The resume as a survey line
The resume is now a ruler from 2002 to today, with each role a span on it that lights up as you scroll past. There's an A4 sheet view for printing too. The component thinks it's 2026 forever, so that's one for January.
CONOR.FYI / LOG / COLLECTION
Short entries from the work: things shipped, lessons, decisions, and what went sideways.
The resume is now a ruler from 2002 to today, with each role a span on it that lights up as you scroll past. There's an A4 sheet view for printing too. The component thinks it's 2026 forever, so that's one for January.
The site started looking like a site: a Making section, a page per app, ListingGrid on every index, CareerLine on the resume, and a footer more fun than it needs to be. Plus a Vercel deploy where the build guards checked the wrong folder.
Two design system snapshots in a day: charcoal grounds and palette tokens, then a lime signal colour and inverse surfaces that follow the OS. The lime is the first colour on the site that feels like mine.
Hundreds of hard-coded UI strings, most written by Claude or carried over from the design system. A small script now pulls them out of the code and a local page lines them up next to a box for my own wording.
Both build guards read from .output/public, but on Vercel the files land in .vercel/output/static, so they were guarding an empty folder. Small fix, and a good reminder that a guard only guards the place it's looking.
I didn't know what the listing pages should look like, so a lab generated sixty versions and I flipped through them with the arrow keys. One got promoted to the landing page, the rest got thrown away, which was the point.
Eight of the first thirty studies survived and bred thirty more. Number 32, a fuzzy title over a full-height focal plane, became HeroFull on the landing page. Also: a Chrome SVG mask that had no idea how big it was.
A permanent local component lab at /_lab, with search, filters and every setting in the URL so a version can be bookmarked. Production builds use a separate route tree that leaves it out entirely, with a guard to keep it that way.
The components finally went into the pages: PageShell and Sections, the header and footer, Prose and CodeBlock for posts, and an article template that splits each post at its ## headings. A synced snapshot also quietly missed a file.
Rather than settle on the first card layout for the index pages, we built a lab with ten versions using the real content, then thirty. Section can now drop its gutters for full-width bands that reach up behind the nav.
Eleven reusable skills and technical references moved out of the writing into their own Agents section, with permanent redirects from the old URLs. The long AI article also got a tabletop hero.
Pages moved onto one standard measure, but the header still sat 48px further in. A stray box-sizing rule in the nav stylesheet was changing how Section counted its gutters. Now everything lines up at x=144.
41 components, one Sonnet agent each, six at a time. Four failures were the check, two were real bugs (a glyph Geist doesn't have, a canvas that couldn't read our colours), and every font weight is self-hosted now.
50 components in, matching the design system in both themes. Getting there meant the parity check borrowing each preview's own styling, a cascade-layer !important bug, and agents holding back two components instead of fudging them.
Every page is now a PageShell holding a stack of Sections, and components respond to their Section's width with container queries, never the viewport. It caught a 404 page with a <main> inside a <main> on its first run.
Every design system component now has a TypeScript port, done by agents in parallel worktrees and merged only when the parity check said identical in both themes. Most of the work was fixing the check rather than the ports.
Code blocks now go through the design system's own CodeBlock and Shiki is gone. Fewer languages and plainer highlighting, in exchange for one less big dependency and code that finally sits on the design tokens.
The topographic background regexed its colours out of computed styles, which falls apart with oklch() inside light-dark(). Now the browser resolves the colour and a 1×1 canvas reads the pixel back. Slightly daft, but it works.
The first real design-system component on the new site. A tiny script in <head> applies the saved theme before anything draws, so prerendered pages don't flash the wrong colours, and a stale hydration warning went with it.
Tearing the old Next.js site back to unstyled HTML on TanStack Start, with a script diffing every head tag against production so no SEO goes missing. The server bundle went from 21MB to 4.3MB along the way.
Native CSS, cascade layers, light-dark() tokens converted to oklch and checked back to the exact hex, and self-hosted fonts at 54KB a page. Also found that base.css was mostly Claude's own work rather than the design system's.
The parity check failed components that had passed all day, but only sometimes. A temp file reloading the dev server, then !important flipping order inside cascade layers. One line fixed it, and then tail ate an exit code.
Porting a 35-line component showed the real problem: 2,600 inline style declarations across 87 components. Instead of rewriting them here, one long prompt sent the system back to Claude Design to rebuild itself on CSS Modules.
All 21 primitives ported and pixel-identical to the design system. A 1/255 mismatch turned out to be page position, not colour, and three agents briefly tried to port from main, which had nothing in it to port.
One loader never stops under reduced motion, so there was no still frame to compare. The fix went in the check, not the port: freeze Playwright's clock on both pages before loading and everything sits at frame one.
The design system lived in a Claude artifact, so Claude swore we hadn't started. 564 files copied into the repo as a reference, fonts pulled down one at a time, and an upgrade to React 19 that came out byte-identical.