Visual polish
The craft lens on the rendered surface: whether what shipped is executed well, not whether its values are on-system or its pairings legible. Most of what you flag will be "acceptable but not elite" — that is the job, and every finding carries the concrete upgrade (the trim, the token, the missing cell), never just the verdict.
When to use
- Judging a rendered surface against a Figma-grade bar inside a design review.
- A component or layout changed and renders exist across themes and states.
- Comparing a build against the design file — what differs is usually craft, not intent.
When NOT to use
- No render. This lens has nothing to read.
- The value is off-system — design-system-consistency. "That gap is 13 px in a 4/8 system" is drift; "those two gaps read identical and should not" is polish.
- The pairing is illegible, or the target too small — frontend-a11y. A polish nit never outranks a measured success-criterion failure.
- Deciding what the bar is — designing-elite-ui.
What to look at
- Optical vs metric alignment. Glyphs carry side bearings and asymmetric mass, so a
box-centred element is not eye-centred: a play triangle in a round button reads
left-heavy, and a cap-height label metric-centred in a pill sits low because the empty
descender space counts. Fix by trimming the text box (
text-box-trim/text-box, where supported) rather than hand-tuningline-height, and nudge a glyph toward its optical centroid — give the direction and the reason, not a magic pixel count. - Hairlines and half-pixel blur. A border landing on a fractional device pixel
renders grey and soft. Usual causes: an odd container height with a centred child,
translate(-50%)on an odd width, a fractionalscaleon an ancestor, a 1 px rule inside a transformed parent. Recognise it as a divider crisp on one side and mushy on the other, or one that vanishes at a particular zoom. Fix the geometry — even dimensions, integer offsets — not the colour. - Rhythm has to read as grouping. Grouping is perceived from relative distance, so the gap inside a group must be visibly smaller than the gap between groups. When both land on the same legal step, every value passes and the hierarchy is gone — a form where each label looks equidistant from its own field and the next one. Fix: two distinct steps, far enough apart to read — the ratio they must clear is empirical-design-principles's, not this lens's. spacing-system owns the scale; this lens owns whether the chosen steps read.
- The state and content matrix. Two axes, both usually under-covered. States: rest, hover, focus-visible, active, disabled, loading, error. Content: empty, one item, long string, overflow, slow network. Flag cells that are missing and cells that are present but indistinguishable — hover and active rendering identically is the same defect as no hover at all. A disabled state built from opacity alone often drops its label under the contrast floor; that half is frontend-a11y's.
- Theme parity beyond colour. Dark is not light inverted. A drop shadow is a
light-surface device — over near-black there is little left to darken, so elevation has
to come from surface lightness, which is why Material's dark guidance makes higher
surfaces lighter rather than shadowing them harder. Low-alpha borders that read on
white disappear on near-black, and assets bake the assumption in: a raster logo with a
white matte, an SVG with a hardcoded fill where
currentColorbelongs. - Direction and locale. In RTL the layout mirrors, and so do direction-encoding
icons (back, next, indentation); icons depicting real objects — a clock, a printer — do
not, and a blanket
scaleX(-1)over the icon set is the usual bug. Long-locale strings are the honest test of a fixed-width label. - Clutter and duplication. Two affordances doing one job, competing focal points, chrome restating what the content already says, decoration carrying no meaning.
Silent failures
- Captured at one DPR. Hairline blur and antialiasing differences only show at 1×; a 2× capture flatters the surface. Say which DPR the evidence came from.
- Captured at seed content. "Jane Doe" and three rows hide every truncation and every empty state. Re-capture with real strings, one item, and none.
- The fix landed as a magic number. A 2 px optical nudge hardcoded is drift the moment density or type changes — it is a finding for design-system-consistency on the next pass. The fix is a trim, a token, or a geometry change.
- Motion judged from stills. A transition that jumps, a hover that shifts layout, an entrance animation replaying on every re-render: none of it is in a screenshot (record it: motion-continuity).
- Nits manufactured to look thorough. A polished surface reported as polished is a valid output; padding the list costs the reviewer's credibility on the findings that matter.
Verification
- Every finding names the element, where it sits in the capture, what is off, and the concrete upgrade. A verdict without an upgrade is not a finding.
- No finding proposes a raw value as the fix where a token, a trim or a geometry change is available.
- Every mode the QA gate's step 6 names was compared; the evidence states its DPR and content state.
- Findings that are really drift or really a11y were routed, not re-filed here — severity never ranks a polish nit above a measured failure.
- A genuinely polished surface gets one line saying so, naming what was exercised.
Cross-references
- REQUIRED BACKGROUND: designing-elite-ui — the bar this lens applies, including the one-role-one-axis rule behind clutter findings. A critic with no encoded bar finds nothing.
- Siblings, and the boundary with each: design-system-consistency — off-system values · frontend-a11y — measured contrast, focus, target failures.
- Scales these judgements sit on: spacing-system (rhythm steps) · type-scale and line-height-grid (why a label sits low in its box) · component-sizing-principles (siblings sharing a rung).
- Code side: web-interface-guidelines-review — state and transition rules readable from the diff. Above: reviewing-design-work orders the lenses and orchestrating-elite-agent-qa drives the browser pass producing this lens's states.
Sources
- CSS Inline Layout 3 —
text-box-trimand MDNtext-box— the standards-track replacement for hand-tunedline-heightcentring, plus its support. - MDN
devicePixelRatio— why a 1 px rule is not one device pixel. - Material 3 — elevation: higher surfaces are expressed lighter, which is why shadow-only elevation fails on dark.
- Apple HIG — right to left for which glyphs mirror.
- Design principle, not a spec: grouping read from relative distance (Gestalt proximity). House convention: "fine but not elite" is a legitimate finding, and so is a clean verdict.