design-ship-gate
Done-gate for a web page built by a small model. Run it before saying a landing page is finished: brief coverage, then nine executable checks, one per defect seen in Qwen3.8-27B pages (fonts, OG tags,
SKILL.mdskills/design/design-ship-gate/SKILL.md
GitHubShip gate (design-ship-gate)
Run this before you say a page is finished. Read the brief again, then run every check below and paste each command's output. A check you did not run is a check that failed. Fix, re-run, then report GATE: 9/9 with the outputs. Run check 9 first: it needs only node, and a page whose script does not parse is not a page.
Why a gate and not more rules: the design rules were loaded and the pages still broke them. Nine pages from one brief, audited with real renders: more than two font families 5 / 9, no Open Graph tags 6 / 9, hero elements colliding 3 / 9, a number split from its unit 3 / 9. Every one of those was found by a command, never by re-reading the rule.
0. Brief coverage — write the list first
Before the checks, write this table in your answer and fill every row:
| the brief names | element(s) that deliver it |
|---|---|
| each style word in the brief (e.g. "Liquid Glass") | selector or section |
| dark mode | prefers-color-scheme block or toggle — a deliverable even when the brief does not say it (4 of 5 pages skipped it unasked) |
| Open Graph tags | <meta property="og:..."> — same, a deliverable when unnamed |
| the brief's language | the page copy is in it (2 / 9 answered a Thai brief in English) |
A row with an empty right-hand cell is a missing deliverable, not a stylistic choice. A page in one generic direction (a SaaS card grid for an editorial brief) fails this table before any check runs.
1–9. The checks
Set P=index.html (the page) and G=<this skill's directory> (where gate-dark.js, gate-390.js, gate-hero.js live). Node with Playwright is needed for 4–6 (npm i -D playwright in the page's folder, or NODE_PATH to an existing install); if it is absent say so in the report instead of skipping silently.
1. At most two font families. Count what is loaded and primary: every family in a Google Fonts family= link, plus the FIRST name of each font-family / --font-* declaration. Fallbacks after the comma (Segoe UI, SF Mono, Helvetica Neue, system-ui) are not families — a first draft of this check counted them and called a two-font page a five-font page. Pass: the list has <= 2 entries.
{ grep -oE "family=[A-Za-z+]+" "$P" | sed 's/family=//; s/+/ /g'; grep -oiE "(font-family|--font[a-z0-9-]*|--ff[a-z0-9-]*)[[:space:]]*:[^;},]+" "$P" | sed -E "s/^[^:]+:[[:space:]]*//; s/['\"]//g; s/[[:space:]]+$//" | grep -viE "^(var\(|sans-serif|serif|monospace|system-ui|ui-|inherit|initial|-apple)" ; } | sort -u
2. Open Graph present. Pass: all three lines print.
grep -oE 'property="og:(title|description|image)"' "$P" | sort -u
3. Dark mode exists. Pass: at least one match.
grep -cE "prefers-color-scheme|data-theme|\.dark\b|theme-toggle" "$P"
4. In dark mode every stat is visible. The failure was color: var(--paper) on a stats band after dark mode was added later — in the dark theme --paper went dark too and the numbers vanished. The script reads text nodes, because a stat is written 9.4<span>B</span> and an element-level test skips it; "invisible" is a luminance distance under 60 to the effective background, which is what the audit saw. Pass: invisible: 0.
Pass: invisible: 0 with candidates: above zero — a run that examined nothing prints the same invisible: 0 as a clean page. A stat is any text node under 40 characters carrying a digit, so 9.4B users counts; and where no element paints a background the comparison is against the real canvas, not white (both #359).
node "$G/gate-dark.js" "$P" # invisible: 0 candidates: n
5. No horizontal overflow at 390 px — with overflow-x:hidden removed for the test. One page hid a 56 px overflow behind that rule. Pass: overflow: 0 and clipped: 0 — an inner wrapper with overflow-x: hidden does not scroll the page, it throws the content away, and reading only the page width called that clean (#359).
node "$G/gate-390.js" "$P"
6. No text collides with the headline. A column ruler 01–12 sat across the h1; a glass lens sat over "EST. 1998". Both were text over text, and the ruler was even below the headline in z-order, so the script does not look at z-index: any positioned element carrying its own visible text whose box overlaps the h1 box is a collision. Decorative blobs and glows carry no text and pass. Pass: colliding: 0. The fix is to move the text (the ruler into the gutter, the lens off the tagline), not to hide it.
node "$G/gate-hero.js" "$P"
A page with no h1 prints colliding: 0 and note: no h1 found — the check cannot run, and reporting that as a collision sent the model looking for one to fix (#359).
Known miss: a glass lens (no text of its own) over a tagline that is not the h1 is not caught — 2 of the 3 audited collisions are. Look at the hero once with your eyes as well.
7. A number stays on one line with its unit. 100B+ split into 100 / B+ because .big span{display:block} also matched the unit's <span>. Pass: nothing prints (no descendant span rule under a numeric block), and ≈5B+/day-style stats use white-space:nowrap.
grep -nE "\.(big|stat|num|n|value)[a-z-]* +span *\{" "$P"
Use .big > span for a deliberate line break, and put the unit inside the same nowrap element as the number.
8. Nothing under 24 px is tracked tighter than −0.025em. A wordmark at 22 px with -0.03em read "Goole". The command lists every tighter value, em and px alike (the first version stopped at -0.09em, so -0.1em and -1px — the tightest values there are — were invisible to the check that exists to find them; review 2026-09-06); grep cannot see the font size, so for each line printed, name the selector's font size in the report. Pass: every printed line is a heading or display number at ≥ 24 px (the tracking hack is for those only); anything smaller goes back to letter-spacing: 0.
grep -nE "letter-spacing: *-(0\.(0(2[6-9]|[3-9])|[1-9])[0-9]*em|[0-9.]+px)" "$P"
9. Every script the page ships parses. A Qwen3.8-27B page passed checks 1, 2, 3, 7 and 8 (no Playwright for 4–6) while js/blob.js could not load — Private field '#rippleCursor' must be declared in an enclosing class, left behind when the model deleted an "unused" field in a self-review edit (#375). An ES module that fails to parse takes every module that imports it down with it, so the page rendered no scene at all, and no grep sees that. A file with import/export is parsed as a module: node --check file.js on a .js with no package.json exits 0 on that very error, so it cannot be the fallback. Pass: nothing prints. A printed line fails the gate whatever the other checks say — fix it and re-run all nine. Needs only node, so run it first.
find "$(dirname "$P")" -type d -name node_modules -prune -o \
-type f \( -name '*.js' -o -name '*.mjs' \) -print |
while read -r f; do
if grep -qE '^[[:space:]]*(import|export)[[:space:]{*]' "$f"; then
node --input-type=module --check < "$f" 2>/dev/null # a module, parsed as one
else
node --check "$f" 2>/dev/null # a classic script
fi || echo "FAIL $f"
done
Report
GATE: n/9 (parse 0 failed · fonts 2 · og 3/3 · dark yes · invisible 0 of n candidates · overflow 0 clipped 0 · colliding 0 · span-rule none · tracking ok)
brief table: k rows, all filled
Paste the command outputs under it. GATE: 9/9 with no outputs is not a pass.
What this does not touch
Design choices, copy, colour, layout — design-rules and design-audit own those. This file only says whether the page may be called finished. It was calibrated on nine pages from one brief and one model; the counts above are what was seen, not a rate.