Blog

4 min readfeature

The craft axes: teaching the agent to see what elite pages never do

The second Website Excellence round: nine elite sites from production CSS, six new enforceable critique gates, and the app that passed on iteration three.

  • #website-excellence
  • #design
  • #critique

The first Website Excellence round taught unown to critique its own design output: screenshot, audit, score, fix, re-score, receipt. The second round — shipped today — taught it to see the things elite pages never do, and to refuse them at gate time.

What the nine-site sweep found

The round's research base reads Apple, Airbnb, Notion, Framer, Raycast, Resend, Anthropic, Mercury and Attio from their live production CSS — the actual token values, not blog posts about them. The pattern that falls out is stricter than any style guide:

  • Type ramps are closed and small. Apple ships 11 sizes on its homepage across 3 weights; Framer 9 sizes across 2. Attio's rendered census: 313 medium-weight elements, one bold. The elite weight ceiling is 600 — bold is the amateur's emphasis.
  • Display type tightens as it grows — mechanically. Apple tunes letter-spacing per size; Notion ships a fluid clamp that tightens tracking to −0.048em at 96px. Below 28px, tracking goes slightly positive; at captions, negative again.
  • Neutrals are tinted, never pure. Anthropic's ivory ramp shares one warm hue across all ten grays; Mercury's ink is a blue-black #272735. Two-hue grays read as two designs stitched together.
  • Exactly one accent. Apple #0071e3, Raycast #ff6363, Framer #0099FF, Anthropic clay #d97757 — one accent plus at most one interactive variant, everywhere.
  • Motion is compositing-only and quiet. Linear's homepage runs 150 hand-authored CSS keyframes and zero JS animation libraries. What died: particles, Lottie-on-marketing, cursor followers. What replaced them: restraint.

The craft axes — research made mechanical

Each of those findings is now an objective gate in the design critic (scripts/design_critic.mjs, receipt schema 2): a weight census with a 600 ceiling, a tracking-vs-size curve over three bands, a tinted-neutral hue-spread check (circular standard deviation ≤ 30°), an accent-cluster census on interactive elements, an animated-property whitelist (layout properties and transition: all are violations), a breakpoint census, and a reduced-motion guard that must exist wherever keyframes do.

Every threshold is cited in the research doc. And the axes discriminate both ways: the landing page passes every craft gate, while the known-bad calibration fixture still fails (it ships weight 700 — the census catches it).

The axes' first run against the real landing found two genuine defects minutes after landing: the benchmark HP bars animated width (a layout property), and the first gallery app shipped four transition[-]alls. Both fixed — the fill now rides scaleX() on a full-width element, identical stepped drain, compositor-only.

Ledger — the validation app

Every capability needs a benchmark. The round's is Ledger: a warm-paper fintech treasury dashboard, deliberately opposite to the landing's dark Game-Boy aesthetic — the proof the gates hold across design languages.

Its critique loop, receipted iteration by iteration:

Iteration Verdict What the tooling caught
v0 FAIL 132 contrast violations; 252px mobile overflow; a weight census tripped by sr-only table defaults; a chart spike flattening its series
iter1 FAIL A render-order crash (the audit measured an error shell) — and a persona bug: the VLM judge, pinned to the landing's brand, dinged a warm-paper app for not being dark-first
iter2 FAIL 2 gates left — and the round's own F24 receipt had re-emitted a banned utility through Tailwind's repo-wide scanner (the F20 trap, again)
iter3 PASS --enforce, exit 0: contrast 0, 8 type sizes, 0% off-grid, VLM 7.0–8.23, every craft axis green

The persona bug became the critic's --brand flag — the judge follows the design language, never the reverse. The receipt trap became ledger entry F25 with a write-time rule: the defusal happens when the prose is written, not as a gate-time follow-up.

The rest of the round

The release that shipped while its own CI was red (v0.0.42 — the e2e lane had failed, the release notes said "all gates green") closed as F21: every release job now depends on a ci-green gate that resolves CI's conclusion for the tagged SHA and refuses to mint artifacts otherwise. Six website skills (landing-page-architect, design-system-generator, ux-reviewer, accessibility-auditor, motion-designer, frontend-craftsman) and four specialist agents (design-reviewer, ux-critic, a11y-auditor, performance-engineer) are committed where every future session picks them up.

The full paper trail — the session log, the gap analysis, the validation log with every failure ledgered — is linked from the docs. Every number in this post is checkable against a committed receipt; that remains the only kind of number this project ships.

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

All posts