading needs work. Here is what I learned building it. 🧵

Real data found what my tests could not. The compact row said latest major for TypeScript, which has 7 releases and no major recorded, because the import stops at 12 months. And a beta sorts above the stable before it, so a preview stood in for a technology.

The release tracker has accounts now. You follow the tools in your stack, and one page tells you what shipped in the last 30 days and whether upgrading needs work. Here is what I learned building it. 🧵

Public pages stay static. The Follow button on a public technology page runs in the browser, and an end to end test downloads the HTML served to a logged in user and checks it contains no Following. Nothing personal can be cached by mistake.

The release tracker has accounts now. You follow the tools in your stack, and one page tells you what shipped in the last 30 days and whether upgrading needs work. Here is what I learned building it. 🧵

The bug the first real refusal found: I had one mutation per row, so the error stayed inside the component that was clicked and the row that had to show it never saw it. One mutation for the page, and the count, the block and the refusal can no longer disagree.

The release tracker has accounts now. You follow the tools in your stack, and one page tells you what shipped in the last 30 days and whether upgrading needs work. Here is what I learned building it. 🧵

Concurrency detail I almost shipped without: two tabs inserting at the same time both count nine and both pass. The trigger takes an advisory lock per developer before counting, so the tenth follow is decided once.

The release tracker has accounts now. You follow the tools in your stack, and one page tells you what shipped in the last 30 days and whether upgrading needs work. Here is what I learned building it. 🧵

The tracker now runs on its own. At 02:00 UTC a cron asks npm and GitHub what is new; today it found Vite 8.3.1 and the public page updated itself. Nobody pressed anything. 🧵

The trap: "new" is not "missing from my database". The initial import stops at 12 months, so that rule would have made the first detection swallow the entire history it had just excluded. New = published after the newest release I already have.

The tracker now runs on its own. At 02:00 UTC a cron asks npm and GitHub what is new; today it found Vite 8.3.1 and the public page updated itself. Nobody pressed anything. 🧵

Detection is expensive if you do it naively. An npm packument is 1 to 5 MB compressed (Vite's is 4.6). With ~40 technologies that is ~80 MB a day to download and parse inside a 300s function, almost always to learn that nothing changed.

The tracker now runs on its own. At 02:00 UTC a cron asks npm and GitHub what is new; today it found Vite 8.3.1 and the public page updated itself. Nobody pressed anything. 🧵

Vercel Hobby gives you 100 cron jobs, each limited to one run a day. I had read that as "one cron a day" and designed a single orchestrating route around it. It is three instead: detect at 02:00, submit at 05:00, collect at 11:00 UTC.