antislop-design
Load before writing, reviewing or fixing any web UI code (HTML, CSS, Tailwind, React) or when auditing accessibility and visual quality. A single deterministic contract: a ban filter for AI-slop defaults (purple gradients, Inter everywhere, 3-card grids, glassmorphism, emoji icons, invented proof), a positive pattern library (layouts, typography, components, charts, web dashboards, award-tier craft), and a CSV design-system generator whose output is post-filtered before use. Not for backend, API, data pipelines, docs-only work, native mobile UI, or data-viz in notebooks and BI tools.
How do I install this agent skill?
npx skills add https://github.com/vanszs/antislopui --skill antislop-designIs this agent skill safe to install?
- Gen Agent Trust Hubpass
AntiSlopUI is a professional design system skill for AI agents that enforces high-quality UI standards while proactively preventing common security flaws in generated code, such as credential exposure and XSS. It includes well-structured Python tools for research and validation that operate entirely within the skill's own directory.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
AntiSlopUI — one skill, one contract
Produce UI that a discerning designer cannot instantly flag as machine-generated.
Everything this skill needs ships in this repository: the ban filter is this file,
the pattern library and the catalogues live in references/, and the research tool
lives in ui-ux-pro-max/.
If you could swap the logo with a competitor and nobody would notice — it's slop.
Delivery gate (read now, apply before shipping)
Before delivery, report PASS, FAIL, N-A, or UNVERIFIED plus an evidence
pointer for every audit item below. N-A is allowed only when the component — or
the capability it audits — does not exist on the page. UNVERIFIED means you could
not measure it; never mark an unmeasured item PASS, and an UNVERIFIED P0 blocks delivery exactly like a
FAIL until you verify it.
Any P0 failure blocks delivery, as does 3+ P0/HIGH tells. MEDIUM and LOW tells are
recorded, not blocking. WARN-BRAND-SWAP is a warning, never a block.
Deterministic terms
A tell is one catalogue row in references/slop-uiux-patterns.md sections 1–10, or one
of the code-level findings the audit lists below (five a11y/form HIGH items and seven code P0
items, the HIGH set being focus states, form labels, alt text, colour-only indicators and
validation states (F1 in the code catalogue)). Severity is exactly P0, HIGH, MEDIUM or LOW; the vocabulary is closed.
Count per element: when one element matches several rows (a 3-card grid with emoji icons
is both K1 and I1), count it once at the higher severity. Rows that describe the whole
page count once for the page; today that is L1, L2, L3, L4, L5, L7, L8, L10, L11, M16, M17,
M20, MO1, MO3, MO13, C10, FO1, and any other row whose tell is a property of the page rather than
of one element. N4 and H8 are
JUDGEMENT — record them, never count them. Absence rows do not match unless the page
claims the capability; otherwise they are N-A.
Banned purple, indigo or violet counts as "real brand" only when the value appears in
user-supplied brand assets — a logo file, existing CSS tokens, or a brand PDF. Otherwise
it is a violation. Record exceptions exactly as OVERRIDE: <from> → <to> [<reason>],
where <reason> is brand: <source>, craft tier §5, domain table, display table,
catalog <ID>, or prose. A gate id is not a reason: gates are prohibitions, not
justifications.
Never invent a product name, customer name, metric, logo or testimonial. If the user
supplied a name, use it exactly; otherwise write Your product in the wordmark, label
it once on the page, and record NO-BRAND: user supplied no name — a page that records NO-BRAND is N-A on
the logo-swap item. Unknown proof is omitted and recorded NO-PROOF: user supplied none;
invention is P0. Placeholder content — including representative numbers — is allowed only when visibly
labelled and structurally distinct (<figcaption>Sample data — replace before launch</figcaption>
or data-sample="true"); [placeholder], lorem ipsum and TODO stay banned. A labelled
sample is not an invented metric.
Hard gates
| ID | Banned | Compliant default | Severity |
|---|---|---|---|
| H2 / T2 | Gradient-text headline (one defect, two catalogue rows) | Solid, high-contrast headline | P0 |
| C1 | Purple/indigo/violet primary accent without a supplied brand asset | A non-purple brand hue from the domain table | P0 |
| C2 | Canonical purple/indigo or blue/cyan gradient combinations | Flat colour, or one intentional compliant gradient per page | P0 |
| C7 | Low-contrast body text | 4.5:1 body, 3:1 large/UI, checked | P0 |
| K6 | Purple left-border card accent | Remove; left borders only for semantic alerts | P0 |
| MO10 | Missing prefers-reduced-motion support | Gate animation; render static under reduced motion | P0 |
| T1 | Inter/Geist for display and body without a second face | Distinctive display face + body | HIGH |
| K1 | Identical 3-col feature grid (icon-in-rounded-square + title + blurb) | Vary size/structure, or use a list/table | HIGH |
| K3 | Any scaling hover (hover:scale-*) | Shadow/border/colour feedback | HIGH |
| I1 | Emoji as functional icons | One SVG icon set | HIGH |
| B3 | Pill buttons everywhere | rounded-md/rounded-lg; pills only when semantic | HIGH |
| N1 | Floating glassmorphic navbar | Fixed solid navbar with a border | HIGH |
Also enforced from the catalog: mono/technical faces as display (§3), indigo-950 or
violet-50 tints as text or background (no dedicated row: treat them as C1 for accents and
C7 for text contrast, and record the substitution from references/research.md), typewriter-style copy animation (MO12) and
multi-layer parallax (MO8), both in §8 MOTION. These are catalog rows, not a
second list.
For the full catalogue: cd <skill-dir> && grep -nE '^## |^\| ?[A-Z]{1,3}[0-9]+' references/slop-uiux-patterns.md.
The five axes
| Axis | Banned default | Required action |
|---|---|---|
| Colour | Purple/indigo accent, purple or blue-cyan default gradients | Domain hue from the table below; document overrides |
| Type | Inter everywhere, gradient headlines, Playfair + Inter | Category display face; solid headline; 2–3 weights |
| Layout | Centred sameness, repetitive 3-col grids, one container width | Vary structure and emphasis by content |
| Copy | Hype, generic CTAs, invented proof | Specific copy; omit unknown proof, record NO-PROOF |
| Motion | Gratuitous motion, missing reduced-motion | Follow the motion table below |
Motion table
| Motion class | Duration / behaviour |
|---|---|
| Hover | 100–150ms |
| Entrances and exits | 150–250ms |
| Page/route transitions | ≤300ms |
| Craft-tier exception | ≤400ms ONLY with OVERRIDE: … [craft tier §5] |
| Ambient | Maximum 1, pause on hover/focus, static under prefers-reduced-motion |
Hover examples use 150ms; entrance and exit examples use 200ms. Outside this table, point here instead of restating durations.
Text measure
max-w-3xl for prose containers, max-w-prose for paragraphs, max-w-5xl for wide
single-column blocks, max-w-6xl for grids. A column in a two-column hero split may use
max-w-2xl; the 65ch paragraph rule applies to paragraphs inside it, not to the column.
A grid section may repeat max-w-6xl; what is banned is every section sharing one
container, which flattens the page into a single column.
Palette and type: who decides
Colour is decided in this order, and the first match wins:
- User-supplied brand assets — logo file, existing CSS tokens, brand PDF. Use the
value found there and record
[brand: <source>]. - This table by product domain — every value is contrast-certified for white text:
| Product domain | Brand hue |
|---|---|
| Money, invoicing, ledger, productivity | #0F766E deep teal |
| Health, clinical, wellness, care | #0E7490 cyan-teal |
| Safety, urgency, logistics, alerts | #B45309 clay |
| Agriculture, environment, sustainability | #3F6212 moss |
| Enterprise, legal, infrastructure, docs | #334155 slate |
| Food, hospitality, craft, heritage | #7F1D1D oxblood |
| Developer tooling, data, AI infrastructure | #4D7C0F olive |
| Anything not listed | #0F766E deep teal |
If several rows match, the row earliest in the table wins. Do not derive a hue by feel.
3. Research output — --design-system colours after the post-filter. A proposal,
never an authority: step 2 already decided the hue, so research supplies systems and
accents only.
references/ supplies structure, spacing, components and colour systems (tokens,
chart palettes, OKLCH recipes). It does not choose your brand hue.
Display face is chosen by product category, not taste:
| Product category | Display face | Body |
|---|---|---|
| Editorial, luxury, culture, portfolio, beauty | Fraunces | Inter or the palette's body face |
| Software, developer tool, data product | Bricolage Grotesque | Inter or the palette's body face |
| Neutral B2B, enterprise, docs | Bricolage Grotesque | Inter or the palette's body face |
| Playful, consumer, education, gaming, e-commerce | Cabinet Grotesk | Inter or the palette's body face |
Two families maximum. Default body is Inter 400/500; any single non-display sans is legal (the generator's body face counts), and Inter is legal only when a different display face is present. If the category is genuinely ambiguous, use Bricolage Grotesque and say why. If the display face is not loadable, use Bricolage Grotesque rather than a vendor face.
Contrast
Compute relative luminance, or use these verified pairs: #0F172A on #F8FAFC = 17.1:1 ·
#475569 on #F8FAFC = 7.2:1 · #FFFFFF on #0F766E = 5.5:1 · #FFFFFF on #B45309 = 5.0:1.
Ship body text at 4.5:1 or better, large text and UI at 3:1 or better. If you cannot
compute a custom pair, move the value until it matches one of these.
Authored defaults (when research output is unavailable)
- Surfaces:
#F8FAFCpage,#FFFFFFcard, border#E2E8F0, text#0F172A, muted#475569. - Type: one display face ≠ Inter, body Inter 400/500, two families, three weights, 1.25 scale.
- Layout: prose
max-w-3xl, gridsmax-w-6xl, section paddingpy-32/py-16/py-12. - Radius scale: badges and avatars
rounded-full(orrounded-smfor rectangular badges), inputs and buttonsrounded-md, cards and tablesrounded-lg, modals and drawersrounded-xl, sections and bannersrounded-noneorrounded-lgto match their card, accordion rows inherit the card radius, tooltipsrounded-md. No other values. - Interaction:
cursor-pointeron anything clickable, visiblefocus-visiblering, hover feedback by colour or shadow only. - Responsive: verify 375 / 768 / 1024 / 1440 with no horizontal scroll.
Which reference to open
| You need | Open |
|---|---|
| Tell catalogues and code-level tells | references/slop-uiux-patterns.md, references/slop-code-patterns.md |
| Page/app skeleton, grid, responsive structure | references/layouts.md |
| Type scale, pairings, fluid type | references/typography.md |
| Cards, buttons, forms, nav, data display | references/components.md |
| Overlays, feedback, motion recipes | references/animation.md |
| Library choice, design tokens | references/design-systems.md |
| Charts, KPI cards, data visualisation | references/charts.md |
| Admin dashboard shell, tables, filters | references/dashboards.md |
| Award-grade craft tier (opt-in) | references/awards-craft.md |
| Industry evidence from the CSV tool | references/research.md |
Anti-monotony rules: never repeat the same section anatomy consecutively; vary container
widths by content; vary section padding (hero py-32, features py-16, CTA py-12);
one dominant element per section; break the rhythm every 2–3 sections; mix grid
structures; at least two deliberate asymmetric moments per page.
Reference budget
Per task: this file plus sections, never whole reference files. A whole reference is 6–36 KB (2–10k tokens). Only the six smallest would fit beside this file and the catalogue; the four largest would not. None may be read whole. A section runs from a few hundred tokens to about 1,300, with the dashboard and code catalogues reaching roughly 2,100.
Ceiling: 70 KB of skill text per task (≈20k real tokens; 3.5 bytes ≈ 1 token, so bytes
÷ 3.5 is the token estimate). Self-check with a shell: wc -c what you have loaded, add
it up, and compare with 70,000. To measure a section rather than a file:
s=$(grep -nE '^#{1,2} ' references/layouts.md | grep -i 'HERO VARIANTS' | cut -d: -f1)
e=$(grep -nE '^#{1,2} ' references/layouts.md | awk -F: -v s="$s" '$1>s{print $1; exit}')
sed -n "${s},$((e-1))p" references/layouts.md | wc -c
Without a shell, use the structural rule: this file plus the tell catalogue plus at most two sections from one library file.
When a number appears here and in a reference, this file wins; reference values are
examples at their own scale. The ## Contents block in a reference is a summary of its sections — grep returns the
same names with line numbers, so grep is the step that saves round-trips:
cd <skill-dir> && grep -nE '^#{1,2} ' references/layouts.md # names + line numbers
sed -n 'START,ENDp' references/layouts.md # END = next match − 1
Some references carry their own checklists (references/dashboards.md 15 principles,
references/charts.md a 15-point scan, references/research.md a post-filter). Run them as
guidance for that artefact; they never add blocking items and never change the tell set.
Where a reference checklist disagrees with this audit, this audit wins.
Craft tier (opt-in)
references/awards-craft.md is the only file that may raise the craft ceiling, and only
when the user explicitly asks for premium, award-grade, "wow", studio or launch work. It
carries a scored rubric (Design 40 · Usability 30 · Creativity 20 · Content 10), hard
performance and accessibility gates that deny the tier on failure, and a cost table for
every craft pattern. It never buys a ban. Its two relaxations are declared OVERRIDE: … [craft tier §5]: a 400ms hero choreography, and one viewport-wide ambient animation. Do not load it for dashboards, docs, checkout or forms.
Research tool
references/research.md documents the CSV generator (commands, domains, stacks, persist
pattern, post-filter, determinism, no-shell path). The script path is
ui-ux-pro-max/scripts/search.py relative to this skill; it resolves its data directory
relative to itself, so an absolute or <skill-dir>-prefixed path works from any cwd.
If the shell is unavailable, do not read the CSVs by hand: build from the authored defaults above, and state that the research step was unavailable.
Pre-Delivery Audit
Report PASS / FAIL / N-A / UNVERIFIED with an evidence pointer per item.
Colour and type
- No banned purple/indigo without a supplied brand asset; exceptions recorded as
OVERRIDE:with source. - No gradient text; at most one intentional gradient element per page.
- Display face is not Inter/Geist; ≤3 weights; modular scale.
- Body text passes 4.5:1 (and 3:1 for large/UI). If the user supplies a dark palette, verify both; otherwise ship light and state that choice.
Layout and components
- Container widths and section padding vary by content; at least one full-bleed break.
- No identical 3-col grid repeated; one dominant element per section.
- Radius is a scale, not one value.
- No scaling hover (
hover:scale-*); no glass except ONE surface per page over imagery atbg-white/80+ with@supports; no emoji icons.
Copy
- No hype, no generic CTAs, no
[placeholder]/ lorem / TODO. - Every metric is real or visibly labelled; unknown proof omitted with
NO-PROOF. - Sample content visibly labelled.
Code-level findings (7 P0)
- (P0) Semantic landmarks: one
<main>,<nav>,<header>,<footer>; page haslang. - (P0) Real controls:
<button>/<a>, never<div onClick>or<span role="button">. - (P0) Overlays are
<dialog>(or a focus-trapped equivalent) and close on Escape. - (P0) Custom inputs follow the ARIA pattern: combobox/select exposes
role="listbox"+aria-expandedand is keyboard operable. - (P0) Icon-only controls carry
aria-label. - (P0) No secrets, keys or tokens in client code, and no unsanitized HTML injection
(
dangerouslySetInnerHTML,innerHTML,v-html). - (P0) Data fetching follows the stack's idiom — server component, loader, or query cache — not a client-side request waterfall in an effect.
Motion and accessibility findings (5 HIGH)
- (HIGH) Images carry
alt. - (HIGH) Focus states are visible.
- (HIGH) Form controls carry labels.
- (HIGH) Forms expose validation states.
- (HIGH) Colour is never the only state indicator.
Cross-cutting checks (counted through the catalogue rows named)
- Motion inside the canonical table; colour/shadow hovers are 100–150ms; any spatial motion
uses
transform/opacityonly; reduced-motion renders static (MO10, MO11). - Body text and UI meet the contrast thresholds above (C7).
Gestalt
- Counted catalog tells per element; any P0 tell or 3+ P0/HIGH tells = block.
- Could a competitor swap the logo without notice? If yes, record
WARN-BRAND-SWAPand fix it if a name exists — never a blocking FAIL.
References
references/slop-uiux-patterns.md— 126 tell rows, TELL → WHY-SLOP → FIX → SEVERITY.references/slop-code-patterns.md— React/TSX/a11y code-level tells.references/layouts.md,typography.md,components.md,animation.md,design-systems.md,charts.md,dashboards.md— the positive pattern library.references/awards-craft.md— opt-in craft tier with hard gates.references/research.md— CSV generator workflow, post-filter, determinism.
Output expectations
The result must read as intentional, branded, content-aware, accessible and production-ready. Every spacing value, colour, font, grid and animation must be defensible with a content-specific reason. If you cannot explain why a choice was made, it is slop — remove or replace it.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/vanszs/antislopui/antislop-design">View antislop-design on skillZs</a>