Child skill under Build Quality UI/UX for the level-up polish pass: when the screen works but looks generic, flat, or AI-generated, this skill applies named craft techniques (two-part shadows, hue rotation, up-pop/down-pop, single-hue HSB palettes, double-your-whitespace, named font pairings, text-on-image methods). Returns evidence-backed findings with severity and before→after token fixes; also the deep corrective lens for v0/Lovable/Cursor/Bolt output after the AI smoke test.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Child skill under Build Quality UI/UX for the level-up polish pass: when the screen works but looks generic, flat, or AI-generated, this skill applies named craft techniques (two-part shadows, hue rotation, up-pop/down-pop, single-hue HSB palettes, double-your-whitespace, named font pairings, text-on-image methods). Returns evidence-backed findings with severity and before→after token fixes; also the deep corrective lens for v0/Lovable/Cursor/Bolt output after the AI smoke test.
[{"id":"visual_craft_fundamentals.depth_color_surfaces","name":"Depth, Color & Surface Craft","summary":"Threshold-bearing rules for shadows and elevation (two-part recipe, xs–xl ladder, inset/outset taxonomy), color systems (HSB single-hue palettes, hue rotation, grey temperature, dark-mode saturation), buttons, form inputs, and icon weight.","when_to_load":["When the finding target is shadows, elevation, color palette, gradients, greys, dark mode, buttons, inputs, or icon weight.","When auditing design-system tokens (shadow scale, color scale) against operator canon."],"path":"references/depth-color-surfaces.md"},{"id":"visual_craft_fundamentals.type_spacing_emphasis","name":"Type, Spacing & Emphasis Craft","summary":"Threshold-bearing rules for typography (closed type scale, inverse letter-spacing and line-height, named font shortlist), the 4-pixel spacing system with whitespace numerics, the up-pop/down-pop emphasis framework, and hierarchy levers.","when_to_load":["When the finding target is typography, font selection or pairing, spacing, whitespace, emphasis, or visual hierarchy.","When picking fonts for a new brand surface or marketing refresh."],"path":"references/type-spacing-emphasis.md"},{"id":"visual_craft_fundamentals.ai_slop_corrections","name":"AI-Slop Deep Corrections & Text-on-Image Methods","summary":"Four additional AI-slop fingerprints beyond the ui_ux_quality_review smoke test, the deep corrective token recipes for all twelve patterns, and the numbered text-on-image methods (Methods 0–4, Scrim, floor blur).","when_to_load":["When rewriting AI-generated UI (v0, Lovable, Cursor, Bolt) after the smoke test has fingerprinted it.","When placing text over photography — heroes, thumbnails, blog covers, OG images."],"path":"references/ai-slop-corrections.md"}]
Level-up polish child. The discipline this skill enforces is craft over decoration: take a screen from "passes the beginner check" to "looks designed by someone who can see," using named, numerical techniques — never vibes-talk. Subtraction beats addition: almost every fix removes weight (fewer borders, softer shadows, hue rotation instead of opacity drop, off-white instead of border, single-hue palette instead of three accents).
This is a procedure skill at domain altitude.
This skill assumes the foundational pass is done. It is the level-up sibling of ui_ux_quality_review (which catches obviously broken; this ships obviously crafted) and the designated deep corrective lens for AI-generated UI after that skill's smoke test names the slop patterns.
Activation
The screen works and passes the beginner-mistake bar but still feels generic, flat, or assembled rather than designed.
The user asks for polish, taste, premium feel, visual craft, "make it look less like AI," spacing, type, color, shadows, or layout refinement.
Rewriting v0 / Lovable / Cursor / Bolt output after ui_ux_quality_review's smoke test fingerprinted it.
Auditing Inkprint or other design-system tokens against operator canon (shadow ladder, color scale, type scale, spacing scale).
Picking fonts for a new brand surface, or applying text-on-image overlays (hero photography, thumbnails, OG images).
Judgment
Severity rubric
high — the surface reads as unmistakably amateur or AI-generated (named slop fingerprint), or the craft issue degrades legibility (untested text-on-image, contrast loss).
medium — a named technique is violated and the fix yields visible polish gain (single flat shadow, full up-pop everywhere, off-scale spacing, pure-grey muted text on color).
low — refinement on an already-competent surface (letter-spacing tune, grey temperature, optical alignment).
Procedure
Preflight. Confirm the foundational pass is complete — if the surface has flow, state, or task-completion problems, escalate to ui_ux_quality_review first and stop; craft polish on a broken flow is wasted work. Capture the surface (screenshots or component code), the tokens in play (Inkprint on BuildOS surfaces), and whether the UI is AI-generated. Run the grayscale-first check (Erik Kennedy, 7 Rules — "black and white first"): mentally strip the surface to B&W and confirm structure, spacing, and hierarchy hold before judging any color. If it reads badly in grayscale, color cannot fix it — log a hierarchy/spacing finding (or escalate) first, and treat color as the LAST pass: one brand hue + grayscale neutrals + at most ONE accent, added only after structure is working.
The Mobbin test (pre-design imitation step — Erik Kennedy, 7 Rules — "steal like an artist"). Before proposing any net-new layout or pattern, pull 5+ real references for that exact pattern (settings page, create flow, onboarding, empty state) from real-app sources — Mobbin (filterable by pattern) over Dribbble, and never AI output, which is the average of average. Copy the moves (spacing, hierarchy, type pairing, color choices), never whole screens. If a proposed layout was not checked against 5+ vetted real references, flag it as unvetted rather than shipping it. (Skip for targeted single-domain fixes — see step 8.)
If the UI is AI-generated or the user says it "looks like AI," load visual_craft_fundamentals.ai_slop_corrections and run the deep correction pass: confirm the smoke-test fingerprints, scan for the four additional fingerprints, and write the corrective token recipe for each hit.
Establish hierarchy and emphasis before touching surfaces: load visual_craft_fundamentals.type_spacing_emphasis and review type roles, the 4-pixel spacing system, whitespace at all three levels, and up-pop/down-pop pairing on every emphasized element.
Then polish depth and color: load visual_craft_fundamentals.depth_color_surfaces and review the shadow ladder (two-part recipe), the four-cue button lighting recipe, brightness-as-elevation for low-shadow surfaces, single-hue palette discipline, grey temperature, buttons (incl. hover recolor), and form inputs. Depth and color polish only lands after structure, type, and spacing are working.
For text-on-image surfaces, load visual_craft_fundamentals.ai_slop_corrections and pick a numbered method (Scrim is the default for hero photography) — never freelance an overlay.
For a targeted question ("fix my shadows," "pick a font"), load only the reference covering that domain and run just those checks — but still flag any foundational issue you notice as Delegated. (Targeted single-domain fixes skip the step-2 Mobbin test.)
Tag out-of-scope findings for escalation instead of dropping them — route each finding to its owner per Routing (→ ui_ux_quality_review · → accessibility_inclusive_ui_review · → marketing_site_design_review).
Assemble the findings report using the Output contract below, ordered by severity then perceived polish gain, ending with the top 5 highest-impact fixes and any token diffs that codify the fix so the same paste does not recur.
Routing
Ownership map for out-of-scope findings. The Procedure sequences and tags; this table assigns the owner. Tag findings Delegated — never drop them.
Concern (out of scope for craft)
Owner
Missing states, broken flows, or beginner-bar issues
ui_ux_quality_review
Any craft fix that would reduce body-text contrast below 4.5:1 or change element semantics (e.g. removing a border that carried focus indication)
accessibility_inclusive_ui_review
Landing-page persuasion structure
marketing_site_design_review
Sequence with siblings: run ui_ux_quality_review first for the beginner bar and the 8-pattern AI smoke test; this skill ships the crafted layer on top and owns the corrective token recipes.
For marketing surfaces: marketing_site_design_review owns section-by-section conversion review, and landing_page_scorecard_funnel owns assessment/quiz funnel design — hand off structure-level findings there and keep only the craft-layer fixes here.
Contract
Every finding follows this canonical shape:
Domain: <shadows | color | typography | spacing | emphasis | hierarchy | images | icons | buttons-forms>
Finding: <named technique violated, e.g. "Two-part shadow not decomposed">
Evidence: <quoted class string, token value, or component snippet>
Severity: <high | medium | low>
Fix: <before → after token or value pair, e.g. `shadow-md` → ambient `0 1px 2px` + direct `0 4px 12px` light-gray>
Delegated: <optional sibling skill id if the fix is out of this skill's scope>
Severity values are defined in Judgment → Severity rubric.
Stop conditions before returning: every reviewed domain has at least one finding or an explicit "no issues"; the top 5 fixes are ranked by perceived polish gain; every finding carries Evidence and Severity; every Fix is expressed in concrete tokens or a named technique, not adjectives; out-of-scope concerns are tagged Delegated, not dropped.
Policy
A finding without evidence is not a finding — if it cannot be tied to a quoted class string, token, or component snippet, do not include it.
Do not produce findings without loading the matching reference module first.
Do not add gradients, glows, or extra cards as the first fix. Subtraction beats addition.
Do not use lightness alone to make a colored element "lighter" — rotate the hue.
Do not stack all up-pop properties on every emphasized element. Only the page title gets full up-pop.
Do not use a single flat box-shadow to communicate elevation — decompose into ambient + direct.
Do not let a single screen carry more than one icon library, more than two fonts (one display + one body), or more than one corner-radius scale.
Do not allow off-scale spacing values (13px, 27px) — every value lives on the 4-pixel system.
Do not ship a craft fix that drops body-text contrast below 4.5:1 or strips semantic affordances — tag it Delegated to accessibility_inclusive_ui_review instead.
Keep every fix implementable in the existing design system (Inkprint tokens on BuildOS surfaces); do not invent parallel token scales.
Primary sources: Steve Schoger / Refactoring UI [PRIMARY] (two-part shadows, hue rotation, single-hue palettes), Adam Wathan / Refactoring UI [PRIMARY] (token-ownership build discipline, uppercase letter-spacing, equal-height card flex recipe, negative-margin border overlap), Erik D. Kennedy / Learn UI Design [PRIMARY] (light-from-the-sky, up-pop/down-pop, text-on-image methods, font shortlist), Kole Jain and DesignSpo [practitioner] (supporting numerics).
Enrichment (2026-06-11): added five Erik Kennedy "7 Rules for Gorgeous UI" rules — grayscale-first preflight + Mobbin pre-design imitation test (workflow steps 1–2), four-cue button lighting recipe + brightness-as-elevation (depth-color-surfaces.md), and the hover recolor rule (depth-color-surfaces.md buttons/forms). Source: docs/research/youtube-library/analyses/2026-04-29_erik-kennedy_7-rules-gorgeous-ui_analysis.md.
Enrichment (2026-06-11): added six Adam Wathan / RefactoringUI build rules — uppercase ⇒ tracking-wide, global type defaults (anti-alias + tight/normal leading), and the equal-height card flex recipe (type-spacing-emphasis.md); own-the-config / extend-the-scale-never-a-magic-number and the negative-margin border-overlap fix (depth-color-surfaces.md). Also marked the hover-recolor rule as converging on a second primary source (Wathan, with the live-tune-then-codify process). Source: docs/research/youtube-library/analyses/2026-06-11_adam-wathan_building-refactoringui-with-tailwind_analysis.md.
Maintainers: the canonical research draft with full lineage lives at docs/research/youtube-library/skill-drafts/visual-craft-fundamentals/ (not available at runtime).
