Repository navigation
Five websites, not only dashboards. examples/sites/ holds a product launch,
a two-page shop, a long read in Thai, a neighbourhood restaurant and a Thai public
service page. Each takes its visual language from one design system in the
library, every brand, person and figure on them is fictional and says so, and each
page passes the same gates as every other example. The design critic reviewed all
five twice, and each was revised against both rounds of findings. The long read
names no real city, river, flood or app, and each footer says in plain words what
was tested.
The contrast gates read oklch. Chrome reports a computed colour in the syntax
it was written in, and measure_render.mjs and verify_states.mjs pulled the
numbers out of the string, so oklch(0.2 0.014 30) read as rgb(0.2, 0.014, 30).
A page whose theme is written in oklch, which every Tailwind v4 palette is, failed
contrast everywhere: a live site checked against the kit came back with 364
failures that were not there. scripts/_color.mjs paints each colour into one
canvas pixel and reads it back, so oklch, oklab, lab, lch, color() and
color-mix() all arrive as the sRGB the screen shows. lint_intent.mjs,
slop_tells.mjs and taste_audit.mjs read colours through it too. A pair of oklch
fixtures holds it: the old gates failed the good one on 6 lines and reported every
line of the bad one as 1.00:1; the new ones pass the good page in light and dark
and report the bad grey at its real 2.01:1. Reading colours right exposed one
thing the misreading had hidden: measure_render measured .sr-only labels, which
are clipped to 1px and have no contrast to meet, so it now skips them.
Spacing is a rule, and a gate. .claude/rules/components.md has a new section,
"Spacing: rows, gaps and gutters", and scripts/verify_spacing.mjs measures the
two parts of it that can be measured on the render:
- Rows of cards share a height. Boxed siblings that start on one line end on
one line. An item that spans two grid rows is not compared. - Nothing touches. A control, badge or card keeps 8px from its neighbour,
stacked or side by side, including after a row wraps; a label may sit 4px from
its own field, and the segments of one control may join inside its box.
"One relationship, one gap" and "one gutter" are in the rule and are read by the
critic, not the gate: both were tried on all 56 example pages and both failed good
pages. The gate runs on every example page in both themes inside the existing
overflow check, so the report still has 52 checks. Fixing what it found touched
seven existing pages and the two shared template stylesheets.
The design critic's second pass found two rows the gate let through, a tile held at
the top of a row of product cards that draw no box of their own, so the rule now
also compares a boxed card with a neighbour that carries a picture, and it reads
rows inside lists. That found two more on the banking and logistics templates.
Thai text is a rule, and a gate. .claude/rules/typography-and-spacing.md has
a new section, "Thai text", and scripts/verify_thai.mjs checks every page with
Thai on it:
- No compound or count split at a line end. Chrome breaks Thai with ICU's
dictionary, which holds the parts of a compound and not the compound, so
ใบ / อนุญาต is a legal break to the browser. The gate knows a list of compounds,
the noun prefixes ความ, การ, ผู้ and ชาว, and numbers with their classifiers.
A name it does not know still needs a Thai reader. - Lines far enough apart for Thai. Measured in the page's loaded font: the room
a Thai line needs above (ที่, นี้) and below (ุ) may not reach the next line. The
first public service page set its heading at 1.3 in Sukhumvit Set; the ink
cleared by 12px and the critic still read the tone marks as part of the line
above. The gate fails that heading and passes the 1.48 it ships with.
It runs inside the same overflow and spacing check, so the report still has 52
checks.
Security and docs: pinned npx versions, SECURITY.md, init --dry-run, graceful
partial-install fallback, consistent stats. An outside review (SkillCritic,
2026-10-29) passed the kit with caveats. Each caveat is answered here, and
checking the claims before writing them turned up more than the review did.
- Every
npxcommand names an exact release. The README, docs, skills,
rules, MCP configs and the CLI's own help all say
npx ux-ui-agent-skills@<version>, andnpm versionrewrites every pin in the
bump commit. Some docs saidnpx ux-ui-skills. That is a bin inside this
package, not a package on npm, so outside a project with the kit installed npx
would have looked the name up on the registry, where anyone could register it.
npm run pins:checkfails on any unpinned or stale command, including one
that prose wraps onto the next line. - SECURITY.md, file by file: what
initwrites, what each of the three
plugin hooks does and when it fires, the one network request (design-system
specs from GitHub, or jsDelivr if that fails, at the indexed tag,
sha256-checked), where the MCP server caches, that there is no telemetry, and
how to report a vulnerability. Writing it found three claims that were false:
init --helpsaid it installs hooks (only the plugin does), the fast-gate hook
said it ran four checks (three), and a clone'sinitcopied
scripts/__pycache__into the user's project. init --dry-runis exactly true. It ended with "N copied" while copying
nothing; it now says how many files would be written and that nothing changed.
A test runs every install command both ways and fails if the lists differ or
the dry run creates anything. Another fails if the tarball ships a file git
does not track, or gains an install-time script.- A skill installed alone degrades instead of stopping.
npx skills add
copies the skill folders and nothing else. Each skill now opens with a
Requirements section, and without the kit 21 of them run in guidance-only
mode: the reply says so in its first line, states no measured number, and ends
by naming the gates it skipped. Tried in a real session on a project holding
onlya11y-audit: the reply opened with the guidance-only line, wrote "not
verified" where a ratio would go, and listed the three gates it skipped. The
five skills whose job is running the kit stop and name the install command. - No browser reads as no browser. With Playwright installed and no browser
that launches, 16 render gates died on a stack trace. All 18 now print
SKIPPED (no browser)and say nothing was measured, orREQUIRED, FAILING
under the release bar, asmeasure_renderalready did. - Every count is derived.
scripts/stats.mjscounts skills, gates and how
many need a browser, eval gates and briefs, indexed and measured design
systems, and the mean SKILL.md size intostats.json, andnpm testfails if
any README, doc, manifest or skill states a different number. The review saw
50 and 52 gates, 138 and 2,144 systems, and 27 skills. The kit has 52, 2,144
and 26: 138 was the library before 3.0, and the 27th SKILL.md is a broken test
fixture. The check found three stale figures of its own: 35 of 51 gates in
CONTRIBUTING.md, 14 eval gates in the guide, and 50 gates in a CLI comment. - Prerequisites up front in the README (Node 20, Python 3, Chrome for 34 of
the 52 gates), and Playwright pinned to an exact version in devDependencies. - Neutral wording in two places a scanner read as prompt injection: the
critique skill's "You are now running in a forked context", and the critic's
claim of fifteen years of experience.