← Back to the screen

Design system portfolio

Design record: Plateworks

The four direction lines behind this page, the generator command that produced its tokens, and the pairs it was measured against.

Packaging artwork review. One label proof on a canvas, annotated, with its colour separations. Brief #10.

Direction

LeverDecision
ArchetypeCanvas. The tool rail is the only permanent chrome. The dock ships open, since the thread is the content, and closing it hands its width back to the proof: 40 % → 89 % at 375 px, 95 % → 102 % at 1280. Above 1400 px the fit is height-bound and closing it changes nothing but the pasteboard, which is what a portrait document does.
Data stanceProse-first. An entry leads with the sentence somebody wrote and carries the replies to it; note 1 is an argument between prepress, account and label compliance, not a finding. The selection is shared: picking an entry draws its measurement on the artwork, picking a pin brings its entry up.
DensitySpectrum desktop scale: 32 px controls, 14 px body. The Interface scale control switches the whole page to the mobile scale (40 px, 17 px).
The one bold thingKeyboard access designed rather than inherited. Arrows pan, +/- zoom, 0 fits. Escape closes the dock and returns focus to whichever control opened it.

Second state

Desktop → mobile is a whole-interface rescale, not a touch-target patch. spectrum_tokens.py --platform-diff reports 922 of 922 platform tokens differ. The page carries the ones it uses:

TokenDesktopMobile
component-height-10032px40px
component-height-7524px30px
font-size-10014px17px
font-size-7512px15px
line-height-font-size-10018px22px
workflow-icon-size-10020px24px
text-to-control-10010px13px
text-to-visual-1006px8px

Measured in Chromium at all three widths: the Undo control goes 32 → 40 px and the body font 14 → 17 px. The page reads the scale from (pointer: coarse) and the control overrides it, which is what the skill asks for.

The sheet is a separate decision, and it is made on width. It was made on the scale, and at 1440 px the Mobile control therefore produced a 1383 px-wide sheet lying over the canvas with the zoom stack underneath it, which is the one thing SKILL.md line 52 forbids ("nothing permanently occludes the work"). Below 760 px the dock floats; the canvas counts the height it takes as spent, so the artwork centres above the sheet and the zoom stack rides on top of it. At 1440 the mobile scale is a rescale and the dock stays a 364 px dock.

Generator

spectrum_tokens.py --both > tokens.css
spectrum_tokens.py --platform-diff
spectrum_tokens.py --theme desktop --format css
spectrum_tokens.py --theme mobile --format css
spectrum_tokens.py --theme default --format css
spectrum_tokens.py --contrast

The generator wrote tokens.css; nothing here edits it. --both emits the colour axis only (671 tokens per set), so the page declares the sizing tokens it uses in its own <style>, from the desktop, mobile and default runs.

Zoom

The artwork is 90 × 130 mm, drawn at 340 px per 90 mm (96 dpi), so --zoom: 1 is the label at print size. On load the page fits the artwork to whatever canvas the open dock leaves, measured: 40 % at 375, 44 % at 390, 95 % at 1280, 119 % at 1440, 149 % at 1920. + and - step the 25 / 50 / 75 / 100 / 150 / 200 / 300 / 400 / 600 ladder from there and disable at its ends, 0 refits, and a resize (or opening, closing or resizing the dock) refits only until the reader picks a zoom. The pan is clamped so 48 px of the proof is always on screen: a canvas that can lose the document has lost it, and nothing on the status bar would say so.

Guides

The pasteboard carries trim marks and the 3 mm safety margin, and the selected pin draws its own measurement on the proof. That is what closes the gap between the thread and the canvas: note 1 says the net weight sits 1.4 mm from the trim, and the artwork now visibly breaks the safety margin at that corner. Without the guides "Show pin 1" moved a coloured dot and the reader had to take the sentence on trust.

The box, the leader and the dimension ticks are in the artwork's coordinate system and scale with it. The label is too (font-size: 3px, three millimetres of proof), but it carries a floor. a11y-probe.mjs --width 390 measured the first version at 7.7 px: 3 units times the 44 % fit is illegible, and a measurement nobody can read is not a measurement. --anno-k counter-scales the glyphs through the scale property, so the label grows with the proof and stops shrinking with it at --anno-floor: 12 px desktop, 15 px mobile, alongside every other sizing token. Measured: 12 px at 390 and at 1280, 13.5 px at 1440, 16.9 px at 1920, 34 px at 300 %. The floor is on the type only; the geometry still belongs to the page, which is what Acrobat and every other proofing tool do with review markup.

Measured contrast

PairDarkLight
neutral-content-color-default on background-layer-1-color (dock prose, 16px)12.44:113.70:1
accent-content-color-default on background-layer-1-color (links, 14px)4.91:14.53:1
--on-solid on accent-background-color-default (primary button, group pin, selected pin): gray-1000 in dark, gray-25 in light4.51:14.81:1
gray-600 control boundary on layer 14.99:14.60:1
negative / notice / positive visual marks4.90–4.92:13.72–3.73:1

Every pair clears WCAG 2.2 AA. An earlier draft of this table claimed gray-25 on the accent in dark at 4.51:1 and the light control boundary at 3.31:1; both were wrong when computed from tokens.css (4.19:1 and 4.60:1), so the label now takes --on-solid, which flips per theme. The three accent pairs fail APCA at 14 px; that is Spectrum's own semantic accent, which the system's own audit records at 4.81:1 with the note that it passes "without much room". No accent-coloured text on this page is smaller than 14 px.

Deviations from CONTRACT.md

  1. Raw hex inside the artwork and the ink swatches. The label SVG and the six plate chips carry the customer's colour data. Spectrum's own rule for a creative tool forbids a UI theme reaching into it (references/color-theming.md § Never color-manage the user's content). Every pixel of chrome is a var(--spectrum-*).
  2. Sizing tokens declared in the page rather than linked. --both emits no sizing token; the values are the verbatim output of the three runs in the Generator section.
  3. Durations authored at 90–190 ms. spectrum_tokens.py --grep duration returns nothing; the set carries no motion token. The range is the system's published 100–250 ms (references/scales-typography.md § Motion).
  4. Source Sans 3 in place of Adobe Clean, which Adobe licenses to its own products. The stack names Adobe Clean first.
  5. Two copy_lint rules kept, recorded in copy-lint.ignore with the reason per occurrence: shouty-caps and title-case fire on the printed cider label inside the SVG, on three one-word table headers read as one string, and on a heading that is already sentence case.

Uplift, 2026-10-29

What changed from the first version, and why each part is a decision rather than decoration.

Round two, 2026-10-29 (first fold)

Round four

Round eight

Round nine

Found by driving the phone as a coarse-pointer device (the review renders use a fine pointer, so they never showed it) and by raising the sheet.

Round ten

A second, cold read. Round nine's 33/40 did not hold: the page's one commit, Approve, was a single press with nothing to confirm it, no way to reopen it and a reason for refusal that lived in a hover tip.

Inventions, none contradicted by the brief: Reopen means only that the approval is withdrawn and the proof is back in review (every property was already editable); comment 4 resolves at 0 mm of overlap; the dialog's wording ("This marks v7 approved for plates") restates the existing stamp; the definition of the 280% limit is the general meaning of total ink and is not tied to a stock. No prices, no new clinical or legal claim.

Scores before and after, ten heuristics: visibility of status 3 to 4, match with the real world 3 to 3, control and freedom 3 to 4, consistency 3 to 3, error prevention 2 to 3, recognition over recall 3 to 3, flexibility 4 to 4, minimal design 3 to 3 (three bands under the sheet on a phone), error recovery 2 to 3, help 2 to 3. Total 28 to 33.

Left: the callout pin 2 still covers the first letter of "Dabinett" at phone fit (severity 1, declined in round nine); the disabled-looking Approve is an icon alone at 390 and relies on the dialog for its name (severity 1).

Round eleven (2026-11-10)

Three severity 2 defects from the independent score (31/40). Each was reproduced first (1440 light, 390 dark and 360 light as touch).

Inventions: none. The 20 % tint and the --hit-target name are the page's own; the desktop stepper is 4 px wider than before (32 against 28).

Round thirteen (2026-11-10)

Reservation from the independent re-score (34/40): "phone pins hide the artwork text they annotate". Reproduced at 390 dark and 360 light (touch, mobile scale): pin 2 covered "Cider apples", "Dabinett" and the INGREDIENTS heading, pin 3 covered the barcode digit "5" (and the first bars at 360), pin 1 covered the estimated mark after "500 ml". The mobile-scale disc is 30 px, which at a 72 % or 51 % fit is 11 to 15 mm of label.

Inventions: none. The 28 px lead and the per-pin factors are layout, not product facts.

Follow-ups on the same reservation (coordinator, same round):

Inventions: the wording "Type 5.2 pt" (the panel names the comment "Ingredient type size").

Round fifteen (2026-11-10)

Independent score of round fourteen (34/40): one severity 2 and two reservations.

Inventions: none. The 76 % threshold, the 1.25 threshold and setOff's choice of where focus goes are layout and keyboard behaviour, not product facts.

Round nineteen (2026-11-10)

Independent score of round eighteen (37/40): one severity 2 and three severity 1 items. The severity 2 reproduced at 1440 light before any edit; the reason at rest and the 0 key were read from the code (opacity 0 at rest; fit() with no say). Letters below are the reviewer's.

Not done: the 28 px dock grip at 390 (reviewer's sev 1 (c), not in this round's list).

Inventions: none. The two announcement strings reuse the page's wording; "Selection cleared." and "Pin N selected." are new status words, not product facts.

Round twenty-one (2026-11-10)

Inventions: none. The tooltip shows the history label the page already writes plus the two shortcuts the page already binds (Ctrl Z, Ctrl Shift Z).