| name | enhance-web-ui |
| description | Artistic, research-grounded UI enhancement skill for making an existing page feel intentional, spacious, and human-crafted. Use when the user asks to make a page nicer, more polished, more premium, more editorial, less crowded, less AI-generated, better laid out or balanced, or mentions typography, visual hierarchy, spacing, empty/dead space, motion, hover states, cards, grids, or density — especially vague-but-visceral feedback ("clunky", "heavy", "incoherent", "wasted space", "feels AI-generated", "monochromatic", "everything stacked left"). Catches hard-to-spot failures: active-state mass mismatch, chrome tautology, card-on-card layering, brand-color competition, full-cell wash where a micro-indicator belongs, wrapper-collapsed tiles, conditional slots reserving space for absent content, and the same datum repeated 3× in one fold. General across web repos. Focuses on content choreography, hierarchy, spacing, and motion; for task-flow usability/data correctness use enhance-web-ux first or alongside it.
|
| license | MIT |
Which enhance skill? (surface router)
| Your surface | Use |
|---|
| Web product page / dashboard — composition, hierarchy, spacing, motion | enhance-web-ui |
| Web product page — UX heuristics, flows, data wiring | enhance-web-ux |
| Web landing / marketing / portfolio (greenfield, anti-slop) | enhance-web-landing |
| Web existing site upgrade (audit-first, preserve behavior) | enhance-web-redesign |
| Web 3D / WebGL / cinematic scroll on an existing site (audit-first) | enhance-web-web3d |
| React Native screen (Expo / bare) | mobile-rn-screen |
| Capacitor / hybrid shell (one web app shipped to iOS + Android) | enhance-capacitor-ui (axis architecture first) → then the web or rn skill |
| Repo README showcase | enhance-readme |
You are here: enhance-web-ui. Native iOS/Android (SwiftUI / Compose, no web layer) is out of scope for all of these — use Apple HIG / Material directly.
Enhance Page UI
Turn an existing screen into a composed interface: calmer hierarchy, smarter
content placement, expressive but purposeful motion, and artful transitions that
support comprehension. This is not a "add more chips/cards/shadows" skill. It
rearranges, subtracts, and stages information before adding decoration.
Vague-but-visceral user feedback ("clunky", "atrocious", "wasted space",
"incoherent height") almost always points to a Hidden Failure Mode from §X
below — start there, not at the surface decoration the user named.
If browser automation is used, first follow the
protocol-browser-anti-stall skill when present.
Critical Rules
Compose before decorating. Fix hierarchy, grouping, alignment, and rhythm
before adding gradients, motion, blur, masks, or shadows.
Replace clutter with structure. Do not stack more badges, labels, cards,
dividers, or helper text onto a crowded surface. Move secondary information
to a quieter region, a footer band, a hover/focus reveal, a drawer, or a
progressive disclosure layer.
Make hierarchy visible at a squint. A user should know the focal element,
supporting content, metadata, and actions before reading every word.
Motion must explain. Animate only feedback, state change, spatial
relationship, affordance, or attention. No looping sparkle unless it carries
status or brand tone and respects reduced motion.
Use the local design system. Reuse existing primitives, tokens, radius,
spacing, typography, animation helpers, and dark/light surfaces. Extend a local
primitive before inventing a one-off.
Never ship mock, dead, or rogue UI. Enhancements must be wired to the
product's real data contracts, backend/API state, database schema, or existing
domain helpers. If the UI needs better visual feedback, first look for the real
status, timestamp, count, owner, error, progress, permission, or lifecycle data
that should drive it. If that data does not exist but is necessary for a solid
user experience, propose or implement the smallest backend/schema/API extension
instead of faking the state in the component.
Respect density by viewport. Desktop can hold layered metadata. Tablet
needs tighter grouping. Narrow screens need progressive disclosure and no
wrapping CTAs.
Tune saturation to product domain. A consumer learning, kids, fitness,
or gamified app needs a vibrant, semantic palette where each colour
carries meaning (Duolingo: green=correct, red=hearts, orange=streak). A
B2B/SaaS/finance dashboard needs restrained colour reserved for status
only (NN/g 60–30–10, Stripe-style). Faint /5–/10 tints on a playful
consumer surface read as monochromatic and dead. Pick a saturation tier
for the product first, then apply it consistently — see Domain Colour
Tier below.
Audit conditional slots for dead-state. Any optional region — media,
aside, eyebrow, footerSlot, cover, illustration, secondaryCTA
— must be tested with sparse / zero-state content. A slot that reserves
min-h: 160px for a rich illustration becomes a giant empty card when
the only content is "0%". Make the slot conditional on content (drop
the wrapper when empty) or collapse the floor (min-h-0 on mobile),
never reserve space for absent content.
Scan for information duplication within one viewport. The same number,
word, date, or status appearing twice in the same fold means one of them
is wasted ink. Eyebrow shows "0/10 today" + inline pill shows "0/10
today" + footer bar shows "0/10 today" → keep the most actionable copy
and delete the rest. Duplicates fragment attention (NN/g #8) and signal
that the content rank in step 2 was skipped.
Patch the primitive, not the site. If a layout/colour/wrap bug
repeats, fix it where the wrapper, helper, or token lives — not at the
consumer. A single Tooltip defaulting to inline-flex will silently
shrink every w-full child across the app; one prop fix beats N
per-page band-aids. See Primitive-First Patch Rule below.
Calm chrome before you decorate content. Wayfinding chrome (header,
breadcrumb, tabs, dock, toolbar) should recede, not announce itself —
Linear's 2025 refresh names this "structure should be felt, not seen". If
chrome competes with content for visual weight, it has too much paint:
strip backgrounds → borders → tints → fills, in that order, until the
chrome reads as a tonal recess that frames the content. The squint test
applies to chrome too: at 10 px blur, the page content should dominate;
the chrome should look like a quiet frame, not a stack of buttons.
One brand-color element per visual zone. A page should have ONE
brand/accent-tinted element per zone (header strip, breadcrumb row, tab
row, dock, page body). If breadcrumb chips, tab badges, AND the actual
primary CTA all use the brand color, the actual CTA loses scent — every
brand-tinted thing claims primacy and the user's eye can't pick a target.
Audit the brand budget per zone: count brand-tinted surfaces in the
screenshot; if > 1 per zone (excluding the count badge / underline / status
chip that carries meaning), demote the chrome ones to neutral
typography. The brand color is a finite resource — spend it on the action
the user is here to take.
Active ≠ heavier layout — active = different cue. When an element
moves to active state, its layout dimensions must not change (no
taller, no wider, no extra padding). The active signal should come from
a micro-cue: text-color shift + font-weight bump (font-medium →
font-semibold), an underline, an icon-wrapping pill (Material 3
NavigationBar pattern: 32×16 bg-secondaryContainer rounded-full pill
wrapping the icon, NOT the whole cell), or a count-badge tint swap. Full
background washes on active tabs/buttons/cells inflate perceived weight
50%+ even when the bounding box is identical; the user reads the active
sibling as "weirdly big and clunky". Linear, Stripe Apps, Vercel, M3 all
obey this — copy them.
Workflow Checklist
Copy and track:
UI ENHANCE /<route-or-component>
- [ ] 0. DOMAIN: product class -> colour-tier + density-tier (see Domain Colour Tier)
- [ ] 1. RECON: route, component tree, data fields, tokens, primitives, wrappers
- [ ] 2. CONTENT RANK: primary, secondary, metadata, actions, ambient
- [ ] 3. LIVE READ: 1440 / 1024 / 800 screenshots + DOM rect measurements
- [ ] 4. PAIN MAP: crowding, dead space, weak hierarchy, dup info, wrapper collapse,
conditional-slot zero-state, monochromatic surfaces, hard cuts, bland motion
- [ ] 5. COMPOSITION PLAN: move, group, pin, crop, reveal, fade, animate
- [ ] 6. IMPLEMENT: primitive-level fixes first, smallest site-level diffs after
- [ ] 7. VERIFY: three viewports + DOM uniformity + squint-for-colour-distinction
+ reduced motion + dark/light if supported
- [ ] 8. WRITE-UP: before pain -> design principle -> after behavior
Step 1 - Recon: Learn the Existing Visual System
Read the route/component in full. Then inspect:
- Layout primitives:
Card, Section, Drawer, Tabs, Badge, Button,
Tooltip, grid/list helpers.
- Styling system: Tailwind config, CSS variables, design tokens, theme context,
animation utilities, existing radii/shadows.
- Data shape: all fields available, especially metadata currently shoved into
titles or hidden in modals.
- Data source: API endpoints, loaders/actions, database schema, cache keys,
backend status enums, validation errors, permissions, and lifecycle events that
can make the UI's feedback real instead of decorative.
- Existing comments: search for "DO NOT", "avoid", "deprecated", "design",
"layout", "spacing", "motion", "overflow".
Output a tiny component map:
<Page>
|-- Header - title, context, primary action
|-- Body/Grid/List - main content
|-- Detail/Drawer - progressive disclosure
`-- Footer/Chrome - metadata, controls, scroll status
Also list every wrapper / layout primitive in the chain from grid cell
down to the rendered child: Tooltip, Trigger, Popover, Slot,
AnimatedDiv, Tippy, motion.div, etc. Note the default display,
width, and flex-shrink of each. This is where wrapper-collapse bugs
hide and where the primitive-first patch in step 6 is decided.
Step 1.5 - Domain Colour & Density Tier
Before touching any colour, name the product class. The same /8 tint
that reads "premium and restrained" on a finance dashboard reads
"monochromatic and dead" on a learning app. Pick a tier, then apply it
consistently — never blend tiers in one product.
| Tier | Examples | Surface bg tint | Accent fill | Categorical hue use | Source |
|---|
| A — Vibrant / Gamified | Duolingo, Drops, Habitica, Strava, kids/learning, fitness, casual gaming | /15–/25 gradient + ring /30–/40 | solid token (bg-success, bg-cta), white-on-solid text | every meaningful category gets a distinct hue (success/error/streak/xp/info/premium) — colour is the gamification | Duolingo design system; Figma 2026 trend report |
| B — Expressive Consumer | Spotify, Notion playful surfaces, Robinhood onboarding, content/lifestyle apps | /10–/18 gradient + ring /20–/28 | semi-solid (bg-cta/85) | 3–5 semantic hues, decorative palette restrained | Spotify, Notion |
| C — Productive / Pro Tools | Linear, GitHub, Vercel, Figma chrome, design tools | /6–/12 tint + ring /15 | one accent + 4 semantic states | accent reserved for primary action; semantic states only for status | Linear, Radix |
| D — Restrained / Data-Dense | Stripe dashboard, AWS console, observability/finance/admin/healthcare | mostly neutral surfaces; /4–/8 only on status pills | one accent for CTA; semantic states 4-slot (success/warn/error/info) | colour reserved for data signal; chrome stays achromatic — NN/g 60-30-10 | Stripe, ColorArchive SaaS dashboards |
How to detect tier (do this before designing):
- Read the product's marketing copy, README, or onboarding — words like
play, streak, level, fun, journey, daily, habit, master → tier A or B;
workspace, dashboard, report, audit, query, pipeline → tier C or D.
- Look at the existing primary CTA: solid saturated colour with chunky
shadow → A; solid muted → B/C; outline or ghost dominates → D.
- Look at semantic state tokens (
success, warning, error,
info, plus app-specific xp, streak, premium, ai, tone-1..6).
More than 5 semantic hues with brand meaning → tier A. Just the basic 4
→ tier C/D.
Tier mismatch is the top hidden cause of "monochromatic" feedback on
consumer apps. A tier-A product with tier-D /8 tints will look
washed-out and AI-generated even though every token is technically
"correct".
Step 2 - Content Rank Before Layout
Classify every visible element:
| Rank | Meaning | Default Treatment |
|---|
| Primary | What the page is about | largest scale, strongest contrast, best space |
| Secondary | Helps interpret primary | near primary, smaller, grouped |
| Metadata | Dates, counts, stats, status | pinned footer/rail/table column, not title stack |
| Actions | What user can do | close to target, stable hit area, no wrap |
| Ambient | Brand, mood, decorative signal | background, mask, low contrast, removable |
Rules:
- If everything is primary, nothing is primary.
- If metadata crowds the headline, move it to a bottom band, side rail, table
column, tooltip, or detail drawer.
- If an action wraps, abbreviate progressively or icon+tooltip it.
- If a grid has dead space, adjust span, dense placement, aspect ratio, or
content anchoring before adding more content.
Step 3 - Live Read
Inspect the page at:
1440 x 900 desktop composition
1024 x 700 tablet / split-screen stress
800 x 700 narrow desktop / large mobile stress
For each viewport, note:
- First focal point.
- What feels crowded.
- What feels empty or unbalanced.
- Which labels wrap or truncate.
- Whether scroll edges are hard cuts or softly communicated.
- Whether hover/focus states explain interactivity.
- Console/runtime errors.
Use the squint test: mentally blur the screenshot. The page should still reveal
its main regions through scale, contrast, grouping, and whitespace.
3a - Two squint passes, not one
Run the squint test twice — both passes are necessary, and the second is
the one most enhancers skip:
| Pass | Question | Failure looks like |
|---|
| Hierarchy squint | "Where is my eye pulled first?" | Several elements equally loud → primary not chosen |
| Category squint | "Can I tell related items apart by colour/shape alone, without reading?" | All tiles same neutral grey → user must read every label = NN/g #6 violation |
For a Tier A/B product (see Domain Colour Tier), a 2-up or 3-up of
metric tiles that all read the same hue when blurred is a bug — colour is
your category signal. For a Tier D product, the same blur should show
near-uniform tiles with colour reserved only for the alert row.
3b - DOM uniformity gate (catches silent wrapper collapse)
Screenshots are necessary but not sufficient. When a page has repeated
elements that should be the same width / height — grid tiles, segmented
control segments, table columns, sidebar nav rows, form fields — measure
them. Many "weird-looking" layouts come from a wrapper higher in the tree
silently collapsing children to content width.
For each repeated group, query rendered rects and assert uniformity:
const rects = [...document.querySelectorAll('<your-tile-selector>')]
.map(el => el.getBoundingClientRect());
const widths = rects.map(r => Math.round(r.width));
const heights = rects.map(r => Math.round(r.height));
const span = Math.max(...widths) - Math.min(...widths);
console.table({ widths, heights, widthSpan: span });
Common wrapper-collapse causes (generic, language/framework agnostic):
- Tooltip / Popover / Trigger primitive defaults to
display: inline-flex
or inline-block → child width: 100% collapses to content width.
- Grid template uses
1fr (which is minmax(auto, 1fr)) and a child has
intrinsic content wider than expected → cell expands beyond 1fr. Fix:
repeat(N, minmax(0, 1fr)).
- Flex child without
min-width: 0 or flex: 1 1 0 → content size
dictates layout instead of the flex track.
- Animation libraries (Framer Motion
m.div, Reanimated, GSAP wrappers)
that inject inline display: inline-block for transform performance.
- Slot patterns (
<Slot>, asChild, cloneElement) where the consumer
drops w-full because the inner component doesn't forward className.
Fix at the wrapper, not the leaf: see Primitive-First Patch Rule.
3c - Conditional-slot zero-state pass
For every conditional region (media, aside, cover, eyebrow,
secondaryCTA, illustration, footerSlot), inspect its sparse
state — render the page with the slot's smallest realistic content
(empty, "0%", a single icon, one short word). If the slot still reserves
its min-h / aspect-ratio / min-w floor, you have a dead-conditional
slot. Move to Composition Move 4 (Stage) and either:
- Drop the wrapper when the slot is empty (
{slot && <Region>{slot}</Region>}).
- Collapse the floor on narrow viewports (
min-h-0 lg:min-h-[160px]).
- Inline the data into the body (chip / pill next to the action).
3d - Information-duplication scan
Within every fold (one viewport-height slice top-to-bottom), search for
the same number, word, date, or status appearing more than once. List
duplicates explicitly:
Fold 1 of /home @ 390×844:
- "0/10 words today" ×3 (eyebrow metric, action pill, footer strip)
- "0%" ×2 (action pill, Today metric tile)
- streak count ×2 (eyebrow metric, Streak tile)
The duplication itself is the diagnosis — pick one canonical home for
each datum (usually the most actionable one) and delete the others. Each
delete is a hierarchy upgrade for whatever stays.
Step 4 - Composition Moves
Prefer these moves, in order:
- Subtract: remove repeated hints, duplicate labels, low-value icons, noisy
chrome, and "just in case" metadata.
- Group: use proximity and common region to make related pieces read as one
unit. Increase spacing between groups; decrease spacing inside groups.
- Pin: move stats, code metadata, timestamps, or progress to a card footer,
side rail, bottom band, sticky summary, or column.
- Stage: reveal secondary content through hover, focus, tabs, accordions,
drawers, details panels, or "show more" only when it helps.
- Rebalance: change card spans, grid density, aspect ratio, order, or
alignment so important items get room and empty space feels intentional.
- Soften: replace hard clipping with fades, masks, scroll shadows, curtains,
sticky gradients, or viewport-aware edge treatments.
- Animate: add microinteractions only after the static composition works.
Hidden Failure Modes (the difficult-to-spot ones)
These are the "user said it but you missed it on first pass" patterns. When
feedback is vague-but-visceral ("clunky", "atrocious", "wasted space",
"incoherent height", "feels AI-generated"), walk this list before you touch
decoration. Each entry has a detection probe (what to look for) and a
fix shape (the smallest move that resolves it). Generic across stacks.
H1. Active-state mass mismatch
Symptom: "Tab/button/dock cell looks weirdly big and clunky", "highlight
area is huge", "heights are incoherent".
Detection: Take a screenshot. Compare bounding boxes of active vs inactive
siblings (DevTools → element rect, or just eyeball). If layout dimensions
match but the active one feels 1.5× heavier, the active state is using a
full-cell background fill. Grep your active state for bg-* on the
container element. If found AND the inactive sibling has bg-transparent,
that's the bug.
Fix shape: Move the active signal to a micro-indicator wrapped inside
the cell, not painted over it:
- Tab → underline + text-color + count-badge swap (Stripe / Vercel pattern).
- Bottom dock → 32×16
rounded-full pill wrapping the icon only, not
the whole button (Material 3 SecondaryContainer pattern).
- Sidebar item → leading 2px brand rail + bg-muted hover only.
- Button → solid fill is fine for primary CTAs (they SHOULD be heavy);
forbidden for navigation siblings (they should NOT compete).
H2. Chrome tautology on root / index routes
Symptom: "Breadcrumb is atrocious", "header is too much", "why is 'Home'
shown 3 times".
Detection: On the root route (/, /dashboard, /home), count how many
chrome zones display the page name or home affordance: company-as-home link,
breadcrumb chip, page title H1, dock active-item label, tab name. If ≥ 2
zones say the same word for the current page, you have tautology.
Fix shape: On root routes, suppress redundant chrome — pass
showContextRow={false} (or your equivalent) to the page-layout wrapper, or
gate the breadcrumb on pathname !== "/". The dock indicator + page H1 +
greeting copy already answer "where am I?". The breadcrumb's job is "how do
I go back?" — at the root, there is no back, so the row is dead weight.
H3. Card-on-card chrome
Symptom: Active page chip "feels heavy", "looks like a button stuck on
another button", a chip on a tonally-recessed row reads as one elevation too
many.
Detection: Inspect the row's background (e.g. bg-muted, bg-chrome-subtle,
or any tonal recess). Then inspect the chip inside that row (e.g. bg-card border). Two distinct elevations within ~6 px of each other = card-on-card.
Fix shape: The chip must adopt the row's elevation, not introduce its own.
Either drop the chip's bg-* + border and let typography (font-semibold
text-foreground) carry the active signal, OR drop the row's tonal recess
and let the chip be the only painted surface. Pick one elevation per row.
H4. Brand-color competition
Symptom: "Page feels noisy even though it's clean", "I can't tell which
button is the primary action", "everything looks important".
Detection: Take a desktop screenshot. Count surfaces tinted in the brand
color (border, fill, ring, text) within one visual zone. Subtract: the
single intended primary CTA + status chips that carry meaning. Anything
left over (a brand-tinted breadcrumb chip, a brand-tinted tab background, a
brand-tinted home icon border) is competition.
Fix shape: Demote the chrome surfaces to neutral (text-muted-foreground,
bg-transparent, border-border/40). Reserve the brand color for: the
single primary CTA per fold, real status indicators (success / progress /
new), and the active-item micro-indicator (not the cell background).
Further reading