Blog

5 min readfeature

The carving round: the letters draw themselves

Glyph draw-on for the decode panel and the spellbox, a decode fill that stays AA at every frame, keycap chips, and copy buttons that finally speak.

  • #landing
  • #design

One round ago the page learned to answer you: the kicker decodes, the headline settles, the route bar scrubs, the nav names your band. This round — v0.0.37, "the carving round" — carries the remainder of the same brief: what effects, animations and elements can make the landing page stunning and intuitive? — asked again with eight effects already shipped, so the only candidates left are the ones that survive the page's four standing motion invariants and do not repeat anything the page already says. The full survey, with citations and the rejection reasons, is docs/landing-carving-research.md.

The round's namesake came from inside the repo. The Glyph component — all 28 hand-authored Unown forms — has shipped a stroke-reveal hook since the alphabet was drawn: a reveal prop that walks each path's strokeDashoffset from nothing to the finished letter. It was never wired to anything. This round points it at the two places where a glyph is the subject.

What shipped

The draw-on, twice. The alphabet band's decode panel — the terminal window that decodes whichever form you pick — now carves its subject when you first scroll to it, and re-carves on every hover and every pick. The finale's spellbox (type a word, read it in the script) draws each letter as it arrives, a touch faster, because typing is the faster gesture. The driver is a new DrawGlyph in effects.tsx, and its honesty contract is the same one every effect on this page signs: the server ships the finished glyph, so no-JS, crawlers, pre-hydration and reduced-motion visitors all see the complete letter; the carve runs only after an IntersectionObserver confirms the glyph is actually on your screen; and the moment the draw finishes, the dash ladder is removed and the resting form is plain paths again.

The decode fill. The finale's lead paragraph — Clone it, run one session, come back tomorrow with memory that already knows what you told it… — resolves word by word as you read into it, a left-to-right wave from the page's faint body token to its dim one. This is the Apple-style scroll-linked word reveal, spelled the only way this page can spell it: those pages animate opacity, and a 40%-alpha frame of body text on this navy canvas measures about 2.4:1 — an instant fail against the WCAG AA contrast gate the page runs in both motion states. So the fill animates color between two AA palette tokens, every frame as readable as the finished state, scrubbed by the reader's own scroll so it fires at no one. Under reduced motion, and on any browser without scroll-driven animations, the paragraph is simply there.

The intuitiveness half. Three smaller answers to the same brief:

  • Keycap chips. The Konami code — the page's one easter egg, and the one signal that this product is keyboard-first — now wears a keycap in both of its homes: the alphabet's hint line and the finale's receipt. Same words, new shape: a mono-faced chip with a 3px keycap depth, sized off the page's own badge grammar. (Not the pixel face — Press Start 2P's latin subset carries no arrows, which is also why every ▶ on this page is drawn as SVG.)
  • "Copied," spoken. The copy buttons have always swapped their icon to a checkmark for two seconds. But an icon swap is silent to a screen reader mid-page, and so is an aria-label change — WCAG 4.1.3 (Status Messages) exists precisely for this gap. Each button now carries a polite status region that says "Copied" for exactly the window the checkmark is up, and says nothing at any other time.
  • Balanced headings. The section headings join the H1 in text-wrap: balance — the browser now breaks a long heading so it blocks into a shape instead of stranding one word on its own line.

What the verifier caught

The round's Agnes VLM review (the repo's review:vlm lane, over the apihub endpoint) earned its keep exactly once: the first cut of the keycap chip measured 21.5px inside a 16.5px text line — it grew its host line and overhung the baseline by 6.5px. Compacted to the badge-proportioned metrics, re-reviewed: PASS. Its other suggestions — scanlines, blinking text — are things this page has already rejected on the record (a Game Boy Advance is a TFT panel; nothing on this page loops for decoration), which is the review lane working as designed: a critic's findings are checked against the contracts, not just shipped.

The verification pass also caught something better than a styling nit: the scroll-budget gates were measuring a moving target. The session replay in the hero autoplays once — about 111 words of transcript typing in over nine seconds — and the gates measured at "page load plus a fixed settle," which meant the same build could record counts up to ~90 words apart depending on when the clock caught the replay. Worse, at the replay's finish the transcript counted twice (the animated screen plus its screen-reader-only copy, and innerText counts both). Both gates — the audit script and the CI spec — now wait for the replay to finish and count readable words once, subtracting the screen-reader lanes. The result is deterministic, the twin gates agree, and the budget number is unchanged: 13.3 viewports / 3,153 readable words against 13.5 / 3,320. The old green numbers were riding on mid-replay measurement, not on a smaller page.

The receipts

  • Research + rejection reasons: docs/landing-carving-research.md
  • The resumable session record, with every gate's measured result: docs/carving-round-session.md
  • The contracts in the page's own doc: src/components/landing/AGENTS.md
  • The tests that pin it all: carving.test.tsx (15), plus the full suite at 374 passing; e2e 15/15 including axe-core zero-violation in both motion states; contrast audited on 1,812 text nodes in both schemes; a 16-expectation real-user browser pass that caught the carve mid-draw, typed HELLO into the spellbox, and sampled the fill's two color states.

The next round's queue, recorded in the research doc: a cmd-K palette over the page's sections, and hover-previews on the blog link. Both need their own surfaces and their own a11y passes — the same bar this round had to clear.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts