Checklist
A website design checklist you can test.
242 rules for a homepage, each one a check a person or a tool can run: accessibility, usability, visual craft, copy, components and what keeps a page from looking generated. Every rule says how bad it is to break, how to check it, and where it comes from.
By Max Zeshut · Rules last checked against their sources on 2026-11-05
| Group | blocker | major | minor | advisory | Rules |
|---|---|---|---|---|---|
| Accessibility | 32 | 18 | 1 | – | 51 |
| Usability | – | 19 | 10 | 1 | 30 |
| Visual craft | – | 9 | 22 | 10 | 41 |
| Copy | – | 12 | 19 | 5 | 36 |
| Components | 3 | 27 | 21 | 8 | 59 |
| Distinctiveness | – | 1 | 17 | 7 | 25 |
| All | 35 | 86 | 90 | 31 | 242 |
- blocker: stops someone from using the page.
- major: makes it clearly worse for many people.
- minor: a flaw a careful reviewer would fix.
- advisory: good practice, not a fault.
Check a page with your agent.
The What a design! MCP server runs these rules on a live page and reports what fails, with the evidence: the review_design tool. See the MCP server.
Accessibility.
WCAG 2.2 at level AA, written as checks a person or a tool can run. 51 rules.
perceivable
blocker · a tool finds it, a person confirms · #a11y-alt-text
Every informative image has a text alternative that serves the same purpose; purely decorative images have an empty alt so they are skipped. (WCAG 1.1.1, A)
Why and how to check
People who cannot see an image still need what it conveys. Missing alt text appears on 53.1% of analyzed home pages.
Check: Automated check finds missing alt (axe `image-alt`); a person judges whether the text actually matches the purpose.
Sources: How to Meet WCAG 2.2 (Quick Reference); The WebAIM Million (2026 report)
blocker · a person checks it · #a11y-captions
Prerecorded video with sound has captions. (WCAG 1.2.2, A)
Why and how to check
Deaf and hard-of-hearing people, and people watching without sound.
Check: Play every video with sound off; check captions exist and are accurate.
blocker · a person checks it · #a11y-audio-description
Prerecorded video has audio description or a text alternative covering the visual content. (WCAG 1.2.3 A, 1.2.5 AA)
Why and how to check
Blind and low-vision people miss information that is only shown.
Check: Check whether key visual information is described by narration, a description track or a transcript. WCAG 1.2.1 requires an alternative or audio track for prerecorded video-only content and its Understanding page does not exempt decorative or background video; 1.1.1 says pure decoration should be ignorable by assistive technology. Treating a purely decorative loop as decoration is an inference from 1.1.1.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.2.1 Audio-only and Video-only (Prerecorded); Understanding SC 1.1.1 Non-text Content
blocker · a tool finds it, a person confirms · #a11y-semantic-structure
Structure that is visible (headings, lists, tables, form labels, regions) is also in the markup, not only in the styling. (WCAG 1.3.1, A)
Why and how to check
Assistive technology navigates by structure.
Check: Automated checks for headings, lists, labels and landmarks; read the page outline in a screen reader or the accessibility tree.
blocker · a person checks it · #a11y-meaningful-sequence
The reading order in the code matches the meaning of the page, even if CSS moves things around visually. (WCAG 1.3.2, A)
Why and how to check
Screen reader and keyboard users follow the code order.
Check: Compare DOM order with visual order. Failure F1 is positioning that changes meaning; technique C27 is making DOM order match visual order. MDN: CSS `order` changes only the visual order, not the logical or tab order, which creates a mismatch for screen reader and keyboard users. A check with CSS disabled is a practical test (not described on WCAG's page).
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.3.2 Meaningful Sequence; CSS order (accessibility concerns)
major · a person checks it · #a11y-no-sensory-only
Instructions do not rely only on shape, color, size, position or sound ("click the green button on the right"). (WCAG 1.3.3, A)
Why and how to check
People cannot always perceive those characteristics.
Check: Read all instructions and labels; each must work without seeing the layout.
major · a person checks it · #a11y-orientation
Content works in both portrait and landscape unless one orientation is essential. (WCAG 1.3.4, AA)
Why and how to check
Some people mount their device in one orientation.
Check: Rotate a device or emulator and check nothing is locked.
major · a tool can check it · #a11y-input-purpose
Fields that collect personal data (name, email, address...) declare their purpose with the right autocomplete value. (WCAG 1.3.5, AA)
Why and how to check
Lets browsers and tools fill and present fields predictably.
Check: Check `autocomplete` attributes on user-data fields (axe `autocomplete-valid`).
blocker · a person checks it · #a11y-color-not-only
Color is never the only way to show information, state, errors or required fields. (WCAG 1.4.1, A)
Why and how to check
Color-blind and low-vision people, and monochrome displays.
Check: Use another cue wherever color carries meaning: text, an icon, underline, weight or pattern (techniques G14, G182). Error and required fields must not rely on color alone (F81); links in text must be distinguishable from surrounding text, for example 3:1 contrast against the text (G183) or another cue (F73). WCAG lists color-vision simulators for checking; grayscale is a quick approximation, not a WCAG test.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.4.1 Use of Color
blocker · a person checks it · #a11y-audio-control
Audio that plays automatically for more than 3 seconds can be paused or stopped, or its volume controlled separately from the system volume. (WCAG 1.4.2, A)
Why and how to check
Autoplaying sound interferes with screen readers and with people who cannot easily find the source.
Check: Load the page with sound on; check for autoplaying audio and its controls.
blocker · a tool finds it, a person confirms · #a11y-contrast-text
Text has a contrast ratio of at least 4.5:1 against its background, or 3:1 for large text (18 pt, or 14 pt bold). (WCAG 1.4.3, AA)
Why and how to check
Low-contrast text is the most common detected failure: 83.9% of analyzed home pages.
Check: Automated check (axe `color-contrast`) misses text over images and gradients; check those by hand, in all states and themes.
Sources: How to Meet WCAG 2.2 (Quick Reference); WCAG 2.2 SC 1.4.3 Contrast (Minimum); The WebAIM Million (2026 report)
blocker · a person checks it · #a11y-resize-text
Text can be enlarged to 200% by the browser without losing content or function. (WCAG 1.4.4, AA)
Why and how to check
Many people enlarge text.
Check: WCAG's pass condition: text can be scaled to 200% using at least one scaling mechanism browsers offer (page zoom, text-only zoom, a larger default font size), with no loss of content or function. Text may shrink at smaller breakpoints if 200% is still reachable; the example is 20 px at 100% in a 1280 px window needing to display at twice that size when enlarged. Viewport-unit-only sizing is a listed failure (F94). Measured 2026-11-05: stepped rem sizes doubled under text-only enlargement; vw-only did not change; a clamp() with a vw term grew 1.14× (see content/learn/measurements).
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.4.4 Resize Text
major · a person checks it · #a11y-images-of-text
Text is real text, not an image of text, except for logos and cases where the look is essential. (WCAG 1.4.5, AA)
Why and how to check
Real text can be resized, recolored and read aloud.
Check: Check headlines, buttons and navigation for text baked into images. Exceptions: the image of text can be customized by the user, the presentation is essential, or it is a logotype.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.4.5 Images of Text
blocker · a person checks it · #a11y-reflow
At 320 CSS pixels wide (400% zoom at 1280 px) content works without two-dimensional scrolling, apart from exceptions such as maps and wide tables. (WCAG 1.4.10, AA)
Why and how to check
People with low vision zoom in and rely on a single column.
Check: Set the viewport to 320 px wide (vertical scrolling content) and 256 px tall (horizontal scrolling content); ordinary content must not need two-dimensional scrolling. Exceptions: maps, diagrams, video, games, presentations, data tables, toolbar-in-view interfaces. A horizontal carousel passes if each panel fits 320 px; a table in its own scroll container passes if the content around it reflows. Also run it at 320 px with the root font size at 200% (and with the WCAG text-spacing overrides applied): fixed rem tracks (minmax(15rem,1fr) is 480 px at 200%), padding that leaves a 96 px column, and display words wider than the viewport are the usual causes. Measured in Chrome on the 2026-11 recipe demos: this extra pass found overflow in 11 of 18 demos that passed the plain 320 px check.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 1.4.10 Reflow
blocker · a tool finds it, a person confirms · #a11y-non-text-contrast
UI components and graphics needed to understand the content have at least 3:1 contrast against adjacent colors. (WCAG 1.4.11, AA)
Why and how to check
Buttons, inputs, focus rings and icons must be perceivable.
Check: Check borders of inputs, icon buttons, toggles and focus indicators; automated tools cover only part.
blocker · a person checks it · #a11y-text-spacing
Content still works when users override line height, paragraph, letter and word spacing. (WCAG 1.4.12, AA)
Why and how to check
People with dyslexia and low vision change spacing.
Check: Apply a text-spacing bookmarklet or user stylesheet; check for clipped or overlapping text.
major · a person checks it · #a11y-hover-focus-content
Content that appears on hover or focus can be dismissed without moving the pointer, can be hovered over, and stays until dismissed. (WCAG 1.4.13, AA)
Why and how to check
Magnifier users and keyboard users trigger tooltips and menus differently.
Check: Test every tooltip, mega menu and hover card with mouse, keyboard and zoom.
Sources: How to Meet WCAG 2.2 (Quick Reference); Mega Menus Work Well for Site Navigation
operable
blocker · a person checks it · #a11y-keyboard
All functionality can be operated with a keyboard alone. (WCAG 2.1.1, A)
Why and how to check
Keyboard, switch and voice users.
Check: Tab, Shift+Tab, Enter, Space and arrow keys through the whole page, including menus and dialogs.
blocker · a person checks it · #a11y-no-keyboard-trap
Keyboard focus can always be moved away from any component. (WCAG 2.1.2, A)
Why and how to check
A trap locks users in.
Check: Tab through every widget, including embeds, modals and 3D canvases. Deliberately restricting focus (a modal dialog) is allowed if users can leave by Esc or buttons; any non-standard exit method must be explained to the user.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 2.1.2 No Keyboard Trap
major · a person checks it · #a11y-char-shortcuts
Single-character keyboard shortcuts can be turned off, remapped, or only work when the component has focus. (WCAG 2.1.4, A)
Why and how to check
Speech-input users trigger shortcuts by accident.
Check: List all single-key shortcuts (letters, numbers, punctuation, symbols, also those needing Shift, and key sequences such as G then A). Each must be turn-off-able, remappable to include a non-printable key, or active only while its component has focus. Exceptions: listbox type-ahead, menus opened with a modified key, access keys.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 2.1.4 Character Key Shortcuts
major · a person checks it · #a11y-timing
Time limits can be turned off, adjusted or extended, unless an exception applies. (WCAG 2.2.1, A)
Why and how to check
People need different amounts of time.
Check: Look for timeouts, auto-advancing content and expiring sessions.
blocker · a person checks it · #a11y-pause-stop-hide
Moving, blinking, scrolling or auto-updating content that starts automatically and runs alongside other content can be paused, stopped or hidden. (WCAG 2.2.2, A)
Why and how to check
Motion distracts and can make content unreadable. WCAG 2.2.2 (see the Understanding page) sets the five-second threshold for automatically starting moving content.
Check: Applies to moving, blinking or scrolling content that starts automatically (including by hover, focus or scrolling into view), lasts more than five seconds and runs alongside other content; auto-updating content has no five-second exception. A control that pauses only while focused does not qualify. Best practice is one control for all moving content. Check carousels, background video, marquees and ambient animations.
Sources: How to Meet WCAG 2.2 (Quick Reference); Homepage design principles; Understanding SC 2.2.2 Pause, Stop, Hide
blocker · a person checks it · #a11y-flashes
Nothing flashes more than three times per second unless below the defined thresholds. (WCAG 2.3.1, A)
Why and how to check
Flashing can trigger seizures.
Check: Pass if no more than three general or red flashes occur in any one second (content within that limit passes without a tool), or the flashes stay below the thresholds: a general flash is an opposing luminance change of at least 10% with the darker state below 0.80, and the flashing area stays within 0.006 steradians of a 10-degree field (WCAG estimates 341×256 px at 1024×768).
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 2.3.1 Three Flashes or Below Threshold
blocker · a tool finds it, a person confirms · #a11y-bypass-blocks
A mechanism, such as a skip link or correct landmarks and headings, lets users skip repeated blocks. (WCAG 2.4.1, A)
Why and how to check
Keyboard and screen reader users should not tab through the full navigation every time.
Check: The skip link should be the first focusable control and may be visible only on focus; WCAG does not require it to be visible when it lacks focus (G1). Landmarks (ARIA11) or headings (H69) are also sufficient. WebAIM: a hidden skip link must become prominently visible on focus; do not use display:none or the hidden attribute; a skip link is not strictly required where a starting h1 or a main region meets the criterion.
Sources: How to Meet WCAG 2.2 (Quick Reference); Understanding SC 2.4.1 Bypass Blocks; axe-core rule descriptions; Technique G1: Adding a skip link to main content; Skip Navigation Links
blocker · a tool can check it · #a11y-page-title
The page has a title that describes its purpose. (WCAG 2.4.2, A)
Why and how to check
The title is the first thing a screen reader says and what appears in tabs and history.
Check: Check `<title>` is present, specific and different from other pages (axe `document-title`).
blocker · a person checks it · #a11y-focus-order
Keyboard focus moves in an order that keeps meaning and operability. (WCAG 2.4.3, A)
Why and how to check
Random jumps disorient.
Check: Tab through the page and compare focus order with the visual and logical order.
blocker · a tool finds it, a person confirms · #a11y-link-purpose
The purpose of each link can be understood from its text, or from its text with its context. (WCAG 2.4.4, A)
Why and how to check
Screen reader users list links out of context. Empty links appear on 46.3% of analyzed home pages.
Check: List all links; flag "click here", "more", and links with no name (axe `link-name`).
Sources: How to Meet WCAG 2.2 (Quick Reference); Homepage design principles; The WebAIM Million (2026 report)
major · a person checks it · #a11y-headings-labels
Headings and labels describe their topic or purpose. (WCAG 2.4.6, AA)
Why and how to check
Helps people scan and navigate.
Check: Read the headings list; each should make sense on its own.
blocker · a person checks it · #a11y-focus-visible
Every keyboard-operable element shows a visible focus indicator. (WCAG 2.4.7, AA)
Why and how to check
Keyboard users must see where they are.
Check: Tab through the page; the focus indicator is always visible with 3:1 contrast against what it sits on, and `outline: none` is never left without a replacement. In forced-colors mode outline and border colors are forced to system colors and stay visible, while box-shadow is removed, so draw focus with outline or border, not box-shadow alone.
Sources: How to Meet WCAG 2.2 (Quick Reference); :focus-visible; forced-colors media feature
blocker · a person checks it · #a11y-focus-not-obscured
A focused element is not entirely hidden behind sticky headers, cookie banners or chat widgets. (WCAG 2.4.11, AA, new in 2.2)
Why and how to check
Keyboard users lose their place when focus lands under an overlay.
Check: Tab through the page with sticky elements and banners present, at several viewport sizes. A focused item may be partly covered but not entirely (failure F110). Technique C43 (CSS scroll-padding) is the sufficient technique listed; a banner can also pass by being modal or closing when it loses focus.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2; Understanding SC 2.4.11 Focus Not Obscured (Minimum)
blocker · a person checks it · #a11y-pointer-gestures
Anything that needs a multi-point or path-based gesture (pinch, swipe) can also be done with a single pointer, unless essential. (WCAG 2.5.1, A)
Why and how to check
People with motor impairments cannot always do complex gestures.
Check: Check swipers, maps and sliders for single-tap alternatives.
major · a person checks it · #a11y-pointer-cancellation
Single-pointer actions can be cancelled or undone, or do not fire on the down event. (WCAG 2.5.2, A)
Why and how to check
Prevents accidental activation.
Check: Check that buttons act on release, and drag-start actions can be aborted.
major · a tool finds it, a person confirms · #a11y-label-in-name
The accessible name of a control contains its visible label text. (WCAG 2.5.3, A)
Why and how to check
Speech-input users say what they see.
Check: Compare visible labels with aria-label and names in the accessibility tree.
minor · a person checks it · #a11y-motion-actuation
Functions triggered by device motion also have a normal control and can be turned off. (WCAG 2.5.4, A)
Why and how to check
Not everyone can shake or tilt a device.
Check: Look for shake, tilt and gyroscope effects.
blocker · a person checks it · #a11y-dragging
Anything that needs dragging can also be done with a single press, without dragging. (WCAG 2.5.7, AA, new in 2.2)
Why and how to check
Dragging is hard or impossible for some people.
Check: Check sliders, sortable lists and reveal-on-drag effects for a single-pointer alternative (buttons, tap-to-place, direct selection, typed value; technique G219). A keyboard alternative alone does not satisfy this criterion; controls must work with a pointer, though they may be visually hidden.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2; Understanding SC 2.5.7 Dragging Movements
blocker · a tool finds it, a person confirms · #a11y-target-size
Pointer targets are at least 24 by 24 CSS pixels, or have enough spacing around them, with listed exceptions. (WCAG 2.5.8, AA, new in 2.2)
Why and how to check
Small targets are hard to hit with a finger, a tremor or a pointer.
Check: Measure icon buttons, close buttons, footer links and inline controls. axe `target-size` is disabled by default, so enable it. Exceptions: spacing (a 24 px circle centered on an undersized target must not touch another target or its circle; 20×20 px buttons with a 4 px gap pass, with no gap they fail), equivalent control, inline text, user-agent controls, essential.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2; Understanding SC 2.5.8 Target Size (Minimum); axe-core rule descriptions
understandable
blocker · a tool can check it · #a11y-lang
The page declares its main language, and passages in another language are marked. (WCAG 3.1.1 A, 3.1.2 AA)
Why and how to check
Screen readers pronounce text using the declared language. 13.5% of analyzed home pages lack it.
Check: Check `<html lang>` and `lang` on foreign phrases (axe `html-has-lang`, `html-lang-valid`).
Sources: How to Meet WCAG 2.2 (Quick Reference); The WebAIM Million (2026 report)
blocker · a person checks it · #a11y-on-focus-input
Moving focus to, or changing the value of, a control does not by itself cause a major change of context such as navigation or opening a new window. (WCAG 3.2.1 A, 3.2.2 A)
Why and how to check
Unexpected changes disorient keyboard and screen reader users.
Check: Tab through selects, radios and inputs; nothing submits or navigates without a deliberate action.
major · a person checks it · #a11y-consistent
Repeated navigation keeps the same relative order, components with the same function are labeled the same way, and help mechanisms appear in the same place. (WCAG 3.2.3 AA, 3.2.4 AA, 3.2.6 A)
Why and how to check
Predictability helps everyone, especially people with cognitive disabilities.
Check: Compare header, footer, help links and component labels across several pages.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2; 10 Usability Heuristics for User Interface Design
blocker · a person checks it · #a11y-error-identification
Detected input errors are identified and described in text, and where possible a correction is suggested. (WCAG 3.3.1 A, 3.3.3 AA)
Why and how to check
People need to know what went wrong and how to fix it.
Check: Submit forms with bad and empty input; each error is announced, specific and linked to its field.
Sources: How to Meet WCAG 2.2 (Quick Reference); Web Form Design Guidelines
blocker · a tool finds it, a person confirms · #a11y-labels-instructions
Every input has a visible label or instructions, and format requirements are stated before the user hits an error. (WCAG 3.3.2, A)
Why and how to check
Missing form labels occur on 51% of analyzed home pages.
Check: Check every field for a programmatic label (axe `label`); placeholders do not count as labels.
Sources: How to Meet WCAG 2.2 (Quick Reference); Web Form Design Guidelines; The WebAIM Million (2026 report)
major · a person checks it · #a11y-error-prevention
Submissions that have legal or financial effect, or change data, can be reversed, checked or confirmed. (WCAG 3.3.4, AA)
Why and how to check
Prevents costly mistakes.
Check: Walk through checkout, signup with payment, and deletion flows.
major · a person checks it · #a11y-redundant-entry
Information already given in a process is filled in automatically or can be selected, not asked again. (WCAG 3.3.7, A, new in 2.2)
Why and how to check
Re-entering data is hard for people with memory or motor difficulties.
Check: Walk through multi-step forms; check for repeated fields.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2
blocker · a person checks it · #a11y-accessible-auth
Login does not require remembering, solving puzzles or transcribing, unless an alternative or helper exists (password managers, paste, passkeys). (WCAG 3.3.8, AA, new in 2.2)
Why and how to check
Cognitive tests block many people.
Check: Check that password fields allow paste and autofill, and that CAPTCHAs have alternatives.
Sources: How to Meet WCAG 2.2 (Quick Reference); What's New in WCAG 2.2
robust
blocker · a tool finds it, a person confirms · #a11y-name-role-value
Every interface component exposes its name, role, state and value to assistive technology. (WCAG 4.1.2, A)
Why and how to check
Custom widgets are invisible or confusing otherwise. Empty buttons appear on 30.6% of analyzed home pages.
Check: Inspect the accessibility tree; check icon buttons for names (axe `button-name`); test custom controls in a screen reader.
Sources: How to Meet WCAG 2.2 (Quick Reference); The WebAIM Million (2026 report)
blocker · a person checks it · #a11y-status-messages
Status messages (success, errors, progress, results counts) are announced by assistive technology without moving focus. (WCAG 4.1.3, AA)
Why and how to check
Without it, screen reader users miss feedback.
Check: Trigger form confirmations, loading and search-result updates with a screen reader running.
practice
major · a person checks it · #a11y-native-first
Use native HTML elements before ARIA. No ARIA is better than bad ARIA.
Why and how to check
Incorrect ARIA misrepresents the interface and can seriously harm screen reader users.
Check: Review custom widgets; replace `div role="button"` with `<button>`, and remove ARIA that duplicates or contradicts native semantics.
major · a person checks it · #a11y-role-is-a-promise
If you add an ARIA role, implement the keyboard behavior and states that role promises, following the matching APG pattern.
Why and how to check
A role without its behavior breaks the user's expectations.
Check: Compare each custom component with its pattern in the APG (tabs, dialog, menu button, carousel, disclosure...) and test with a screen reader.
Sources: ARIA Authoring Practices Guide: Read Me First; ARIA Authoring Practices Guide: Patterns
major · a person checks it · #a11y-test-manually
Test accessibility manually as well as with automated tools: keyboard-only, a screen reader, zoom and text-size changes, and color checks.
Why and how to check
Automated tools do not catch all errors; GOV.UK states this explicitly.
Check: Run axe or Lighthouse for a baseline, then do the manual checklist above.
Sources: Testing for accessibility
major · a person checks it · #a11y-test-with-people
Test with people who use assistive technology, including disabled and older users, and get a formal audit before launch for significant services.
Why and how to check
Only real use shows what actually works.
Check: Recruit participants with different needs; include assistive technology in usability sessions.
Sources: Testing for accessibility; Inclusive Design
major · a tool finds it, a person confirms · #a11y-reduced-motion
Respect the user's reduced-motion setting: provide a way to reduce or remove non-essential animation, and keep the logo, tagline and headline static.
Why and how to check
Motion can cause discomfort for people with vestibular disorders and can distract. MDN: honor the setting by replacing large-scale motion (scaling, panning) with gentler effects such as fades, without losing what the animation conveys. NN/g also recommends controls to reduce motion.
Check: Enable "reduce motion" in the operating system and check that animations respond: large movement is replaced by fades or removed, and information the animation conveyed remains. WCAG 2.3.3 (Level AAA, Animation from Interactions) names parallax and scroll-triggered motion and points to the CSS media query (technique C39) and its JavaScript equivalent (SCR40).
Sources: prefers-reduced-motion; Homepage design principles; Understanding SC 2.3.3 Animation from Interactions (AAA)
Usability.
The message, navigation, forms and speed: what lets a visitor do what they came for. 30 rules.
message
major · a person checks it · #ux-purpose-clear
The homepage says who you are, what you offer and how you differ, quickly, in the audience's own words.
Why and how to check
When the purpose is not visible at a glance, potential customers leave.
Check: Five-second test: show the first screen and ask what the site is for; compare answers with the brief.
Sources: Homepage design principles
major · a person checks it · #ux-tagline-specific
Use a plain, specific tagline and headline instead of generic welcome messages or jargon.
Why and how to check
Users scan and spend most of their time at the top; vague text wastes it.
Check: Read the headline without the logo: could any competitor say the same?
Sources: Homepage design principles
minor · a person checks it · #ux-informative-imagery
Images carry information about the offer or brand rather than serving as decorative stock.
Why and how to check
Decorative imagery adds weight without meaning.
Check: For each major image ask what it communicates that the text does not.
Sources: Homepage design principles
minor · a person checks it · #ux-real-content-examples
Show specific samples of real content (products, articles, projects) instead of only generic category links.
Why and how to check
Concrete examples reveal what the site offers.
Check: Check that the first two screens show actual items, not only labels.
Sources: Homepage design principles
major · a tool finds it, a person confirms · #ux-descriptive-labels
Links and buttons have specific descriptive labels, not "Click here" or "Learn more".
Why and how to check
Users scan, and screen reader users hear links out of context.
Check: List every link and button label out of context.
Sources: Homepage design principles; How Users Read on the Web
major · a person checks it · #ux-scannable-text
Make text scannable: meaningful subheadings, short paragraphs with one idea each, lists, and the conclusion first.
Why and how to check
In NN/g's study (Nielsen, 30 Sep 1997; sample size not stated) 79% of users always scanned and 16% read word by word. Concise text scored +58%, scannable layout +47% and the combined version +124% on usability, relative to a promotional-writing control.
Check: Read only headings, first sentences and bold text; the page should still make sense.
Sources: How Users Read on the Web; F-Shaped Pattern For Reading Web Content
minor · a person checks it · #ux-objective-language
Prefer plain factual statements to boastful marketing language.
Why and how to check
Objective language scored +27% on usability relative to a promotional control (Nielsen, 30 Sep 1997; sample size not stated); users spend effort filtering out hype.
Check: Highlight superlatives and claims; each should be backed by a fact or removed.
Sources: How Users Read on the Web
major · a person checks it · #ux-front-load
Put the most important information in the first screens and the first two paragraphs; begin headings, bullets and paragraphs with meaningful words.
Why and how to check
Eyetracking of 232 users found a rough F-shaped pattern (Nielsen, 16 Apr 2006; a later study 11 years on reportedly confirmed it); it is a shape, not a law. In a 2018 study (Fessenden, 15 Apr 2018; 120 participants, over 130,000 fixations) 57% of viewing time was above the fold and 74% in the first two screenfuls.
Check: List the top three things a visitor must learn and check they appear in the first two screens.
Sources: F-Shaped Pattern For Reading Web Content; Scrolling and Attention
major · a tool finds it, a person confirms · #ux-primary-cta-above-fold
The primary call to action is visible without scrolling, with visual hierarchy built around the top task.
Why and how to check
Attention drops sharply after the fold: 57% of viewing time was above the fold in 2018 versus 80% in 2010; 74% fell in the first two screenfuls (NN/g, 15 Apr 2018, 120 participants).
Check: Automated: find the primary call to action (the first visible link or button in the hero, else in main, that starts with an action verb or is drawn as a button; else the first button-like element on the page) and fail when its bounding box is not fully inside the first viewport at 1440x900 and 390x844. The evidence names the element it took and why (for example 'first filled button in the hero'). When the page's main action is something else (a board of rows, a form, a single link), mark that element with a data-primary-cta attribute: it overrides the heuristic and its bounding box is measured instead. A person confirms the element found is the page's top task and that the visual hierarchy is built around it.
Sources: Scrolling and Attention; Homepage design principles
minor · a person checks it · #ux-false-floor
Avoid false floors: signal that there is more below, for example with cut-off content.
Why and how to check
Minimalist layouts can look finished before they are.
Check: Look at the first screen: does anything suggest more content below?
Sources: Scrolling and Attention
navigation
minor · a person checks it · #ux-logo-home
The logo links to the homepage and sits where people expect it (top-left in left-to-right languages); the homepage looks distinct from interior pages.
Why and how to check
Easy access to the starting point is a basic orientation need.
Check: Click the logo from several pages.
Sources: Homepage design principles
advisory · a person checks it · #ux-conventional-layout
Use conventional layouts and patterns where the user's goal is practical, so existing mental models apply; depart from convention deliberately.
Why and how to check
Users bring expectations from other sites. NN/g recommends standard layouts for homepages.
Check: Ask whether each unconventional choice is worth its cost, and write the reason down.
Sources: Homepage design principles; 10 Usability Heuristics for User Interface Design
minor · a person checks it · #ux-consistency
Keep labels, components and behavior consistent within the site and with common conventions.
Why and how to check
Heuristic 4 (consistency and standards); GOV.UK says "consistent, not uniform".
Check: Audit repeated components and labels across pages.
Sources: 10 Usability Heuristics for User Interface Design; Government design principles
major · a person checks it · #ux-carousel-caution
Prefer a single static hero to a rotating carousel. If you use a carousel: five frames or fewer, controls inside it with a clear current state, no auto-advance on mobile, timing based on reading speed (about 3 words per second), and the important content repeated elsewhere.
Why and how to check
Users often scroll past carousels and miss frames after the first. This is a 2013 article (reviewed 2026) with no stated sample size, so the guidance is guideline-level, not study-level.
Check: Check frame count, controls, auto-rotation and whether any critical message exists only inside the carousel. APG carousel pattern: an auto-rotating carousel has a stop/start button as its first tab stop (its label changes, so no aria-pressed), rotation stops when focus enters or the mouse hovers, and previous and next buttons are required.
Sources: Designing Effective Carousels; How to Meet WCAG 2.2 (Quick Reference); Carousel Pattern
major · a person checks it · #ux-no-ad-lookalikes
Do not style important content like an ad (banner shapes, right-rail boxes, animated strips).
Why and how to check
Users ignore elements they perceive as ads, including content nearby and large distinct images (NN/g, 22 Apr 2018, 26 participants in the latest eyetracking study). One participant's gaze on one page gave the right rail 33 times fewer fixations than its size suggested (1 of 132 content-area fixations); that is an illustration, not a statistic.
Check: Check that key messages are not in banner-like containers.
major · a person checks it · #ux-no-popups
Avoid popups and splash screens on the homepage, except where legally required (cookie consent, age checks).
Why and how to check
They are among the most disliked web design elements, per NN/g.
Check: Load the page fresh; note anything that interrupts the first view.
Sources: Homepage design principles
minor · a person checks it · #ux-minimal-motion
Keep motion minimal on the homepage; moving elements are often mistaken for ads, and the logo, tagline and headline should stay static.
Why and how to check
Motion draws attention away from the message and can look like advertising.
Check: Watch the first 10 seconds with and without motion.
Sources: Homepage design principles; Banner Blindness: Old and New Findings
forms
major · a person checks it · #ux-form-length
Keep forms as short as possible: remove fields that can be derived, collected later or dropped; limit optional fields to one or two and label them.
Why and how to check
Fewer fields mean higher completion.
Check: List every field and justify it.
Sources: Web Form Design Guidelines
minor · a person checks it · #ux-form-layout
Use a single column with labels close to their fields (above on mobile); only short related fields share a row.
Why and how to check
A single column keeps people moving down the form.
Check: Check layout at desktop and mobile widths.
Sources: Web Form Design Guidelines
major · a tool finds it, a person confirms · #ux-form-no-placeholder-label
Do not use placeholder text as the label or as the only instruction.
Why and how to check
NN/g reports usability problems with placeholders; the text disappears on input.
Check: Check every field has a visible, persistent label.
Sources: Web Form Design Guidelines; How to Meet WCAG 2.2 (Quick Reference)
minor · a person checks it · #ux-form-controls
Use radio buttons instead of dropdowns for two or three options; size inputs to the expected value; state format rules up front; avoid Reset and Clear buttons.
Why and how to check
Reduces hidden options and accidental data loss.
Check: Review each control type.
Sources: Web Form Design Guidelines
major · a person checks it · #ux-form-errors
Signal errors with several cues, not color alone; explain the problem specifically; keep what the user typed.
Why and how to check
Users must understand and fix errors without redoing work.
Check: Submit invalid data and read the messages.
Sources: Web Form Design Guidelines; How to Meet WCAG 2.2 (Quick Reference)
performance
major · a tool can check it · #perf-lcp
Largest Contentful Paint is 2.5 seconds or less at the 75th percentile of page loads, for mobile and desktop separately.
Why and how to check
Reflects whether the page feels like it is loading. NN/g cites a Google finding that bounce likelihood rises 32% as load time goes from 1 to 3 seconds; the method is not given, so treat it as secondary evidence.
Check: Check field data (CrUX) and lab data (Lighthouse). LCP candidates are img, image in svg, video (poster load or first frame, whichever is earlier), url() background images (not gradients) and block-level text.
Sources: Web Vitals; Homepage design principles; Largest Contentful Paint (LCP)
major · a tool can check it · #perf-inp
Interaction to Next Paint is 200 ms or less at the 75th percentile.
Why and how to check
Reflects whether the page feels responsive.
Check: Check field data; lab tools only approximate it.
Sources: Web Vitals
major · a tool can check it · #perf-cls
Cumulative Layout Shift is 0.1 or less at the 75th percentile.
Why and how to check
Reflects whether the page feels stable.
Check: Check field and lab data. Common fixes: width and height or aspect-ratio on images and video, reserved space for late content, font-display: optional with a close fallback, animate transform instead of top, left or box-shadow.
Sources: Web Vitals; Optimize Cumulative Layout Shift
major · a person checks it · #perf-core-first
Core content loads immediately and does not depend on a heavy effect layer; build effects on top of working content.
Why and how to check
Progressive enhancement keeps the page usable when effects fail or are slow.
Check: Disable JavaScript and throttle the network; the message and navigation should still appear.
Sources: Homepage design principles; Resilient Web Design (chapters 2 and 5)
process
major · a person checks it · #ux-test-iteratively
Test usability with about five people per audience group per round, fix problems, and test again; several small rounds beat one large one.
Why and how to check
Nielsen (18 Mar 2000): the first user finds about a third of problems (31% averaged over projects studied), five users find about 85%, and three rounds of five beat one round of fifteen; test 3 to 4 users per group for two groups and at least 3 per group for three or more. Norman (2013, ch. 6) also advises five and iterating.
Check: Run the planned rounds; record problems and fixes.
major · a person checks it · #ux-start-with-needs
Start from real user needs and test assumptions with real people from the beginning.
Why and how to check
GOV.UK and the U.S. Web Design System both list it as their first principle.
Check: Check the brief cites research, not only internal opinion.
Sources: Government design principles; U.S. Web Design System design principles
Visual craft.
Type, colour, spacing, imagery and motion: the details that separate careful work from a template. 41 rules.
typography: line length
major · a tool finds it, a person confirms · #craft-measure-body
Running-text paragraphs are never wider than 75 characters per line at any viewport width from 320px to 1920px (the target is about 66 characters; in CSS that is about 50ch to 55ch for proportional fonts, because ch is the width of the digit 0).
Why and how to check
web.dev (citing Bringhurst) calls 45-75 characters satisfactory and about 66 ideal for single-column text; USWDS says 45-90 with a long-text target of 66; Butterick says 45-90; GOV.UK keeps main content in a two-thirds column for no more than about 75; WCAG 1.4.8 (AAA) caps at about 80. The upper edge differs between sources (75, 80, 90), so the rule uses the strictest common figure. Butterick calls missing line-length checks the main flaw of many responsive designs. Our own measurement (Chrome 156, macOS, 2026-11-06): max-inline-size 66ch produced 87 to 96 characters per line with system-ui, Georgia, Arial and Helvetica Neue; 50ch gave 67 to 74. So set the container to about 50ch to 55ch, not 66ch, and count real characters per line.
Check: At widths 320, 768, 1280 and 1920 px, find every paragraph/list-item/blockquote with more than two lines of text. Compute characters per line (text length divided by rendered line count, or compute container inline-size divided by the width of one '0' glyph, i.e. in ch). Fail if any exceeds 75. For the lower bound, where the viewport allows it (>= 768px), flag single-column text below 45 characters per line for manual review.
Sources: Learn Design: Typography; Typography (USWDS); Line length; Layout; Understanding SC 1.4.8 Visual Presentation
typography: units
major · a tool can check it · #craft-relative-units-text
Font sizes are declared in rem or em (not px), the root font-size is not fixed in px, and text containers are width-limited in ch, rem or em.
Why and how to check
web.dev says to use relative units for line length (66ch rather than 700px) so it scales with user font settings, and for font size so text follows user preferences; Roselli advises not to set the base font size in px but to start from the browser default; MDN defines rem as the root font size, which users can change in preferences.
Check: Scan CSS for font-size declarations with px units (allow px only in fallback lines inside clamp() bounds that also use rem), for `html { font-size: <n>px }`, and for max-width/max-inline-size on text containers in px.
Sources: Learn Design: Typography; Accessibility (course module); Responsive type and zoom; `<length>` (ch, em, rem)
typography: size minimum
major · a tool can check it · #craft-body-size-min
Running text (paragraphs, list items, form text, table cells) computes to at least 16px at every viewport width; only captions, footnotes and legal text may be smaller.
Why and how to check
USWDS asks for an effective size of at least 16px for most text and reserves smaller sizes for captions and footnotes; GOV.UK sets body at 19px with its smallest style at 16px, to be used sparingly; Butterick gives 15-25 px as the web range for body text. Sources differ on the floor (15 vs 16), so the rule takes 16.
Check: At the audit's desktop (1440 px) and mobile (390 px) viewports, collect computed font-size of every visible text node inside p, li, td, th, label, button, and of every visible input, textarea and select (invisible and screen-reader-only text is ignored). Fail when any is under 16px, unless the node is inside figcaption, small, footer or a caption. List every such exempt text under 14px for manual review (14px is our review threshold, not from a source).
Sources: Typography (USWDS); Paragraphs; Type scale (GOV.UK Design System); Point size
typography: line spacing
minor · a tool can check it · #craft-line-height-body
Body text uses a unitless line-height between 1.3 and 1.7, and long-form reading text uses 1.5 or more.
Why and how to check
Butterick recommends 120-145% of point size and unitless CSS; web.dev gives 1.65 at a 66ch measure and says shorter lines can take more; USWDS says longer texts need at least 1.5 and gives tokens up to 1.75 for extended reading; GOV.UK body is 19px on 25px (about 1.32); WCAG 1.4.8 (AAA) asks that users can get 1.5. Sources overlap only loosely (1.2 to 1.65+), so the band is wide and the 1.5 floor applies to long-form only.
Check: For p, li, blockquote read the computed line-height and the authored value. Fail if the authored value has a unit (px, rem) rather than a unitless number, or if the computed ratio to font-size is outside 1.3-1.7. For article/long-form containers (more than about 300 words) require >= 1.5.
Sources: Line spacing; Learn Design: Typography; Typography (USWDS); Type scale (GOV.UK Design System); Understanding SC 1.4.8 Visual Presentation
typography: heading spacing
minor · a tool can check it · #craft-heading-line-height
Headings that wrap to two or more lines use a unitless line-height of 1.35 or less, and lower than the body line-height.
Why and how to check
USWDS says headings can use line-height 1 to 1.35 and body reading text at least 1.5; GOV.UK heading styles pair 48/50, 36/40 and 24/30 px (ratios about 1.04-1.25) against 19/25 for body.
Check: For h1-h3 and display text, read computed line-height / font-size; fail if > 1.35 or if it is >= the body ratio. Skip single-line headings.
Sources: Typography (USWDS); Type scale (GOV.UK Design System)
typography: type scale
minor · a tool finds it, a person confirms · #craft-type-scale-tokens
All text sizes come from a defined scale of named tokens with at most 8 distinct sizes in use on a page (a fluid token counts once, not per viewport).
Why and how to check
web.dev tells designers to build a typography scale into the design system; GOV.UK defines 7 scale points with line heights that are multiples of 5px and says to align any new style to an existing point; Utopia generates a named scale (the example in its post uses 5 steps up and 2 down). The count 8 is derived from those two (7 and 8 steps) and is our threshold.
Check: Collect computed font-size of all rendered text nodes at 1280px, round to 0.5px, and count distinct values. Fail above 8. Then confirm the sizes map to named tokens (custom properties) in the CSS.
Sources: Learn Design: Typography; Type scale (GOV.UK Design System); Fluid responsive typography: easy when you know how
typography: hierarchy
advisory · a tool finds it, a person confirms · #craft-heading-levels-limit
Long-form content shows at most three visually distinct heading levels, and the size step between adjacent levels is small enough that a heading is not mistaken for display text.
Why and how to check
Butterick says to use no more than three heading levels (two is better), to raise size only slightly and to emphasize with space; GOV.UK's default pairing uses just three heading styles (l, m, s) for h1-h3, with a larger fourth only for long-form pages.
Check: In article/long-form containers, list the distinct computed (font-size, font-weight) pairs of h1-h6. Fail if more than three look distinct. Manual: confirm each level is distinguishable from the next by size, weight or space.
Sources: Headings; Headings, text and captions
typography: capitalization
advisory · a tool finds it, a person confirms · #craft-heading-sentence-case
Headings and button labels are written in sentence case and are not set in all caps when longer than one short line.
Why and how to check
GOV.UK says to write headings in sentence case; Butterick advises against title case and against all-caps headings that are full sentences, allowing caps only for text shorter than one line.
Check: Extract heading text; flag headings where most words are capitalized (title case) for manual review. Flag any heading with computed text-transform: uppercase that wraps to more than one line.
Sources: Headings, text and captions; Headings; All caps
typography: weight
minor · a tool can check it · #craft-weight-thin-limited
Body text and any text under 24px uses font-weight 400 or heavier; weights of 300 or lighter appear only at display sizes (24px and above).
Why and how to check
Apple HIG says to avoid Ultralight, Thin and Light because they are hard to see at small sizes; WCAG 1.4.6 notes thin fonts render fainter due to anti-aliasing so a thicker font or a stronger color pair is advised; Butterick says heavier black/ultra weights are for display, not body. The 24px cut-off is WCAG's large-text size and our threshold.
Check: Collect computed font-weight and font-size of every text node. Fail where weight <= 300 and size < 24px. Variable fonts: check the computed value, not the file.
Sources: Typography (Apple HIG); Understanding SC 1.4.6 Contrast (Enhanced); Bold or italic
typography: families
advisory · a tool can check it · #craft-font-families-few
The page uses at most two font families (plus an optional monospace for code), with hierarchy built from size, weight and color within them.
Why and how to check
Apple HIG says to minimize the number of typefaces because many faces obscure hierarchy and hurt readability; web.dev notes each added font file increases load time. The figure of two is our own threshold: neither source states a number.
Check: Collect distinct computed font-family first entries across all rendered text, ignoring generic fallbacks and monospace on code/pre. Fail above two.
typography: fluid type
major · a tool finds it, a person confirms · #craft-fluid-type-rem-clamp
No font-size is set in vw, vh or vi alone. Text that must reach 200% on zoom (body, UI, and headings you cannot afford to clip) uses rem sizes stepped at em or rem breakpoints, with the smallest step at least 40% of the largest. A clamp() with a vw term is allowed only for decorative display type, and only after you measured that it still grows at 200% zoom.
Why and how to check
web.dev says never to use vw alone and to combine it with rem/em/ch in calc(), clamping with rem bounds; MDN says the maximum should be a relative unit at least twice the minimum so text can reach 200% (WCAG 1.4.4); Roselli shows `clamp(1.5rem, 5vw, 3rem)` growing from 64px only to about 68px at 200% zoom in a 1024px window, so the vw term can still fail 1.4.4 even when max = 2x min; Utopia cites the same warning and says to test before production. Our own measurement (content/learn/measurements/type-zoom-2026-11-05.json, Chrome): viewport-unit and clamp(rem + vw) sizes did not reach 200% at 200% zoom, because the vw term shrinks as the viewport in CSS px shrinks; stepped rem sizes with a small size of at least 40% of the large size did. The rule therefore prefers stepped rem and no longer recommends clamp(rem + vw) for text that must scale.
Check: Static: reject any font-size value that is a bare vw/vh/vi length or a clamp() with px bounds. Dynamic: at a 1024px-wide window measure the computed size of fluid headings, then emulate 200% and 300% browser zoom (viewport halves/thirds) and confirm the on-screen size at 200% is close to double; flag headings whose size grows by far less (Roselli's example: 6.25%) for manual review. Mark a clamp() with a vw term as failing for any text larger than 24px unless the measured size at 200% zoom is at least 1.5 times the 100% size.
Sources: Learn Design: Typography; clamp(); Responsive type and zoom; Clamp
typography: line breaking
advisory · a tool can check it · #craft-text-wrap-balance-pretty
Headings, captions and short display text set text-wrap: balance, and long paragraphs set text-wrap: pretty, with no layout depending on either.
Why and how to check
MDN says balance evens line lengths but browsers apply it only to blocks of six lines or fewer in Chromium and ten or fewer in Firefox (so it suits headings, captions, blockquotes), and pretty uses a slower algorithm to reduce orphans for body copy. MDN's text-wrap page lists Baseline Widely available (March 2024) while the text-wrap-style page lists Baseline 2024 newly available (October 2024), so treat it as progressive enhancement. Only MDN was read, so this is labeled our-take.
Check: Check computed text-wrap-style on h1-h3, figcaption, blockquote (expect balance) and on long p (expect pretty). Confirm nothing breaks when the declaration is removed.
Sources: text-wrap-style; text-wrap
typography: capitals
minor · a tool can check it · #craft-caps-letterspacing
Text set in uppercase or small caps has letter-spacing between 0.05em and 0.12em.
Why and how to check
Butterick: capitals are designed to sit next to lowercase and look too tight set together, so add 5-12% letterspacing (CSS 0.05-0.12em), always for all caps; the 12% upper end also matches the WCAG 1.4.12 override value of 0.12 x font size that layouts must survive.
Check: Select elements with computed text-transform: uppercase or font-variant-caps small-caps (or text that is all uppercase); fail when letter-spacing / font-size is below 0.05 or above 0.12.
Sources: Letterspacing; All caps; Summary of key rules; Understanding SC 1.4.12 Text Spacing
minor · a tool can check it · #craft-caps-short-only
Uppercase text is limited to strings shorter than one line (labels, eyebrows, buttons, short headings); no paragraph or multi-line block is uppercase.
Why and how to check
Butterick: all caps is harder to read at body sizes because readers recognize lowercase word shapes, suits text shorter than one line, and whole paragraphs set in caps get skimmed; emphasize with rules, a label-style heading or a larger size instead.
Check: Find elements with text-transform: uppercase (or uppercase source text) and compute rendered line count; fail when any has two or more lines or more than about 40 characters (40 is our review threshold).
Sources: All caps; Summary of key rules
typography: tracking
advisory · a tool can check it · #craft-body-tracking-normal
Body paragraphs in lowercase text keep letter-spacing at normal (0); added tracking on body text is limited to small text, and large display headlines may be tightened but not widened.
Why and how to check
Butterick: lowercase usually needs no letterspacing, text smaller than about 9pt may benefit from slight added spacing, and large lowercase headlines are often tightened.
Check: For p, li and td at 16px or more, fail if computed letter-spacing is not 'normal' or 0 (allow within +-0.01em). For text 24px and above, flag positive letter-spacing on lowercase headlines for review.
Sources: Letterspacing; Summary of key rules
typography: alignment
minor · a tool can check it · #craft-align-left-body
Multi-line body paragraphs are start-aligned (left in LTR); none is center-aligned or justified, except justified text that also has hyphens: auto and a declared lang.
Why and how to check
Butterick left-aligns web text because browser justification is crude, says justified text must be hyphenated, and says to use centered text sparingly; WCAG 1.4.8 (AAA) asks that users can avoid full justification because it can make reading harder for some people. Butterick's remark that CSS hyphenation support is inconsistent is undated and may be stale.
Check: For paragraphs with 3 or more rendered lines, fail if computed text-align is center or justify (for justify, pass only if hyphens is auto and html has lang).
Sources: Justified text; Hyphenation; Summary of key rules; Understanding SC 1.4.8 Visual Presentation
typography: emphasis
advisory · a tool can check it · #craft-emphasis-single
Inline emphasis uses either bold or italic on a run of text, never both, and is limited to short phrases; underline is reserved for links.
Why and how to check
Butterick: use bold or italic, not both, only sparingly, because both are harder to read than regular text and stacking leaves no room for further emphasis; do not underline text except possibly web links.
Check: Find inline em/strong/b/i/span runs with both font-weight >= 600 and font-style italic; find text-decoration: underline on non-link elements; flag emphasized runs over about 12 words for review (12 is our threshold).
Sources: Bold or italic; Summary of key rules
color: contrast beyond minimum
advisory · a tool can check it · #craft-contrast-body-enhanced
Primary body text reaches at least 7:1 contrast against its background, and secondary or muted text at least 4.5:1.
Why and how to check
WCAG 1.4.6 (AAA) sets 7:1 (4.5:1 for large text); Apple HIG says 4.5:1 minimum and 7:1 for custom foreground/background colors, especially small text. The 4.5:1 floor for secondary text is WCAG 1.4.3 AA (already covered by a11y-contrast-text); this rule only adds the 7:1 target for primary text, which is AAA and so advisory.
Check: Compute the WCAG 2.x contrast ratio of the main body text color against its actual background in both light and dark themes; fail the target if under 7:1, and list muted text under 4.5:1 as a failure of a11y-contrast-text.
Sources: Understanding SC 1.4.6 Contrast (Enhanced); WCAG 2.2 SC 1.4.3 Contrast (Minimum); Dark Mode (Apple HIG)
color: text over images
major · a tool finds it, a person confirms · #craft-contrast-text-on-media
Text placed over photos, video or gradients keeps 4.5:1 (3:1 for large text) against the worst-case backdrop beneath each line of text, using an overlay or scrim where needed.
Why and how to check
WCAG 1.4.3 Understanding says background images that fail to give sufficient contrast with text are a failure, and contrast is judged on the color pairs actually adjacent; no gradient-specific method is given, so sampling the lightest/darkest backdrop pixel is our reading of it. This refines the existing a11y-contrast-text for hero imagery.
Check: Screenshot each section with text on imagery at 375, 768 and 1440 px (and at the video's brightest frame). Sample the backdrop under each text line's bounding box, take the pixel closest to the text color, compute the WCAG ratio. Fail below 4.5:1 (3:1 for 24px+ or 18.66px+ bold).
Sources: WCAG 2.2 SC 1.4.3 Contrast (Minimum)
color: palette roles
minor · a tool can check it · #craft-color-role-tokens
Colors are defined once as named role tokens (text, secondary text, background, surface, border, accent/link, focus, error, success) and components reference the tokens; literal color values appear only in the token definitions.
Why and how to check
GOV.UK says to reference functional colours by name, not hex, so palette updates propagate, and lists role colours such as text, secondary text, border, focus, error and success; web.dev's theming module stores colors in custom properties redefined for dark mode; Apple HIG prefers semantic colors over hard-coded values.
Check: Parse the CSS: collect hex/rgb/hsl/oklch literals outside :root, [data-theme] and @media (prefers-color-scheme) token blocks. Fail if components contain literals; confirm the token set includes at least text, background, border, accent, focus, error and success.
Sources: Colour; Theming; Color (Apple HIG)
color: semantic color
minor · a tool finds it, a person confirms · #craft-color-meaning-reserved
A color that signals interactivity or a state (link/accent, focus, error, success, warning) is not reused for non-interactive decoration or for a different state.
Why and how to check
GOV.UK says to use functional colours only in their intended context, to reserve the focus yellow for keyboard focus and to use the error colour only for errors; Apple HIG says not to use the same color for different meanings, for example a brand color that marks a button as interactive should not style non-interactive text.
Check: List the computed colors of links, buttons, focus rings, error and success messages. Fail if any of those exact colors also appears as text or fill on non-interactive elements, or if error and success use the same hue. Confirm by reading the token definitions.
Sources: Colour; Color (Apple HIG)
color: dark mode and theming
minor · a tool can check it · #craft-dark-mode-scheme
If the site has a dark theme, it declares color-scheme (meta tag before stylesheets and CSS) and follows prefers-color-scheme by default, so scrollbars, form controls and canvas render in the matching scheme.
Why and how to check
MDN: color-scheme makes the browser adjust canvas, scrollbar and form control colors, the meta tag placed before stylesheets avoids a flash of wrong colors, and prefers-color-scheme (Baseline widely available since Jan 2020) is what styles the rest; light-dark() (Baseline 2024) only works when color-scheme is set. web.dev's theming module redefines custom properties in the dark media query. Apple HIG says to avoid an app-specific appearance setting that overrides the system choice.
Check: Emulate colorScheme 'dark' and 'light'. Confirm getComputedStyle(html).colorScheme includes the scheme, a <meta name='color-scheme'> exists in head before stylesheets, native inputs/scrollbars are dark in dark, and no first-paint flash of the light theme occurs (screenshot at DOMContentLoaded).
Sources: color-scheme; prefers-color-scheme; light-dark(); Theming; Dark Mode (Apple HIG)
major · a tool can check it · #craft-dark-mode-reverify
In the dark theme every text and UI color pair is re-checked against the same contrast thresholds (4.5:1 text, 3:1 large text and UI components), and the palette is not made by inverting or filtering the light one.
Why and how to check
Apple HIG: dark colors are not simple inversions, keep 4.5:1 in every appearance and test with Increase Contrast; web.dev's theming module redefines each token for dark and darkens photos with `brightness(.8) contrast(1.2)` rather than inverting; WCAG 1.4.3 thresholds apply to the page as rendered. White-background images should be dimmed so they do not glow (Apple).
Check: Run the contrast audit (text 4.5:1, large 3:1, UI components 3:1) twice, once per emulated color scheme, over all pages and states (hover, focus, disabled excluded). Fail any pair that passes light but fails dark. Flag use of `filter: invert()` on the page root.
Sources: Dark Mode (Apple HIG); Theming; WCAG 2.2 SC 1.4.3 Contrast (Minimum)
color: perceptual color spaces
advisory · a tool finds it, a person confirms · #craft-palette-oklch-ramps
Tints, shades and state variants (hover, active, disabled) of a brand color are generated from one base by changing lightness with hue and chroma held fixed (OKLCH, color-mix() or relative color syntax) instead of hand-picked unrelated hex values.
Why and how to check
web.dev: in OkLCh changing hue does not change perceived lightness or saturation, which makes adjustments predictable; MDN lists lightness ramps and theming variants from one base as use cases and gives oklch() Baseline Widely available since May 2023; USWDS shows the payoff of lightness-ordered grades, where a grade gap of 40/50/70 predicts WCAG AA large/AA/AAA (written for WCAG 2.0). Contrast must still be computed with the WCAG 2.x formula; OKLCH lightness is not a contrast guarantee.
Check: Inspect the color tokens: for each ramp, confirm hover/active/tint values are derived with oklch()/color-mix()/relative color from one source token, or that their OKLCH hue differs by less than about 5 degrees and chroma follows a smooth curve (our tolerance). Still run the WCAG ratio on each text pairing.
Sources: Learn CSS: Color; oklch(); Color: overview (USWDS)
color: APCA and WCAG 3 status
color: user contrast preferences
minor · a tool finds it, a person confirms · #craft-contrast-preferences
Custom components stay legible under prefers-contrast: more and forced-colors: active (visible borders, focus and state) and the site strengthens borders or text colors when prefers-contrast: more is set.
Why and how to check
web.dev's media-features module: prefers-contrast: more lets you strengthen colors and borders, Windows contrast themes replace page colors so custom components need fixing under forced-colors, and forced-color-adjust: none is for elements where color carries meaning; its accessibility module says to test with the OS increased-contrast setting. Apple HIG: custom colors need an increased-contrast variant with significantly more differentiation.
Check: Emulate forcedColors: active and contrast: more. Screenshot key pages; confirm buttons, inputs, cards, tabs and focus rings keep a visible boundary and no text disappears. Confirm a @media (prefers-contrast: more) block exists that raises border/text contrast.
Sources: Media features; Accessibility (course module); Color (Apple HIG)
layout: spacing scale
minor · a tool can check it · #craft-spacing-scale-tokens
All margin, padding and gap values come from one spacing scale defined as tokens on a single base unit, with at most 12 distinct non-zero spacing values in use on a page.
Why and how to check
web.dev: pick a spacing system (a fixed gutter or a multiple of em) and store values as custom properties; GOV.UK defines units 0-9 in 5px steps up to 15px then larger; USWDS tokens are multiples of 8px with 4px half-steps; Utopia builds a named space palette and prefers as few sizes as possible. Base units differ (5px, 8px, em multiples), so any consistent base is acceptable. The cap of 12 is our threshold (GOV.UK has 9 non-zero units).
Check: At 1280px collect computed margin, padding, row-gap and column-gap (block/inline) of all elements, round to 0.5px, ignore 0 and 1-2px borders, and count distinct values; fail above 12 or when a value is not a multiple of the base unit (4 or 8, or the project's token base). Confirm values come from spacing custom properties.
Sources: Learn CSS: Spacing; Spacing; Spacing units (USWDS); Designing with a fluid space palette
layout: responsive spacing
advisory · a tool can check it · #craft-space-responsive-large
Section-level spacing (section padding, page gutters, large gaps) is smaller on a 360px viewport than on a 1280px viewport, while the smallest steps stay constant.
Why and how to check
GOV.UK's responsive scale keeps units 0-3 the same at every size and lets units 4-9 grow above 640px (unit 9 is 40px on small screens and 60px on large); Utopia interpolates its named space steps between a small and large viewport so layouts breathe without per-breakpoint rules.
Check: Measure the computed vertical padding/margin of the main section containers at 360 and 1280 px; require the 1280 value to be greater. Measure small component gaps (4-16px) and require them to be unchanged within 2px.
Sources: Spacing; Designing with a fluid space palette
layout: hierarchy and whitespace
minor · a tool finds it, a person confirms · #craft-group-by-proximity
Space between items inside a group (label and field, heading and its paragraph, card contents) is smaller than the space between groups or sections.
Why and how to check
Apple HIG: group related items with negative space, container shapes or separators and keep unrelated items visually distinct; alignment and indentation tell people what is related. Only this one source was read for this claim (the sources.json gestalt-secondary page was not re-read), so it is labeled our-take.
Check: For each heading, compare its margin/padding above with below; for each card or form group, compare the internal gap with the gap to the neighbouring group. Fail where the internal gap is greater than or equal to the external gap. Visual review of screenshots for ambiguous groupings.
Sources: Layout (Apple HIG); Headings
layout: responsive behavior
major · a tool finds it, a person confirms · #craft-breakpoints-content-based
Layout breakpoints are written in em/rem at the widths where the content breaks, not at named device widths, and the page has no overflow, clipped content or line-length violation at any viewport width between 320px and 1920px.
Why and how to check
web.dev: set breakpoints from your content rather than popular device sizes, use em/ch for text-based content, and prefer intrinsic grid rules (auto-fill/minmax(15em, 1fr)) and container queries over a ladder of media queries; GOV.UK says to design for a range of screen sizes rather than specific devices and to start with one column on small screens.
Check: Automated: resize one loaded page from 320 to 1920 px in 40px steps and fail at any width where the document scrolls sideways (scrollWidth > clientWidth); list the px media-query breakpoints of the readable stylesheets and flag any matching device presets (375, 414, 768 and 1024 px) for review. A person checks what the audit does not: no element overlapping its sibling's box, the measure rule at each width, and that breakpoints are written in em or rem where the content breaks.
Sources: Media queries; Macro layouts; Micro layouts; Layout
layout: container widths
advisory · a tool can check it · #craft-container-max-width
Page content sits in a centered container with a maximum width of 1400px or less; only backgrounds and full-bleed media may exceed it.
Why and how to check
GOV.UK's default maximum page width is 1020px (wider only when content needs it) and USWDS names widths up to desktop 1200px and widescreen 1400px. These are the two design systems' own choices; neither states the number as a universal rule, so the rule is advisory.
Check: At a 1920px viewport, measure the bounding box of the main content wrapper(s): fail when text, cards or controls span wider than 1400px or touch both viewport edges; allow elements flagged full-bleed (backgrounds, images, video).
Sources: Layout; Spacing units (USWDS)
imagery: layout stability
major · a tool can check it · #craft-image-dimensions-reserved
Every img and video element has width and height attributes or a CSS aspect-ratio, so its box is reserved before the file loads.
Why and how to check
web.dev: always include width and height so the browser reserves the right space and avoids layout shift even when CSS scales the image; MDN: `aspect-ratio: auto 3/2` keeps the ratio before load and the natural one after (Baseline Widely available since Sept 2021). This is the cause-level rule behind the existing perf-cls metric.
Check: For every img, video, iframe: fail if it lacks both (width and height attributes) and a computed aspect-ratio other than auto. Run with images blocked and check CLS stays at or below 0.1.
Sources: Responsive images; aspect-ratio
imagery: treatment
minor · a tool can check it · #craft-image-no-distortion
No image is shown stretched or squashed: its rendered aspect ratio matches its natural ratio unless object-fit: cover or contain crops or letterboxes it, and crops that lose the subject on small screens use object-position or a different crop via picture.
Why and how to check
web.dev: block-size: auto keeps the ratio, forcing another ratio with aspect-ratio can squash or stretch, and object-fit avoids distortion; MDN: the default fill stretches content while cover crops and contain letterboxes; web.dev's picture module says to serve a different crop, not a smaller copy, when an image loses impact at small sizes.
Check: For every img/video compute renderedW/renderedH and naturalWidth/naturalHeight. If computed object-fit is fill (or none specified) and the ratios differ by more than 2%, fail. For object-fit: cover images in hero/portrait slots, screenshot at 375 px and confirm the subject (face, product) is inside the visible area.
Sources: Responsive images; object-fit; The picture element
iconography: consistency
minor · a tool finds it, a person confirms · #craft-icons-one-family
All UI icons come from one family with a single stroke weight, corner style and fill treatment, are drawn on a shared size grid, and their stroke weight visually matches adjacent text.
Why and how to check
Apple HIG: use the same size, level of detail, stroke weight and perspective across all interface icons, and match icon weight to adjacent text weight; web.dev: icons are part of the brand identity so their style should match the rest of the site.
Check: Extract all inline SVG icons; collect distinct stroke-width, stroke-linecap/linejoin, fill-vs-outline treatment and viewBox sizes. Fail if more than one stroke-width or more than one outline/filled style is mixed in the same context. Manual: confirm icons beside text look the same weight.
Sources: Icons (Apple HIG); Icons
iconography: implementation
minor · a tool can check it · #craft-icons-svg-currentcolor
UI icons are vector (inline SVG or SVG sprite) that take their color from currentColor or theme tokens, so hover, focus and dark-theme changes apply to them; no raster icons are used for UI controls.
Why and how to check
web.dev: inline SVG can be styled with CSS (fill set to match button text, hover and focus changes applying to icon and text together) while SVG background images are harder to recolor, and SVG fill/stroke can follow the theme custom properties; Apple HIG: use vector formats so icons scale and adapt, with dark-mode variants when needed.
Check: List all icon elements: fail any PNG/JPG/GIF used as a control icon; for inline SVG, fail hard-coded fill/stroke hex values instead of currentColor or var(--token). Toggle dark theme and hover state and confirm icon color changes with the text.
Sources: Icons; Theming; Icons (Apple HIG); Dark Mode (Apple HIG)
motion: duration
minor · a tool can check it · #craft-motion-duration-range
Functional UI transitions (hover, press, toggle, dropdown, tooltip, modal, drawer) last between 100ms and 400ms and none exceeds 500ms.
Why and how to check
NN/g: about 100 ms for simple feedback, 200-300 ms for large screen changes, 100-400 ms as the general range, 400 ms only for big movements, and over 500 ms starts to feel like a drag; Material 3 tokens run from 50 ms (short1) to 1000 ms (extra-long4) with duration growing with the area or distance covered, so longer values exist for large transitions. Sources differ at the top end; the rule follows the stricter NN/g figure for functional UI and leaves narrative or hero motion to its own review.
Check: Collect computed transition-duration (+ delay) and animation-duration of interactive elements and overlays (button, a, input, nav, dialog, menu, tooltip, accordion). Fail durations under 100 ms (except color changes under 100 ms, review) or over 500 ms; flag 401-500 ms for review.
Sources: Executing UX Animations: Duration and Motion Characteristics; Material motion tokens (Material Components Android, Motion.md); Motion (Apple HIG)
motion: easing
minor · a tool can check it · #craft-motion-easing-direction
Elements entering the screen use a decelerating curve (ease-out or equivalent), elements leaving use an accelerating curve (ease-in), and linear easing is limited to non-spatial changes (opacity, color) and continuous or looping motion.
Why and how to check
NN/g: linear motion looks unnatural, ease-out is the usual default for entrances, ease-in suits exits, and exits should be slightly shorter than entrances (for example about 300 ms in, 200-250 ms out); Material 3 tokens pair standard-decelerate cubic-bezier(0, 0, 0, 1) with entering and standard-accelerate cubic-bezier(0.3, 0, 1, 1) with exiting motion, and keep a linear token for simple non-stylized motion.
Check: For transitions that animate transform, top/left or height, read transition-timing-function: fail 'linear' on those properties. For overlays, menus and drawers, check the open state uses ease-out/decelerate and the close state ease-in/accelerate and has a duration no longer than the open duration.
Sources: Executing UX Animations: Duration and Motion Characteristics; Material motion tokens (Material Components Android, Motion.md)
motion: purposeful motion
minor · a tool finds it, a person confirms · #craft-motion-frequent-quick
Animations on frequently repeated interactions (hover, press, toggle, menu open, tab change) are 200ms or shorter and never block input: the next click, key press or scroll works while the animation is running.
Why and how to check
NN/g: the more often users see an animation, the shorter and subtler it should be, with about 100 ms for simple feedback; Apple HIG: avoid adding motion to frequently used interactions, keep feedback brief, and let people cancel motion so they do not wait for it to finish. The 200 ms ceiling combines NN/g's 100 ms feedback and 200-300 ms large-change figures.
Check: Automated: collect durations of hover/active/focus/toggle transitions and fail above 200 ms. Manual/scripted: trigger an animation (open menu) and immediately click the next target, press Escape or scroll; fail if the input is ignored or delayed until the animation ends.
Sources: Executing UX Animations: Duration and Motion Characteristics; Motion (Apple HIG)
motion: reduced-motion handling
minor · a tool finds it, a person confirms · #craft-reduced-motion-replace
Under prefers-reduced-motion: reduce, large spatial motion (parallax, scroll-linked movement, large translate or scale, autoplaying loops) is removed or replaced with opacity or color changes, while essential and feedback animations remain in a gentler form.
Why and how to check
web.dev's accessibility module: limit scaling and movement, color changes and opacity or cross-fades are generally acceptable; its media-features module: keep reduced-motion alternatives free of the problematic movement and do not remove meaningful feedback entirely; web.dev's reduced-motion article lists parallax, autoplaying video, animated images and attention-grabbing motion as things to reduce and says essential movement is outside the preference; NN/g says to remove or minimize animations; Apple HIG says custom motion should respect accessibility settings. This refines the existing a11y-reduced-motion by saying what to do under the preference.
Check: Emulate reducedMotion: reduce. List computed transform/animation on elements that move more than about 20px or scale more than 5% (our review thresholds); fail if any remains. Confirm state changes (open, selected, error) still show a visible, non-spatial cue such as opacity or color. Confirm autoplay media and GIF/WebP loops are stopped or static.
Sources: Accessibility (course module); Media features; prefers-reduced-motion: Sometimes less movement is more; Executing UX Animations: Duration and Motion Characteristics; Motion (Apple HIG)
layout: balance
minor · a tool finds it, a person confirms · #craft-layout-balance
At desktop widths no section keeps all its content to one side and leaves a wide empty region next to it, unless the space is meant (a deliberate margin, a sticky rail, a pause between sections).
Why and how to check
Found when reviewing a real redesign: an accordion sat in a narrow left column at 1440 px and left the right side blank, which read as unfinished next to full-width sections. No source gives a number; the 40% content width and the 25%-of-width asymmetry limits are our own working thresholds, and the result is for a person to judge, not a fail.
Check: At the 1440 px viewport, for each top-level section at least 240 px tall, take the left and right edges of its text (measured with a text range, not the block width) and of its images, controls and media. Flag when the content covers less than 40% of the width and the empty gaps on the two sides differ by 25% of the width or more. A person decides whether the space is meant.
Sources: How I Design with AI (part 2)
layout: lead images
minor · a tool finds it, a person confirms · #craft-lead-image-visible
Images in the first screen are at least 40% visible at desktop and mobile widths: not covered by another element and not cut off by the screen edge, unless the layering is the design and enough of each image still shows to read it.
Why and how to check
Found when reviewing a real redesign: on a phone the hero's scatter of frames sat mostly behind the text plate, so the mobile page lost the personality the desktop page had. No source gives a number; the 40% visible limit is our own working threshold, and the result is for a person to judge, not a fail.
Check: At 1440x900 and 390x844 take each img, video and canvas of 80 px or more that starts in the first screen. Sample a 5 by 5 grid of points over its box and count the points where the element itself is on top (elementFromPoint) and inside the screen. Flag when fewer than 40% show. A person decides whether the cover is meant.
Sources: How I Design with AI (part 2)
Copy.
Plain, specific words, and the tells of copy that was written by a model and never edited. 36 rules.
plain language
major · a tool finds it, a person confirms · #copy-sentence-length
No sentence in body copy is longer than 25 words, and the average sentence length across a page's body copy is 20 words or fewer.
Why and how to check
GOV.UK says to split any sentence over 25 words; NN/g (Loranger, 2017) suggests 15 to 20 words at most. The two sources differ slightly, so the rule uses 25 as the hard cap and 20 as the average. NN/g also reports that users read only about 20 to 28% of the text on a page.
Check: Automated (apps/audit): extract the visible text of p, li, blockquote and dd elements (headings, navigation, forms, headers, footers and buttons left out), split on sentence terminators and line ends, count words per sentence. Needs 80 words and <html lang> starting with en, otherwise the result is manual. Fail when the mean exceeds 20 words or more than 25% of sentences are over 25 words; needs a person when only some sentences are over 25 (the splitter can join sentences around abbreviations); pass when none is. A person checks the long sentences it lists and any legal text.
Sources: Use clear language (GOV.UK content guidance); Plain Language Is for Everyone, Even Experts; Microsoft's brand voice: above all, simple and human
minor · a tool can check it · #copy-paragraph-length
Each body paragraph has one idea and at most 5 sentences; marketing and product pages aim for 3 sentences or fewer.
Why and how to check
GOV.UK sets 5 sentences as the maximum; NN/g (Loranger, 2017) says one idea per paragraph. The 3-sentence marketing target is our choice, not stated by either source.
Check: Count sentences in each p element. Fail above 5; warn above 3 on landing and product pages. Manual spot check that each paragraph carries one idea.
Sources: Use clear language (GOV.UK content guidance); Plain Language Is for Everyone, Even Experts
major · a tool finds it, a person confirms · #copy-reading-grade
Body copy for a general audience scores at or below US grade 9 on a readability formula (target grade 6 to 8); specialist audiences may go to grade 10 to 12 only when the audience is documented. For ESL, low-literacy or UK public-service audiences the target is lower: grade 5 or below, or CEFR B1 and GOV.UK-style plain language (our take).
Why and how to check
NN/g says aim for grades 6 to 8 for general audiences and 10 to 12 for experts, and that plain language helps experts too (2017 article; 2026 FAQ). WCAG 3.1.5 uses more than 9 years of school as the trigger for needing a summary. digital.gov says to write for the 8th grade only if the readers are 8th graders, so the target depends on audience. The lower target for ESL, low-literacy and UK public-service readers (grade 5 or below, CEFR B1, GOV.UK plain language) is our take: grades are a US schooling measure and a poor fit for a UK service, so the brief should name the audience's target and how it was checked. ESL group note: offer a translation or an interpreter where the service really can (translated pages linked with lang and hreflang, passages in another language marked with lang), avoid idioms and phrasal-verb puns, write dates and numbers in one unambiguous form, and put the action first.
Check: Automated (apps/audit): Flesch-Kincaid grade, with an English-only syllable heuristic, on the page's visible p, li, blockquote and dd text (headings, navigation, forms, headers, footers and buttons left out). It needs 80 words and <html lang> starting with en, otherwise the result is manual. Generic thresholds only: fail above grade 10, needs a person from 8 to 10, pass below 8. A pass never shows that a plain-language target (grade 5 or lower, CEFR B1, ESL or low-literacy readers) is met, and proper names and technical terms inflate the score. A person confirms the audience and the target the brief names, and scores the page by hand where the scorer did not run.
Sources: UX Writing: FAQs from Practitioners; Plain Language Is for Everyone, Even Experts; Product-Specific GenAI Needs to Write for the Web; Understanding SC 3.1.5 Reading Level (AAA); Plain language guide (principles, writing, design pages)
advisory · a tool finds it, a person confirms · #copy-reading-level-companion
Any text that needs more than lower secondary education (more than 9 years of school) to read has a plain-language summary or an easier version on the same page, or the page records the conscious exception.
Why and how to check
WCAG 2.2 SC 3.1.5 (AAA) requires a companion for text above lower secondary level; proper names and titles are excluded, and each language that is at least 5% of the content is assessed separately. AAA is not required for conformance, hence advisory.
Check: Compute grade per page as in copy-reading-grade. For any section above grade 9, confirm a summary at grade 9 or lower exists (technique G86) or an illustration or easier version (G103, G153).
minor · a tool finds it, a person confirms · #copy-plain-words
Copy uses short everyday words (for example 'buy', 'help', 'about') instead of 'purchase', 'assist', 'approximately', and avoids stacking nominalizations (-tion, -ment, -ance nouns) where a verb works.
Why and how to check
GOV.UK lists those exact swaps and notes that -ion and -ment words lengthen sentences; digital.gov calls hidden verbs a weakness and prefers the strongest verb form; Microsoft says to use short everyday words and prune every excess word.
Check: Grep body copy for a banned-word list (utilize, leverage, purchase, assist, approximately, commence, facilitate, in order to, prior to, subsequent to, endeavour, numerous) and flag sentences whose main verb is 'make/take/give/have/achieve + -tion noun'. Review flags manually.
Sources: Use clear language (GOV.UK content guidance); Plain language guide (principles, writing, design pages); Microsoft's brand voice: above all, simple and human; Google developer documentation style guide: Tone and content
minor · a tool finds it, a person confirms · #copy-active-voice
Headings, buttons, error messages and calls to action are written in the active voice, and body sentences use the passive only where the actor is unknown or unimportant.
Why and how to check
GOV.UK and digital.gov both prefer active voice; digital.gov says passive hides who does what and calls it one of the biggest problems in government writing; Mailchimp and Microsoft also say active. Neither GOV.UK nor digital.gov gives a percentage limit.
Check: Run a passive-voice detector (be-verb plus past participle) over all copy. Any hit in a heading, button, error or CTA fails; review body hits for a named actor.
Sources: Use clear language (GOV.UK content guidance); Plain language guide (principles, writing, design pages); Mailchimp Content Style Guide: Voice and Tone
major · a tool finds it, a person confirms · #copy-jargon-acronyms
Every acronym and specialist term that is not universally known is expanded or explained where it appears in headings, navigation and buttons, and at first use in body copy; branded feature names are always accompanied by a plain description.
Why and how to check
NN/g says to spell out acronyms on every use because readers scan and miss first-use definitions, and to treat branded labels as jargon; GOV.UK says explain at first use and write unfamiliar abbreviations out in titles. The sources differ on 'every use' versus 'first use'; the rule takes the stricter reading only for the places people scan.
Check: Extract all-caps tokens of 2 or more letters and capitalised product names from headings, nav and buttons; confirm each is expanded or described adjacent. Confirm first body use is expanded. Whitelist PDF, FAQ, URL.
Sources: UX Writing: FAQs from Practitioners; Use clear language (GOV.UK content guidance); Write clear titles; Plain Language Is for Everyone, Even Experts; Homepage design principles
headlines and value proposition
major · a person checks it · #copy-headline-swap-test
The H1 and hero subhead name a concrete outcome, offer or audience, and fail a swap test: a direct competitor could not publish the same text unchanged without lying; template headlines of the form 'The [adjective] [category] for modern [audience]' are not used.
Why and how to check
NN/g says to explain what the organisation does and why to choose it, and to replace generic welcome lines with specific taglines (2024). NN/g's headline guidance replaces 'Make better, faster decisions' with a named benefit (2015). Two practitioner write-ups (Superdesign 2026, Landing Doctor) independently flag the mad-lib hero and 'platform for teams' headlines as generic. Overlaps existing ux-tagline-specific, which does not give the swap test or the template pattern.
Check: Read the H1 and subhead alone. (1) Does it say what is offered and for whom? (2) Paste it with a competitor's name: does it still read true? If yes, fail. (3) Fail if it matches the template 'The [AI-powered|modern|all-in-one] [noun] for [teams|businesses|everyone]'.
Sources: Homepage design principles; Headings Are Pick-Up Lines: 5 Tips for Writing Headlines That Convert; Fix a Generic AI Landing Page: 6 Tells and the Prompts That Kill Them; Why AI landing pages convert worse
headings
major · a tool finds it, a person confirms · #copy-heading-informative
Each H2 and H3 describes the content beneath it in literal terms, with generic headings ('Introduction', 'Overview', 'Features', 'Our approach', 'Why choose us') replaced by topic-specific ones, and the list of headings read alone summarises the page.
Why and how to check
NN/g (FAQ 2026) recommends literal, informative headings that summarise what follows; GOV.UK adds that headings must be descriptive, front-loaded and removable without breaking sense, and says not to use questions as headings. Overlaps a11y-headings-labels and ux-scannable-text; this rule adds the generic-heading blocklist and the 'read headings alone' test.
Check: Extract the heading outline. Fail any heading matching the blocklist (case-insensitive: introduction, overview, features, benefits, our approach, why us, about, welcome). Then read the outline alone: a reader should be able to say what the page offers.
Sources: UX Writing: FAQs from Practitioners; Create a clear structure for your content; Mailchimp Content Style Guide: Web Elements; Plain language guide (principles, writing, design pages)
minor · a person checks it · #copy-heading-front-load
Headlines and headings make sense out of context (in a search result, a link list or a feed), put the key term in the first few words, and drop nonessential words.
Why and how to check
NN/g (Loranger, 2015) says headlines often appear alone, that readers focus on the first words, and that long headlines get truncated on small screens; GOV.UK says front-load headings. The article gives no study numbers. Overlaps ux-front-load, which covers the idea generally.
Check: For each H1 and H2, read it with no surrounding text. Fail if the topic cannot be named, if the key term is not within the first 3 words, or if removing a word changes nothing.
Sources: Headings Are Pick-Up Lines: 5 Tips for Writing Headlines That Convert; Create a clear structure for your content; Write clear titles
minor · a tool can check it · #copy-page-title-length
Each page's title element is unique, 65 characters or fewer, names what the page offers with the key term first, and meta descriptions are about 160 characters.
Why and how to check
GOV.UK sets 65 characters including spaces and says titles must be unique and use search terms; NN/g (2026) says page titles over 60 characters tend to be truncated and descriptions should be about 160 characters. The sources differ by 5 characters. Existing a11y-page-title only requires that a title exist and describe purpose.
Check: Crawl every page; assert title length is 65 or less, titles are unique across the site, and description length is 150 to 160.
Sources: Write clear titles; The 3 C's of Informational Microcopy
tone and voice
minor · a person checks it · #copy-no-idioms-puns
Headlines, navigation labels and calls to action contain no idioms, puns, slang, pop-culture references or metaphors that need decoding; any playful line sits next to a plain one that states the offer.
Why and how to check
NN/g (2015) says cute, faddish or idiomatic headlines sound inauthentic and confuse non-native speakers; Google's style guide says to avoid slang, figurative language and pop-culture references; Mailchimp says avoid slang. NN/g's 2026 microcopy article ranks 'character' last, after clarity and concision, so humour is allowed only after clarity.
Check: Read each headline, nav label and button aloud to someone who does not know the product. Fail where a literal reading misleads or the meaning needs a cultural reference. Confirm one plain statement of the offer is within the first screen.
Sources: Headings Are Pick-Up Lines: 5 Tips for Writing Headlines That Convert; Google developer documentation style guide: Tone and content; Mailchimp Content Style Guide: Voice and Tone; The 3 C's of Informational Microcopy
minor · a tool can check it · #copy-no-all-caps-text
Sentences, paragraphs, headings of more than 3 words and button labels are not set in all capitals, either in the source text or with text-transform: uppercase; short eyebrow labels of 3 words or fewer are the only exception.
Why and how to check
GOV.UK says no block capitals for large amounts of text because they are hard to read and can read as shouting; digital.gov says the same; UX Movement says all caps is loud and harder to read. The 3-word allowance is our threshold; the sources only say 'large amounts of text'.
Check: Flag text nodes in source with 4+ words entirely uppercase, and elements with computed text-transform: uppercase whose text has more than 3 words, or that are buttons.
Sources: Use the right tone; Plain language guide (principles, writing, design pages); 5 Rules for Choosing the Right Words on Button Labels
advisory · a tool can check it · #copy-sentence-case
Buttons, headings and navigation labels use sentence case consistently across the site.
Why and how to check
Mailchimp, GOV.UK (buttons), Microsoft and UX Movement all prescribe sentence case; Google says to match the page's existing capitalization. Mailchimp itself uses title case for form titles, so sources are not perfectly aligned. Wikipedia's AI-writing guide lists title-case section headings as a possible tell, but that is a single source about encyclopedia articles.
Check: For headings, buttons and nav items, flag strings where 3 or more words after the first start with a capital letter (excluding proper nouns); confirm the style is the same across all pages.
Sources: Mailchimp Content Style Guide: Web Elements; Button component; Microsoft's brand voice: above all, simple and human; 5 Rules for Choosing the Right Words on Button Labels; Google developer documentation style guide: UI elements
minor · a tool finds it, a person confirms · #copy-voice-consistent
The site has a written voice sheet (3 to 4 trait words, each with a do and a don't example), and copy sampled from hero, navigation, buttons, errors, footer and 404 page is consistent on person (you/we), contractions, spelling variant and formality.
Why and how to check
NN/g (2026), Mailchimp and Microsoft all say keep voice constant while tone varies; GOV.UK says address the reader as 'you'. Inconsistent copy is also a sign of copy pasted from different sources, which is a common failure of generated pages (hypothesis; no source measures it).
Check: Automated: count 'we/our' versus third-person ('users', 'customers') per page, contraction use, and US versus UK spelling pairs (color/colour); fail on mixed spelling. Manual: sample 6 surfaces and compare against the voice sheet.
Sources: UX Writing: FAQs from Practitioners; Mailchimp Content Style Guide: Voice and Tone; Microsoft's brand voice: above all, simple and human; Use the right tone
minor · a person checks it · #copy-tone-fits-context
Tone is adjusted to the reader's situation: payments, errors, legal, security and account-deletion text is neutral and literal, with no jokes, slang or hype.
Why and how to check
Mailchimp says tone should follow the reader's emotional state and not to force humour; NN/g (2023) says avoid humour in errors; Microsoft says adapt tone from serious to lighthearted to fit context; NN/g (2026) warns AI can approximate tone but cannot predict how it lands.
Check: List the copy on checkout, error, security, deletion, cancellation and legal surfaces; fail any joke, exclamation point or hype word.
Sources: Mailchimp Content Style Guide: Voice and Tone; Error-Message Guidelines; Microsoft's brand voice: above all, simple and human; UX Writing: FAQs from Practitioners
buttons and links
minor · a person checks it · #copy-cta-next-step
Each primary call to action is followed within the same viewport by a short line saying what happens next or what it costs (for example 'No card needed' or 'Takes 2 minutes'), and a repeated generic CTA is not used as filler.
Why and how to check
NN/g (2025) says CTAs should reduce uncertainty and give an easy next step. Landing Doctor lists the repeated decorative 'Get started' with no stated next step as a tell of generic page copy and recommends reassurance under the button. The second source is a vendor and gives no data, so treat the exact pattern as practice, not proof.
Check: For each primary CTA, check the adjacent text states the next step, cost or commitment truthfully. Fail if the same unexplained CTA label repeats in 3 or more sections.
Sources: The 3 I's of Microcopy; Why AI landing pages convert worse; No One Wants to "Get Started"
major · a tool can check it · #copy-link-text-unique-destination
Link text makes sense out of context and is brief; identical link text always goes to the same destination, different destinations never share the same text; the word 'link' and 'click here' are not in link text; external, new-window and file links say so, with file format and size for downloads.
Why and how to check
GOV.UK: do not reuse the same text for different pages, state when a link leaves the site, give file type and size (its example is 'PDF, 19.5KB'). WebAIM: omit 'link', keep it brief, make it understandable out of context, put 'opens in a new window' after the main text. Overlaps a11y-link-purpose (WCAG 2.4.4) and ux-descriptive-labels; this adds the duplicate-text, destination-notice and file-notice checks.
Check: Collect all anchors with accessible names. Fail if two different hrefs share the same name, if a name matches /^(click here|here|more|read more|link|this)$/i or contains 'link', if an href ends in a document extension without format and size in the name, or if target=_blank lacks a notice placed after the main text.
Sources: Write effective links; WebAIM: Hypertext Links; Mailchimp Content Style Guide: Web Elements
error messages and microcopy
major · a tool finds it, a person confirms · #copy-error-specific-actionable
Every error message names the field or problem and tells the user how to fix it in plain words; generic messages such as 'An error occurred', 'This field is required', 'Something went wrong' and raw error codes are not used without a cause and next step.
Why and how to check
GOV.UK says to say what happened and how to fix it, with the pattern 'Enter your first name' for empty fields and 'Name must be 35 characters or less' for wrong values, and lists 'This field is required' as a generic message to avoid. NN/g (2023) says be specific, offer remedies and hide error codes. Overlaps ux-form-errors and a11y-error-identification, which require specificity but give no wording patterns.
Check: Trigger every validation and failure state (empty, wrong format, too long, network failure, server error). Fail any message that does not name the problem and a fix, matches the generic list, or shows a raw code. Automated part: grep message strings for the generic list.
Sources: Error message component, "Be clear and concise"; Error-Message Guidelines; Validation pattern
minor · a tool can check it · #copy-error-no-blame-words
Error and warning text contains none of: invalid, illegal, incorrect, forbidden, prohibited, 'you forgot', 'please', 'sorry', 'oops', and uses no jokes; humour is allowed only for a rare complete system failure.
Why and how to check
GOV.UK lists these as words to avoid (it says 'please' implies a choice and 'sorry' does not help fix the problem); NN/g avoids 'invalid, illegal, incorrect', says humour wears thin on repeated errors, and allows novelty only for rare total failures. Mailchimp's guide also says forced humour can be worse than none. Sources agree.
Check: Grep all validation, toast and error-page strings for the banned list (case-insensitive) and for exclamation marks. Manually confirm that any jokes appear only on a whole-site failure page.
Sources: Error message component, "Be clear and concise"; Error-Message Guidelines; Mailchimp Content Style Guide: Voice and Tone
minor · a tool can check it · #copy-no-please-filler
Instructions and interface text do not use 'please', 'please note', 'simply', 'just', 'easy', 'quickly', 'it's that simple', or exclamation points.
Why and how to check
GOV.UK: no 'please' or 'please note'. Google: no 'please', no 'simply/easy/quickly/it's that simple', avoid exclamation points, cut filler such as 'please note' and 'at this time'. Two independent guides agree. Marketing headlines are out of scope; this is for instructions and UI text.
Check: Grep UI strings, form hints, instructions and help text for the banned list; exclude marketing headlines and testimonials.
Sources: Use the right tone; Google developer documentation style guide: Tone and content; Error message component, "Be clear and concise"
advisory · a tool can check it · #copy-microcopy-length
Interface microcopy (button and link labels, form hints, tooltips, error messages, empty states) is under 3 sentences, with the key information in the first words.
Why and how to check
NN/g defines microcopy as under 3 sentences and ranks clarity first, then concision, then character (2025 and 2026 articles); it warns that key details get truncated when copy is too long. Microsoft says to prune every excess word.
Check: Collect microcopy elements (labels, hints, tooltips, alerts, empty states) and count sentences; fail at 3 or more.
Sources: UX Copy Sizes: Long, Short, and Micro; The 3 C's of Informational Microcopy; Microsoft's brand voice: above all, simple and human
structure and scannability
minor · a person checks it · #copy-summary-first
Every article, guide or long section (more than roughly 300 words) opens with a one- or two-sentence summary that states the conclusion, and each section's first sentence carries its main point.
Why and how to check
GOV.UK says write a summary at the top and put the most important information first; digital.gov says summarize lengthy documents at the start; NN/g (2025, 2026) recommends the inverted pyramid and found AI answers often buried the overview in the conclusion. The 300-word threshold is our choice. Overlaps ux-front-load.
Check: For each long page or section, read only the first sentence: does it state the main point? Fail if the summary appears last or is missing.
Sources: Create a clear structure for your content; Plain language guide (principles, writing, design pages); UX Writing: FAQs from Practitioners; Product-Specific GenAI Needs to Write for the Web; Plain Language Is for Everyone, Even Experts
minor · a tool finds it, a person confirms · #copy-lists-for-parallel-items
Sequential steps are a numbered list and 3 or more parallel items are a bulleted list, not a comma-run sentence or a long paragraph.
Why and how to check
GOV.UK and digital.gov say to use bullets and numbered steps to reduce duplication and aid scanning; NN/g (2017) lists bullets and tables for scanning; NN/g (2025) found AI output often presented lists as plain paragraphs. The '3 or more' threshold is our choice. Overlaps ux-scannable-text.
Check: Flag paragraphs containing 3 or more comma-separated parallel items or the words 'first, second, third', and steps written in prose. Confirm real ul or ol elements are used.
Sources: Create a clear structure for your content; Plain language guide (principles, writing, design pages); Plain Language Is for Everyone, Even Experts; Product-Specific GenAI Needs to Write for the Web
specificity and credibility
major · a person checks it · #copy-proof-verifiable
Statistics, testimonials, client logos, awards and superlatives ('#1', 'trusted by 10,000 teams', 'increase conversions 300%') are either backed by a named, checkable source on the page or removed; nothing is invented to fill a layout.
Why and how to check
Landing Doctor lists fabricated or unsourced proof (invented testimonials, placeholder logos, precise unsourced stats) as a tell of generated pages; Wikipedia's guide lists vague attribution ('experts argue') as a pattern; NN/g (2026) says AI never certifies that a claim is correct, so a human must check. Mailchimp says to keep claims grounded. Overlaps ux-objective-language; this adds the verification step and the ban on invented proof. Plain facts that are given or agreed (a price, an hour) but not yet confirmed are not proof claims: show them in a prototype with the visible assumption marker and legend (guide protocol 3.2) and list them as open questions; this rule still bans invented customers, numbers, quotes and press, which are removed, not marked.
Check: List every number, quote, logo and superlative on the page. For each, record the source and confirm the customer or study exists and consented. Remove any with no source. Also grep for placeholder names (John Doe, Acme, Company Name).
Sources: Why AI landing pages convert worse; Wikipedia: Signs of AI writing; AI Can Help Write an Article, but It Can't Stand Behind It; Mailchimp Content Style Guide: Writing Goals and Principles
major · a tool can check it · #copy-no-placeholder-text
Shipped pages contain no lorem ipsum, bracketed placeholders ('[Company]', 'Your tagline here'), default template text, or chatbot residue such as 'Certainly!', 'I hope this helps', 'Would you like me to', or 'As an AI'.
Why and how to check
Wikipedia's guide lists chatbot leftovers ('Certainly!', 'I hope this helps') as a sign of pasted output. Landing Doctor lists placeholder logos as a tell. A published placeholder is a functional content failure. No source measures how often this happens. Not a placeholder: a real fact the owner has not confirmed yet, shown in a prototype with the visible, non-bracket assumption marker and legend from the guide protocol (3.2), listed in OPEN-QUESTIONS.md; it is a draft state, not a ship state, and the launch step refuses while any remain.
Check: Grep rendered text and alt attributes for: lorem ipsum, \[.*\], 'your (company|tagline|text) here', 'certainly!', 'I hope this helps', 'as an AI', 'let me know if', 'sample text', 'placeholder'.
Sources: Wikipedia: Signs of AI writing; Why AI landing pages convert worse
major · a tool finds it, a person confirms · #copy-specific-over-abstract
Feature and benefit copy replaces abstract adjectives (powerful, seamless, robust, intuitive, innovative, cutting-edge, next-generation, world-class) with a measurable change, a named workflow or a concrete example, and each feature block opens with a concrete detail rather than an icon plus a two-word title.
Why and how to check
NN/g (2015) turns 'Make better, faster decisions' into a named benefit and warns against hype; NN/g (2024) says avoid feature-driven wording and jargon and prefer concrete examples over umbrella terms. Landing Doctor, Superdesign and Microcopy Secrets independently recommend replacing vague adjectives and 'fits any product' benefit lines with specifics. Overlaps ux-real-content-examples and ux-objective-language.
Check: Grep for the abstract-adjective list. For each hit, require a number, a named object or a verifiable example in the same sentence, otherwise rewrite. Check that no three feature cards share the same structure of icon, two-word title and filler sentence.
Sources: Headings Are Pick-Up Lines: 5 Tips for Writing Headlines That Convert; Homepage design principles; Why AI landing pages convert worse; Fix a Generic AI Landing Page: 6 Tells and the Prompts That Kill Them; Avoid "landing page words" (Microcopy Secret #22)
minor · a tool can check it · #copy-no-buzzwords
Headlines, buttons and subheads contain none of the stock marketing verbs and metaphors (unlock, unleash, empower, supercharge, elevate, revolutionize, transform, harness, leverage, streamline, game-changer, journey, tapestry, landscape, realm, delve), and body copy has no more than 2 distinct ones per page.
Why and how to check
GOV.UK, Google and Microsoft separately tell writers to avoid buzzwords, jargon and figurative language; Microcopy Secrets (2024) and ContentBeta (2026) list unlock, unleash, empower, supercharge, game-changer as empty; Wisp (2024) and Wikipedia list tapestry, delve, realm, crucial, pivotal as overused. The thresholds (zero in headlines, 2 in body) are ours. Word lists drift by model era (Wikipedia) and single words are weak evidence (Wikipedia, TechCrunch), so this rule is about empty wording, not about detecting AI.
Check: Grep headings, buttons, hero text and meta descriptions for the list (case-insensitive, word-boundary); fail on any hit. Count distinct hits in body copy; fail above 2 per page.
Sources: Use clear language (GOV.UK content guidance); Google developer documentation style guide: Tone and content; Microsoft's brand voice: above all, simple and human; Avoid "landing page words" (Microcopy Secret #22); List of Words Overused by AI; How to Spot AI-Generated Content: A Guide to Identifying Common Phrases; Wikipedia: Signs of AI writing
minor · a person checks it · #copy-explicit-option-differences
Where a page offers plans, tiers or products, the differences between them are stated explicitly with units and precise labels (for example 'Up to 5 users' not 'Small team'), and all costs are shown in one place.
Why and how to check
NN/g (2024) says implicit differences lead to wrong choices, recommends explicit units and precise labels, and says to show all costs together. Only one source (NN/g, updating a 2013 Nielsen article), so this is not marked craft-consensus.
Check: For each comparison, confirm each plan label states what differs, numbers have units, and the total price including recurring fees can be read without leaving the page.
major · a person checks it · #copy-expert-review-before-launch
Copy that gives legal, medical, financial or safety advice (anything where a wrong or out-of-date sentence could harm the reader) is reviewed by a qualified person before launch; until then each such passage carries a visible draft marker with a legend on the page, and the page does not ship it as final.
Why and how to check
This is our judgement, often called 'your money or your life' (YMYL) content in search-quality writing; no source was read for this rule. NN/g (2026) says a human must check that AI-written claims are correct, which supports the principle for generated copy but is not specific to advice. A model cannot hold the licence, the jurisdiction or the current law that the advice depends on, so a draft written by a model is a draft. The marker is the same visible, non-bracket convention as for unconfirmed facts (guide protocol 3.2), so the draft is never presented as checked. Severity is major, not blocker, because an our-take rule cannot be a blocker; realize-launch applies it as a launch gate.
Check: List every passage that advises on law, health, money or safety (rights, deadlines, dosages, eligibility, 'you should', 'you must'). For each, confirm that a named qualified reviewer and a date are recorded (design/OPEN-QUESTIONS.md or the QA report), or that the passage carries the visible draft marker and its legend and the launch gate is still closed. Search the built page for the marker before launch.
Sources: AI Can Help Write an Article, but It Can't Stand Behind It
AI copy tells
advisory · a tool finds it, a person confirms · #copy-no-stock-rhetoric
A page uses at most one negative-parallelism construction ('not X, it's Y', 'not only... but also', 'No X. No Y. Just Z.'), at most one slogan of three parallel short items ('Save time. Work smarter. Scale faster.'), and no self-answered rhetorical question openers ('The result?').
Why and how to check
Corroborated by two independent sources: Wikipedia's guide lists negative parallelism and the rule of three; ContentBeta (2026) lists 'It's not about X, it's about Y', 'No X. No Y. Just Z.' and 'The result?'. RAXXO and Superdesign add the triad ('works with any product'). Wikipedia warns that humans also write these, so frequency, not presence, is the signal. The 'one per page' limit is ours.
Check: Regex for /not (just|only|about)? .{3,40}[,;] (it's|but|it is)/i, /\bno \w+\. no \w+\. just/i, sentences of the form 'X. Y. Z.' with each under 4 words, and short-question-then-answer lines (looked for only in headings, paragraphs, list items and quotes, not in form labels, legends or buttons, where a question with a hint such as '(optional)' is not rhetoric). Count hits per page; fail above 1 of each kind.
Sources: Wikipedia: Signs of AI writing; List of Words Overused by AI; Why your AI-generated content looks like AI and how to fix it; Fix a Generic AI Landing Page: 6 Tells and the Prompts That Kill Them
advisory · a tool can check it · #copy-em-dash-restraint
Em dashes are absent from headings, buttons and labels and appear no more than about once per 300 words of body copy.
Why and how to check
Three sources report em-dash overuse in AI drafts: Wikipedia's guide (heavy use in place of commas, parentheses or colons, formulaic punch-up), Bullock (her own count: about 1 per 130 words in 24 AI-first drafts against about 1 per 850 words in 15 human pieces), and Monteiro (asserted, no data). All are anecdotal or small, Wikipedia says the sign works only with others, and Wikipedia also notes newer models reportedly avoid em dashes. The 300-word threshold is ours.
Check: Count U+2014 in visible text; fail any in headings, buttons, nav or labels; fail if body density exceeds 1 per 300 words.
Sources: Wikipedia: Signs of AI writing; Does AI writing use a lot of dashes?; Why the F does AI always use em dashes
minor · a tool can check it · #copy-no-filler-openers
Copy does not open or close with filler framing such as 'In today's fast-paced world', 'In the realm of', 'It is important to note', 'Whether you're a X or a Y', 'In conclusion', 'stands as a testament to', or an inspirational closing line.
Why and how to check
Wisp (2024) lists 'In the realm of' and 'It is important to understand'; ContentBeta (2026) lists 'in today's world' and 'digital age'; Wikipedia lists 'stands as a testament', 'pivotal', canned 'It's important to note' (2023 to 2024) and 'In conclusion' summaries (about 2023, since less frequent); RAXXO lists inspirational closings. Google says to cut filler like 'please note'. Corroborated by at least three, but the exact phrases date quickly.
Check: Grep for the phrase list at the start and end of paragraphs and sections; fail on any hit. Manually review any section whose final sentence only restates the heading.
Sources: How to Spot AI-Generated Content: A Guide to Identifying Common Phrases; List of Words Overused by AI; Wikipedia: Signs of AI writing; Why your AI-generated content looks like AI and how to fix it; Google developer documentation style guide: Tone and content
minor · a person checks it · #copy-section-content-drives-structure
The section order and number of feature blocks on a page follow the audience's decisions and objections (documented in the brief), not the default sequence hero, logo bar, three feature cards, testimonial, pricing, FAQ, CTA, and feature cards do not carry emoji as icons.
Why and how to check
Superdesign (2026) and Landing Doctor independently name the stock section order and three identical feature cards as tells; Superdesign also names emoji or thin-line icons on cards, and Wikipedia lists emoji at the start of headings or list entries as a possible sign. Landing Doctor recommends answering the buyer's top reasons not to convert near each CTA. No data is given by either vendor.
Check: Write the page's section order and the question each section answers. Fail if the order matches the stock sequence with no documented reason, if three or more cards share identical structure, or if emoji are used as feature icons or heading prefixes.
Sources: Fix a Generic AI Landing Page: 6 Tells and the Prompts That Kill Them; Why AI landing pages convert worse; Wikipedia: Signs of AI writing
copy and accessibility
minor · a tool can check it · #copy-alt-text-writing
Alt text states the image's purpose in context in about 150 characters or fewer, starts with the key words, does not begin with 'image of' or 'photo of', does not repeat adjacent text, does not contain keyword lists, and is empty for decorative images.
Why and how to check
NN/g (2024) gives about 150 characters, purpose from context, no leading 'image', empty alt for decorative images and says AI is unreliable at writing alt text from scratch; WebAIM (2021) says brief, no 'image of', no keyword stuffing, no repeat of nearby text, and gives no numeric limit. Overlaps a11y-alt-text; this adds length and wording checks.
Check: For every img with non-empty alt: fail if length exceeds about 150 characters, starts with /^(image|photo|picture|graphic) of/i, equals the text of a nearby caption or heading, or contains 3 or more comma-separated keywords. Review decorative images for alt="".
Components.
How common parts must behave: tabs, accordions, dialogs, menus, forms, carousels and more. 59 rules.
Tabs
major · a tool finds it, a person confirms · #comp-tabs-semantics
A tabbed interface uses role=tablist containing role=tab elements, each tab has aria-selected (true on exactly one) and aria-controls pointing at a role=tabpanel that has aria-labelledby pointing back to its tab, and the tablist has an accessible name.
Why and how to check
The APG tabs pattern requires these roles and properties so assistive technology announces the set, the selected tab and the panel relationship.
Check: Query the DOM: every [role=tab] is a child of [role=tablist]; exactly one tab has aria-selected=true; each tab aria-controls resolves to a [role=tabpanel] whose aria-labelledby resolves to that tab; tablist has aria-label or aria-labelledby. Confirm with a screen reader that the selected state is announced.
Sources: Tabs Pattern; How to Meet WCAG 2.2 (Quick Reference)
major · a person checks it · #comp-tabs-keyboard
In a tab list only the active tab is in the Tab order (others tabindex=-1), Left/Right arrows move focus between tabs with wrap, Home and End jump to first and last, and a tabpanel whose first content is not focusable has tabindex=0.
Why and how to check
The APG gives this as the keyboard model; the single tab stop keeps long tab lists from trapping keyboard users, and the focusable panel lets Tab reach the content.
Check: Press Tab into the tablist: focus lands on the active tab. Press Right/Left/Home/End and verify focus moves; press Tab and verify focus lands in the panel (or its first focusable element). Check inactive tabs have tabindex=-1.
Sources: Tabs Pattern; Tabbed Interfaces; How to Meet WCAG 2.2 (Quick Reference)
Accordion / collapsible sections
major · a tool can check it · #comp-accordion-markup
Each accordion header is a native button that is the only child of a heading element, with aria-expanded reflecting the panel state, and the button's accessible name is the visible title.
Why and how to check
APG accordion and Inclusive Components agree: put the button inside the heading so the heading keeps its structure role; do not give the heading role=button. Progressive enhancement: when content must also work without JavaScript (perf-core-first, a11y-native-first), write plain heading + div markup (the content is visible) and let the script insert the button inside the heading and set the hidden state; this is the pattern the 2026-11 pilot used. Native details/summary is an acceptable no-JS alternative where the design allows it.
Check: For each toggle: tagName is BUTTON, parent is h1-h6 (or role=heading with aria-level), the button is the heading's only child, aria-expanded is 'true' or 'false' and flips on click, the hidden panel is display:none or [hidden] when collapsed.
Sources: Accordion Pattern; Collapsible Sections; Disclosure (Show/Hide) Pattern
minor · a tool can check it · #comp-accordion-region-limit
Accordion panels carry role=region with aria-labelledby only when the accordion has fewer than about six panels; larger accordions leave panels unlabeled as regions.
Why and how to check
APG warns that region on many panels (it cites roughly six or more) creates landmark clutter. The number is approximate in the source.
Check: Count [role=region] / named section elements that are accordion panels on the page; flag if 6 or more per accordion.
Sources: Accordion Pattern; Landmark Regions (practice)
minor · a tool finds it, a person confirms · #comp-disclosure-state-not-color-only
The open or closed state of a disclosure control is shown by a shape or icon change (such as plus to minus or a rotating chevron) that is hidden from assistive technology, never by color alone, and the icon uses currentColor.
Why and how to check
Inclusive Components hides the decorative icon with aria-hidden because state is carried by aria-expanded, and uses currentColor so high-contrast modes recolor it. COGA 4.3.2 asks for consistent, clear indicators for expandable content. Primarily one practitioner plus COGA Note.
Check: Toggle each control and compare screenshots: the indicator shape changes. Inspect the icon SVG for aria-hidden=true and fill/stroke currentColor. Emulate forced-colors and confirm it stays visible.
Sources: Collapsible Sections; Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Modal dialog
major · a tool can check it · #comp-dialog-semantics
A modal dialog has role=dialog (or alertdialog), aria-modal=true, and an accessible name from aria-labelledby pointing at its visible title (or aria-label), and every control needed to operate it is inside the dialog element.
Why and how to check
The APG makes the label required and the modal flag part of the pattern. Prefer the native dialog element with showModal(), which supplies modality; the APG page itself does not discuss inert or the native element.
Check: Open each dialog; query [role=dialog],[role=alertdialog],dialog[open]; assert accessible name is non-empty and that no interactive control needed by the dialog sits outside it.
Sources: Dialog (Modal) Pattern; Alert Dialog Pattern; How to Meet WCAG 2.2 (Quick Reference)
blocker · a person checks it · #comp-dialog-focus
Opening a modal dialog moves focus into it, Tab and Shift+Tab cycle only within it, Escape closes it, a visible close or Cancel button exists, and closing returns focus to the control that opened it (or to a logical next target if that control is gone).
Why and how to check
These are the APG keyboard and focus rules for modal dialogs; the close button is 'strongly recommended' there. Initial focus guidance: first focusable by default; a static start element for long or structured content; least destructive action for irreversible steps.
Check: Keyboard only: activate the opener, verify focus is inside; Tab through the whole dialog and confirm wrap; press Escape and confirm close and focus on the opener; confirm a visible close control exists.
Sources: Dialog (Modal) Pattern; How to Meet WCAG 2.2 (Quick Reference)
major · a person checks it · #comp-dialog-truly-modal
aria-modal=true is set only when code actually blocks interaction with the page behind the dialog and the page behind is visually obscured; otherwise the surface is built as a non-modal popover.
Why and how to check
The APG says aria-modal can hide outside content from some assistive technology users, so a dialog marked modal that behaves non-modally seriously harms them.
Check: With the dialog open, try to Tab, click and screen-reader-navigate to background content. If any works, aria-modal must be absent. If aria-modal is present, confirm a dim or blur overlay covers the page.
Sources: Dialog (Modal) Pattern; Alert Dialog Pattern
Alert dialog
major · a person checks it · #comp-alertdialog-confirmations
Confirmations of destructive or irreversible actions use role=alertdialog with aria-describedby pointing at the message, and initial focus goes to the least destructive control (Cancel) when undoing is hard.
Why and how to check
APG alertdialog is for interrupting the workflow to get a response and requires aria-describedby on the message; the modal-dialog page advises focusing the least destructive action for irreversible steps. COGA 4.8.3 asks that consequences of actions are stated.
Check: Trigger each delete/purchase/discard confirmation: verify the role, aria-describedby target text, and that focus starts on Cancel (not the destructive button).
Sources: Alert Dialog Pattern; Dialog (Modal) Pattern; Making Content Usable for People with Cognitive and Learning Disabilities
Alert / notifications
major · a person checks it · #comp-alert-no-focus-no-autodismiss
Alerts and notifications do not move keyboard focus, are not removed by a timer shorter than the user needs to read them, and error or action-required messages stay until the user resolves or dismisses them.
Why and how to check
The APG alert page says alerts must not take focus and should not vanish automatically (it ties this to time-limit criteria). WCAG 2.2.1 already requires adjustable timing.
Check: Trigger each alert; confirm document.activeElement does not change. Time how long it stays; any auto-dismiss under about 10 seconds or on an error is a finding unless the user can extend or recover the message elsewhere.
Sources: Alert Pattern; Notifications; How to Meet WCAG 2.2 (Quick Reference)
major · a tool finds it, a person confirms · #comp-live-region-exists-before-content
Live regions (role=status, role=alert or aria-live) exist in the DOM before their message is inserted, and messages that must be announced are injected after page load rather than rendered in the initial HTML.
Why and how to check
APG says alerts already present at page load are not announced; Inclusive Components says do not add the region and its content at the same moment, leave a short gap.
Check: Inspect the initial DOM for an empty live-region container; trigger the message and confirm a screen reader announces it (NVDA, VoiceOver). Flag a region created together with its text.
Sources: Alert Pattern; Notifications; Understanding SC 4.1.3 Status Messages
minor · a person checks it · #comp-notification-recoverable
Any transient notification (toast, snackbar) refers to something the user can find again elsewhere (list, history, page state), and notifications are not used for information needed to complete a task.
Why and how to check
Inclusive Components advises that each notification points to something discoverable elsewhere so people who missed it are not stranded. Sources disagree on dismiss buttons and auto-dismiss (see disagreements); this rule takes the safe common ground. Single practitioner source.
Check: For each toast: after it disappears, can the same information be found on the page (cart count, order list, inline status)? If not, flag.
Sources: Notifications; Alert Pattern
Status messages
major · a tool finds it, a person confirms · #a11y-status-role-choice
Non-blocking status messages (results count, success, progress, cart updated) use role=status or an equivalent polite live region; role=alert or aria-live=assertive is reserved for important, time-sensitive messages such as errors; none of them take focus.
Why and how to check
Understanding 4.1.3 maps success/results to role=status (ARIA22) and errors to role=alert (ARIA19), and lists assertive use on unimportant content as a failure. It adds that over-announcing makes apps chatty.
Check: List every dynamic message on the page. Assert each has a live-region role or property; assert role=alert / assertive appears only on errors or urgent warnings; test with a screen reader that it is announced once.
Sources: Understanding SC 4.1.3 Status Messages; How to Meet WCAG 2.2 (Quick Reference); Alert Pattern
Breadcrumb
Tooltips and toggletips
major · a tool finds it, a person confirms · #comp-tooltip-not-sole-label-no-interactive
A tooltip appears on both hover and keyboard focus, closes with Escape, stays open while the pointer is over it, is linked to its trigger with aria-describedby, and contains no links, buttons or other interactive content; where space allows, a visible text label is used instead.
Why and how to check
APG: Escape dismisses, focus stays on trigger, hover persistence, aria-describedby (the page says the pattern has no consensus yet). Inclusive Components: tooltips are a last resort and must not hold interactive content, which belongs in a dialog or menu. WCAG 1.4.13 covers dismiss, hover and persist.
Check: Hover and Tab to each tooltip trigger; press Escape; move the pointer onto the bubble; check aria-describedby; check the bubble DOM for a, button, input.
Sources: Tooltip Pattern; Tooltips & Toggletips; How to Meet WCAG 2.2 (Quick Reference)
minor · a person checks it · #comp-toggletip-click-live-region
Help bubbles that must work on touch (info icons) are click-activated 'toggletips': a button with type=button and a visible or visually hidden name, an adjacent role=status region filled on click, no aria-describedby on the button, and dismissal by Escape, outside click or blur.
Why and how to check
Inclusive Components: hover and focus tooltips do not work reliably on touch; describedby would read the content before the user asks. This differs from the APG tooltip, which has no click variant (single practitioner source, so hypothesis).
Check: Tap and keyboard-activate each info button: bubble appears, screen reader announces it, Escape closes. Check the button is a real button with an accessible name and that the bubble text is not exposed before activation.
Sources: Tooltips & Toggletips
minor · a tool can check it · #comp-no-title-attribute-reliance
The title attribute is not the only source of an accessible name, instruction or tooltip text; essential information is also available as visible text or a labelled control.
Why and how to check
Inclusive Components says title fails for keyboard and touch users and has inconsistent screen reader support; COGA 4.5.6 says labels should not depend only on title or ARIA attributes.
Check: Search for [title] on interactive elements; for each, assert there is also visible text or aria-label/labelledby and that no instruction exists only in title.
Sources: Tooltips & Toggletips; Making Content Usable for People with Cognitive and Learning Disabilities
Menus
Links
Listbox / custom select
major · a tool finds it, a person confirms · #comp-listbox-select-semantics
A custom listbox has role=listbox with an accessible name, role=option children each with aria-selected (or aria-checked, never both), aria-multiselectable=true when multi, short option names, and Up/Down, Home/End and type-ahead keys; options contain no interactive controls.
Why and how to check
APG listbox: names are read as a flat string, long or repetitive names slow screen reader users, interactive content in options is unsupported. Where a native select would do, a11y-native-first applies.
Check: Inspect roles and states; keyboard test Up/Down/Home/End and typing a letter; scan options for nested a/button/input.
Sources: Listbox Pattern; How to Meet WCAG 2.2 (Quick Reference)
Combobox / autocomplete
major · a tool can check it · #comp-combobox-semantics
An autocomplete input has role=combobox on the input with aria-expanded, aria-controls pointing at the popup, aria-autocomplete matching its behavior, aria-activedescendant naming the highlighted option, and a real label.
Why and how to check
APG combobox requires these relationships; DOM focus stays on the input while virtual focus moves in the popup (except for dialog popups).
Check: Open the popup and assert on the input: role=combobox, aria-expanded=true, aria-controls resolves to a visible listbox, aria-activedescendant changes with Down arrow and resolves to an option; label via label/aria-labelledby.
Sources: Combobox Pattern; How to Meet WCAG 2.2 (Quick Reference)
major · a person checks it · #comp-combobox-keyboard
In a combobox, Down Arrow opens the popup and moves into it, Enter accepts the highlighted option, Escape closes the popup (restoring the typed value), typing keeps editing the input, and Tab moves on without stopping inside the popup.
Why and how to check
APG lists these keys; Escape restoring the previous value is named as a distinction from listbox and menu button.
Check: Type, press Down, Enter, Escape, Tab in order and verify each outcome and that focus never leaves the input.
Sources: Combobox Pattern; How to Meet WCAG 2.2 (Quick Reference)
Switches and toggles
major · a tool can check it · #comp-switch-label-stable
An on/off control is a native checkbox, a button with aria-pressed, or role=switch with aria-checked, has a fixed accessible name, toggles with Space, and its name does not change when its state changes.
Why and how to check
APG switch calls a stable label critical; Inclusive Components says never change the label and the state together (voice users say the visible label), and explicit aria-pressed true/false. Support for role=switch varied in 2017 (ChromeVox, NVDA) per Inclusive Components; check current support.
Check: Toggle each control and assert its accessible name is identical before and after, aria-checked/aria-pressed/checked is explicit 'true' or 'false', and Space toggles it.
Sources: Switch Pattern; Toggle Buttons; Checkbox Pattern
major · a tool finds it, a person confirms · #comp-toggle-visual-state-not-color
A toggle's on/off state is visible through more than color (position of a knob, an icon, or text) and the visual style is driven by the ARIA state attribute so they cannot diverge.
Why and how to check
Inclusive Components cites WCAG 1.4.1 and styling from [aria-pressed=true]; COGA 4.5.10 says toggles should expose state programmatically and every state change needs visual feedback.
Check: Grayscale screenshot of both states; confirm still distinguishable. Check CSS selectors key off aria-pressed/aria-checked/:checked.
Sources: Toggle Buttons; Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Theme switcher
advisory · a person checks it · #comp-theme-switcher-control
If the site offers a light/dark theme control, it is a labelled toggle (button with aria-pressed, or radio group in a fieldset with legend), lives in a landmark (header or footer) in the same place on every page, and the choice persists across visits.
Why and how to check
Inclusive Components (2017) describes a toggle button, placement in a landmark, same location everywhere and localStorage persistence; it does not cover prefers-color-scheme, so default to the user's OS setting is not from this source. Single practitioner source.
Check: Toggle, reload, confirm persistence; confirm control is within header/footer landmark and identical on two pages; check name and state announced.
Sources: A Theme Switcher; Toggle Buttons; prefers-color-scheme
Checkboxes and radios
major · a tool can check it · #comp-checkbox-radio-group-labels
Related checkboxes or radios are wrapped in a fieldset with a legend (or role=group / radiogroup with aria-labelledby) that names the group, each input has its own label, and a mixed-state parent checkbox uses aria-checked=mixed.
Why and how to check
APG checkbox, radio and switch pages require a group label for visible group labels; the radiogroup must be labeled; mixed is the third state for tri-state checkboxes.
Check: For each set of inputs sharing a name or visual group, assert a fieldset>legend or [role=group|radiogroup][aria-labelledby|aria-label]; every input has a label.
Sources: Checkbox Pattern; Radio Group Pattern; Switch Pattern; How to Meet WCAG 2.2 (Quick Reference)
major · a person checks it · #comp-radio-keyboard-model
A radio group is one Tab stop (landing on the checked radio, else the first) with arrow keys moving and checking the next/previous radio with wrap; use native input type=radio with a shared name so the browser provides this.
Why and how to check
APG radio keyboard model; the page notes some browsers differ on Shift+Tab landing when none is checked.
Check: Tab into the group, press arrows, confirm focus and checked state move together and the group is one tab stop.
Sources: Radio Group Pattern; How to Meet WCAG 2.2 (Quick Reference)
Slider
major · a tool finds it, a person confirms · #comp-slider-semantics
A slider is a native input type=range or an element with role=slider carrying aria-valuenow, aria-valuemin, aria-valuemax and an accessible name, uses aria-valuetext when the number is not self-explanatory (a price, a time), and responds to arrow keys, Home and End.
Why and how to check
APG slider requirements; it also warns touch-based assistive technology may not support slider gestures well, so test on touch screen readers.
Check: Inspect value attributes; keyboard test arrows/Home/End/PageUp/PageDown; check aria-valuetext shows units; test with VoiceOver on iOS or TalkBack.
Sources: Slider Pattern; How to Meet WCAG 2.2 (Quick Reference)
Toolbar
minor · a tool finds it, a person confirms · #comp-toolbar-threshold
role=toolbar is used only for a group of 3 or more controls, has an accessible name, exposes a single tab stop with arrow-key navigation among controls, and contains at most one control that needs the same arrow keys (placed last).
Why and how to check
APG toolbar states the 3-control threshold and the single-tab-stop roving model; otherwise a plain group of buttons with normal Tab order is simpler.
Check: Find [role=toolbar]; count controls (3 or more); Tab once into it and use arrows; check name.
Sources: Toolbar Pattern; How to Meet WCAG 2.2 (Quick Reference)
Landmarks
major · a tool can check it · #comp-landmarks-complete
The page has exactly one main and one top-level banner (header) and one contentinfo (footer), uses nav, aside and search elements for those roles, and all perceivable content is inside some landmark.
Why and how to check
APG landmarks best practice says all content belongs in a landmark; banner, main and contentinfo are one per page and top-level. header and footer only count as banner/contentinfo in the body context.
Check: Query landmarks with axe rules landmark-one-main, landmark-no-duplicate-banner, landmark-no-duplicate-contentinfo, region (content outside landmarks).
Sources: Landmarks Pattern; Landmark Regions (practice); How to Meet WCAG 2.2 (Quick Reference); axe-core rule descriptions
minor · a tool can check it · #comp-landmarks-labels
Landmarks of the same role each have a unique accessible name (except identical-content pairs), section and form elements become landmarks only when named, labels do not repeat the role word, and the page has about seven or fewer landmark regions.
Why and how to check
APG landmark regions: unique labels, name required for region and form, 'Site' not 'Site navigation'; the landmarks page gives a seven-or-fewer rule of thumb because value drops as the count grows.
Check: List landmarks with names; flag duplicates of the same role with the same or missing name; flag labels containing their own role word; count landmarks.
Sources: Landmark Regions (practice); Landmarks Pattern
Cards
minor · a tool can check it · #comp-card-single-primary-link
Each card is one list item with a heading at the same level across the set; the heading text is the only primary link (not a link wrapping the whole card); a whole-card click area, if wanted, is made with a stretched ::after on that link; images that are decorative have alt="".
Why and how to check
Inclusive Components: whole-card links make screen reader announcements verbose and break with nested links; list markup gives item counts; the heading link has a unique descriptive name. Trade-off: the stretched-link technique prevents text selection, mitigated by a click-duration check.
Check: For each card: parent is li inside ul/ol; one heading; exactly one a whose text matches the heading; no a ancestor wrapping heading+image+text; images have alt (empty if decorative).
Sources: Cards (Inclusive Components)
advisory · a tool can check it · #comp-card-limited-tab-stops
A card has one tab stop, or two only when a secondary link is genuinely useful, and 'read more' cues are non-interactive text hidden from assistive technology (aria-hidden) and tied to the link with aria-describedby rather than being a second link.
Why and how to check
Inclusive Components: extra tab stops slow keyboard users and cards should not be miniature web pages; a visual 'read more' cue helps people recognize interactivity.
Check: Count focusable elements per card (max 2); flag 'read more' / 'learn more' links duplicated per card.
Sources: Cards (Inclusive Components)
Data tables
major · a tool can check it · #comp-data-table-semantics
Tabular data uses table with a caption (or a figure and figcaption), th with scope=col / scope=row, and no layout tables; the grid role is used only for interactive cell-by-cell widgets.
Why and how to check
Inclusive Components lists these basics; WCAG 1.3.1 requires the structure to be in the markup, which the existing a11y-semantic-structure rule covers.
Check: Each table has caption or figcaption/heading, header cells are th with scope, no table used purely for layout (role=presentation only as legacy fallback).
Sources: Data Tables; How to Meet WCAG 2.2 (Quick Reference)
major · a tool can check it · #comp-data-table-overflow-focusable
A table wider than its container scrolls inside a labelled container (aria-labelledby the caption) that is keyboard focusable (tabindex=0) only when it actually overflows; table display properties are not changed to reflow it.
Why and how to check
Inclusive Components: keyboard users cannot scroll unfocusable scrollers; tabindex=0 only on overflow avoids useless tab stops; changing display breaks table semantics. Inclusive Components uses role=group for the container.
Check: At 320 px and 1280 px: for each table container with scrollWidth > clientWidth assert tabindex=0 and an accessible name; for non-overflowing containers assert no tabindex.
Sources: Data Tables; Keyboard focusable scrollers; axe rule: Scrollable region must have keyboard access; How to Meet WCAG 2.2 (Quick Reference)
minor · a tool can check it · #comp-data-table-sortable-headers
Sortable column headers contain a button labelled with the column name, and the th carries aria-sort (ascending, descending or none); sort arrows use currentColor.
Why and how to check
Inclusive Components: button in each sortable header, aria-sort for state, and clear labels; icons that show current state can mislead about what pressing does, so the button text must say what happens.
Check: For each sortable column: th[aria-sort] contains button with name; activating it changes aria-sort and row order.
Sources: Data Tables
Content slider / carousel
major · a tool finds it, a person confirms · #comp-content-slider-no-autoplay
A horizontal slider uses native scrolling with a focusable (tabindex=0) labelled region, list markup for slides, labelled previous/next buttons, does not autoplay or move without user action, and slides that are out of view have their links removed from the tab order.
Why and how to check
Inclusive Components says a slider should never move without the user's say-so and that WCAG 2.2.2 is a minimum; it also recommends native overflow scrolling, a region with aria-label and tabindex=-1 on links in hidden slides. APG carousel (existing source) covers rotating carousels.
Check: Load the page for 30 seconds without input: nothing moves. Tab through: arrows reachable, scroller focusable with visible focus, hidden slides' links not focusable. Check prev/next buttons have names.
Sources: A Content Slider; Carousel Pattern; How to Meet WCAG 2.2 (Quick Reference); Designing Effective Carousels
Icons in components
Navigation
blocker · a person checks it · #a11y-multiple-ways
Each page in a set of pages (other than steps or results of a process) can be reached by at least two of: links from related pages, table of contents, site map, search, a list of links to all pages, or links to all pages from the home page.
Why and how to check
WCAG 2.4.5 (AA). The Understanding page says very small sites can satisfy this with links from the home page. Example pairings: search plus category list.
Check: Pick 3 inner pages and show two independent routes to each (nav + search, nav + sitemap, home page links + previous/next). Process pages (checkout steps, search results) are exempt.
Sources: Understanding SC 2.4.5 Multiple Ways; How to Meet WCAG 2.2 (Quick Reference)
Media
blocker · a person checks it · #a11y-captions-live
Live audio in synchronized media (live stream, webinar, event broadcast) has captions that include speech, speaker identification and significant non-speech sounds.
Why and how to check
WCAG 1.2.4 (AA). It targets broadcasts, not two-way calls in web apps; live means captured from a real event with at most a broadcast delay; fully generated content is not live.
Check: If the page embeds a live stream, enable captions in the player during the live event and confirm they track speech; otherwise mark not applicable.
Sources: Understanding SC 1.2.4 Captions (Live); How to Meet WCAG 2.2 (Quick Reference)
Readable content
advisory · a tool finds it, a person confirms · #a11y-reading-level-support
Where text needs more than lower-secondary reading ability (roughly school years 7 to 9), a plain summary, illustrations or a simpler version is offered.
Why and how to check
WCAG 3.1.5 is Level AAA (not required for AA conformance); the Understanding page lists a text summary, illustrations, spoken version and simplified text as techniques. The EAA Annex I sets B2 (CEFR) for consumer banking information specifically.
Check: Run a readability score (for example Flesch-Kincaid) on main content, ignoring proper names; if above grade 9, check for a summary or simpler version.
Sources: Understanding SC 3.1.5 Reading Level (AAA); Directive (EU) 2019/882 on accessibility requirements for products and services; Making Content Usable for People with Cognitive and Learning Disabilities
Cognitive accessibility: finding things
minor · a person checks it · #a11y-coga-key-tasks-prominent
The one to three key tasks of the site are visually distinct, placed near the top of the page, each under a descriptive heading, and reachable from the top level of the main navigation.
Why and how to check
COGA patterns 4.3.1 and 4.3.4 ask for key tasks in visually distinct boxes near the top, in top-level navigation and early in the reading order, visible without scrolling or hovering. Overlaps ux-primary-cta-above-fold.
Check: At 1280x800 and 375x812 without scrolling, identify the key task controls; check each is in the nav and early in DOM order.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; Homepage design principles
minor · a tool finds it, a person confirms · #a11y-coga-hierarchy-visible-no-hover-reveal
Sub-menus and expandable content show a visible indicator that more exists and open on click or tap, not only on hover; heading levels form an outline of the page and same-level headings share a style.
Why and how to check
COGA 4.3.2 (clear indicators for hidden or expandable text, no hover-only reveal), 4.3.3 (headings as an outline) and 4.2.3 (consistent visual design).
Check: Tab and click each menu: it opens without hover; indicator present. Extract heading outline and check levels do not skip and same-level headings have the same computed font-size/weight.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: understanding controls
minor · a tool finds it, a person confirms · #a11y-coga-controls-recognizable
Links are visually distinguishable from body text (underline or equally strong cue), controls have a clear boundary or fill, non-interactive items do not look clickable, and nothing styled as a control is crowded against its neighbor.
Why and how to check
COGA 4.2.5 asks for recognizable controls, underlined links, and for non-clickable items not to look like links; 4.2.6 asks that interactive elements are not crowded. Overlaps WCAG 1.4.1 and 1.4.11.
Check: Screenshot in grayscale: can every link and button be found? List elements with pointer-style hover or link color that are not links.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
minor · a person checks it · #a11y-coga-icons-with-text
Icons that carry meaning are paired with a visible text label, each icon has one consistent meaning across the site, and no section is crowded with decorative icons.
Why and how to check
COGA 4.2.7: familiar icons next to content, one meaning per icon, no cluttered pages. Inclusive Components says tooltips are a last resort for icon-only controls.
Check: List every icon-only control on the homepage; each needs a visible text label or a proven universal meaning; verify the same icon is not reused for different things.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; Tooltips & Toggletips; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: clear content
advisory · a tool can check it · #a11y-coga-short-paragraphs
Body paragraphs carry one topic each and are about 50 words or fewer (English), with the point stated first; sequences of three or more items use bulleted lists and steps use numbered lists.
Why and how to check
COGA 4.4.5 gives about 50 words per paragraph as a guide for English, one point per sentence, lists for three or more items. It is a guideline, not a hard limit.
Check: Count words per p in main content; report paragraphs over about 60; flag long comma-chained runs of 3+ items that should be lists.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How Users Read on the Web
minor · a person checks it · #a11y-coga-plain-literal-language
Headings, labels, buttons and error messages use common words, literal wording, active voice and positive phrasing; acronyms and jargon are expanded or explained on first use; metaphors or idioms in navigation and calls to action are avoided.
Why and how to check
COGA patterns 4.4.1 to 4.4.4 (clear words, simple tense and voice, no double negatives or nested clauses, literal language) apply to headings, labels, navigation, instructions and error messages.
Check: Read every nav label, heading, button and error string: list jargon, unexplained acronyms, idioms, double negatives; each must be replaced or explained.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities
minor · a tool finds it, a person confirms · #a11y-coga-text-not-over-busy-background
Text is placed on solid or low-detail backgrounds (or a solid panel/overlay), not directly over photos, video, gradients or patterns with varying luminance.
Why and how to check
COGA 4.4.11 says not to place words over busy backgrounds, to use solid backgrounds for text blocks and strong contrast. Overlaps WCAG 1.4.3 measurement, which varies across a photo.
Check: Screenshot hero and banner text; sample background luminance across the text box (min and max) and confirm the worst case still meets 4.5:1; flag text over video.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: focus
advisory · a person checks it · #a11y-coga-limit-choices-per-screen
Each screen or section offers about five or fewer main choices, secondary links sit in the footer or under a titled 'more' link, and one feature is visually dominant.
Why and how to check
COGA 4.6.3 (avoid too much content) gives five or fewer main choices; also suggests one or two relevant images per page. This conflicts with ux-visible-navigation on desktop if read literally (see disagreements).
Check: Count primary actions/links competing in each viewport-height section of the homepage; flag more than about 5 equal-weight choices.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; Homepage design principles
minor · a person checks it · #a11y-coga-limit-interruptions
No pop-up, chat widget prompt, subscribe overlay or moving ad-like element interrupts the first task; moving content can be paused; notification settings, if any, are reachable from every screen.
Why and how to check
COGA 4.6.1 (limit interruptions) says users should control reminders and changing content and that blocking pop-ups such as subscribe prompts should be avoided. Overlaps ux-no-popups and a11y-pause-stop-hide.
Check: Load the homepage fresh, wait 30 s, scroll to 50%, and record any overlay, auto-opened chat, or banner that takes attention or blocks controls.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: predictability
minor · a tool can check it · #a11y-coga-no-surprise-changes
Content and controls do not move, open or change on load or without a user action: no layout shift after paint, no auto-opening windows or pop-ups, no auto-submitting forms; a loading indicator shows while content loads.
Why and how to check
COGA 4.5.1 and 4.9.1 list these. WCAG 3.2.x and the Core Web Vitals CLS limit (0.1) already cover parts of this.
Check: Record page load; measure CLS (<= 0.1); open page with no input for 10 s and assert no new windows, dialogs or focus changes; check that selecting a control does not submit or navigate.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; Web Vitals; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: help and support
advisory · a tool can check it · #a11y-coga-visible-help-contact
A way to reach a person (contact page, email or phone) is available from every page within one click, and a contact form has at most three required fields.
Why and how to check
COGA 4.8.1 and 4.8.5: human help easy to find on every page, few steps, not behind complex menus; feedback forms simple with no more than three required fields; chatbots not the only route.
Check: On five different pages assert a link to contact/help exists in header or footer; open the contact form and count required fields.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities
Cognitive accessibility: feedback
minor · a person checks it · #a11y-coga-action-feedback
Every user action produces immediate visible feedback (pressed or loading state, confirmation message after sending), and multi-step flows show completed, current and upcoming steps.
Why and how to check
COGA 4.5.10 (provide feedback) and 4.2.4 (make each step clear): confirmation after sending a form, visible status per step, buttons with pressed state.
Check: Submit each form and click each state-changing control; confirm a visible confirmation or state change within about 1 second; in multi-step flows, check a step indicator exists.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: avoid mistakes
minor · a person checks it · #a11y-coga-costs-up-front
On e-commerce and booking pages, price, fees, charges and key conditions are shown at the start of the task and the final total matches what was shown earlier; consequences and disadvantages of each option are stated next to it.
Why and how to check
COGA 4.5.3 (fees at the start) and 4.8.3 (state results and disadvantages of options). Relevant to the EAA e-commerce scope.
Check: Walk product page to checkout: record each price shown; confirm no new fee appears at the last step; check options list their inclusions and exclusions.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: memory
major · a tool can check it · #a11y-coga-allow-paste-autofill
Password and code fields allow pasting and password-manager autofill; fields use correct autocomplete attributes; the site does not block browser extensions or assistive add-ons.
Why and how to check
COGA 4.7.1 says password fields allow pasting; 4.9.2 says make sure password managers and spell checkers work and do not block extensions; 4.9.4 asks for HTML autocomplete on common fields. Overlaps a11y-accessible-auth and a11y-input-purpose.
Check: In login/signup: paste into password and confirm it is accepted; check autocomplete tokens (username, current-password, new-password, email, one-time-code); search for onpaste preventDefault handlers.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Cognitive accessibility: media
advisory · a person checks it · #a11y-coga-media-chunks-transcripts
Long video or audio is split into labelled segments of about six minutes or less where there are natural breaks, and a transcript or summary is easy to find next to it.
Why and how to check
COGA 4.3.5 (break media into chunks: segments generally six minutes or less, unique descriptive labels, transcripts easy to find) and 4.4.8 (summaries).
Check: For each media item over 6 minutes, check for chapter links or segmentation and a transcript or summary link adjacent.
Sources: Making Content Usable for People with Cognitive and Learning Disabilities; How to Meet WCAG 2.2 (Quick Reference)
Distinctiveness.
What keeps a page from looking like every other generated site. 25 rules.
Direction before build
major · a tool finds it, a person confirms · #distinct-written-direction
Before building, a written direction file exists that names 4 to 6 colours with roles, the display, body and (if used) mono font roles, a radius scale, and the subject-derived lead element of the hero, and every colour, font family and border-radius value in the shipped CSS appears in that file.
Why and how to check
Four independent practitioner sources recommend fixing tokens up front (a 4 to 6 colour plan; exact tokens instead of adjectives like 'modern and clean'), and Impeccable ships rules that flag a font, colour, radius or font size outside a DESIGN.md. Vague adjective prompts are said to return the statistical average.
Check: Read the direction file. Extract distinct computed colours, font-family primaries, border-radius values and font sizes from the built page's CSS. Fail any value that is absent from the file (allow neutrals derived from the stated ramp). Confirm the file names the hero's lead element and why it fits the subject.
Sources: How to Fix AI Slop in Web Design, Tell by Tell; frontend-design SKILL.md; Escaping AI Slop: Fix the 4 Overused AI UI Patterns; Why Your AI Keeps Building the Same Purple Gradient Website; pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); design-taste-frontend SKILL.md (local)
minor · a person checks it · #distinct-justified-exceptions
Every use of a flagged default in this family (overused font, purple/cyan palette, gradient text, glow, glass, cream+terracotta, identical card grid, eyebrow labels, marquee) is listed in the direction file with a subject-specific reason.
Why and how to check
The tells are defaults, not banned techniques; Capital & Compute and Laith's tool both provide an explicit way to mark an intentional use, and Impeccable offers inline ignores. An unrecorded use is indistinguishable from a reflex.
Check: Run the automated checks of the other distinct- rules. For each finding, confirm the direction file records the reason. Findings without a recorded reason fail.
Sources: How to Fix AI Slop in Web Design, Tell by Tell; How to detect AI slop in your design (ux-skill wiki); pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); design-taste-frontend SKILL.md (local)
Typography
minor · a tool can check it · #distinct-display-font-chosen
The primary font-family of the display headings (h1 to h3) is not on the overused list (inter, roboto, open sans, lato, montserrat, arial, helvetica, fraunces, instrument sans, instrument serif, geist, geist sans, geist mono, mona sans, plus jakarta sans, space grotesk, recoleta) unless it is the brand's own face or the direction file justifies it; the body face may be a neutral sans.
Why and how to check
Six independent groups of sources name default faces as the top typographic tell; Impeccable's list was current as of 2026-11-05. The list is contested (one local skill recommends Geist; Google ranks Inter 5th in popularity), so the stable rule is a chosen and recorded face, not membership of a list.
Check: In a headless browser, read getComputedStyle(fontFamily) for h1, h2, h3 and body; take the first non-generic family, lowercase it, ignore spaces, hyphens and underscores, and compare with the list. Flag a match on display headings that is not recorded as an exception.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; How to Fix AI Slop in Web Design, Tell by Tell; Why Your AI Keeps Building the Same Purple Gradient Website; Escaping AI Slop: Fix the 4 Overused AI UI Patterns; How to detect AI slop in your design (ux-skill wiki); unslop-ui (Jonathon's AI Wiki); design-taste-frontend SKILL.md (local)
minor · a tool can check it · #distinct-type-hierarchy-step
The rendered type roles (body, h3, h2, h1) contain at least one step of 1.25x or more between neighbours, and headings are visibly separated from body by size or by weight (at least 200 weight units heavier for at least 80% of headings).
Why and how to check
Impeccable flags a flat hierarchy when heading and body roles differ by less than 1.25x at every step (it needs at least 3 roles) and weight does not separate them. Other sources push stronger contrast (prg.sh suggests size jumps of 3x or more and extreme weights); those numbers come from one blog.
Check: For h1 to h3 and p, li, td, blockquote read computed font-size and font-weight. Compute the largest ratio between neighbouring size roles and the share of headings at least 200 weight units heavier than the body weight. Fail if every step is below 1.25 and fewer than 80% of headings clear the weight gap.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); Why Your AI Keeps Building the Same Purple Gradient Website; wystack agent-kit anti-patterns reference v0.12.0; How to detect AI slop in your design (ux-skill wiki); Learn Design: Typography
minor · a tool can check it · #distinct-hero-headline-proportion
The hero h1 is no larger than 6rem (96 px), is not both 72 px or larger and 40 characters or longer, uses letter-spacing no tighter than -0.04em, and wraps within about 45 characters per line.
Why and how to check
Impeccable's detector flags h1 at 72 px or more with 40 or more characters and its craft floor caps display type at 6rem with a tracking floor of -0.04em; Laith's wiki gives 30 to 45 characters per line and about -0.03em at 48 px or more. These thresholds conflict with very large, short headlines only when the headline is long.
Check: At 1440 px and 390 px viewports, read the hero h1 computed font-size, letter-spacing (in em), text length, and the longest rendered line length in characters. Fail any threshold above; review one-to-three-word headlines as allowed.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
Colour
minor · a tool can check it · #distinct-no-ai-palette
No gradient stop or primary accent has a hue between 260 and 310 degrees (purple, violet, indigo), and no cyan accent sits on a near-black ground, unless the brand is that colour and the direction file says so.
Why and how to check
Six independent source groups name purple-to-blue gradients as the most recognised tell. Impeccable tests purple text at hue 260 to 310 (and treats a declared colour as purple within 250 to 320). The Reddit-mining dataset (3,214,533 posts scanned, 3,033 comments) ranks the 'AI purple' gradient among its top tells; that measures what people complain about, not cause.
Check: Parse all gradients and accent colours in computed styles; convert stops to HSL or OKLCH hue; flag any in 260-310 degrees with meaningful saturation, plus Tailwind classes from/to purple, violet, indigo. Flag cyan accent text or fills on a background with very low luminance.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; Why Your AI Keeps Building the Same Purple Gradient Website; Escaping AI Slop: Fix the 4 Overused AI UI Patterns; unslop-ui (Jonathon's AI Wiki); JCarterJohnson/vibecoded-design-tells; How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
advisory · a tool can check it · #distinct-palette-not-reflex-cream
A warm cream or beige page background (minimum channel 209 or higher, red at least green at least blue, red minus blue between 6 and 48) combined with a terracotta or clay accent and a serif display face is used only when the direction file gives a subject reason.
Why and how to check
Impeccable has a rule for the cream surface; Anthropic's skill names cream with a terracotta accent near #D97757 as a default to avoid; Capital & Compute reports it as the second-wave cluster that appears after purple is banned; the local design-taste skill bans a family of cream, brass and espresso hexes for premium-consumer briefs. These are all one lineage plus one local skill, so the evidence is thin.
Check: Read body and main-section background colours; test the cream condition above; if true, also check for a terracotta/clay accent (hue roughly 10-30 degrees, moderate saturation) and a serif display face. Report all three together as one finding and require an exception record.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); frontend-design SKILL.md; How to Fix AI Slop in Web Design, Tell by Tell; design-taste-frontend SKILL.md (local)
minor · a tool can check it · #distinct-no-gradient-text
No text (headings, metrics, body) is filled with a gradient: no element combines background-clip: text with a gradient background-image.
Why and how to check
Gradient text appears in the Impeccable detector as its own rule, in wystack and the Reddit dataset, in Laith's list and in the local skills; emphasis can come from weight or size instead.
Check: Query all elements; flag those whose computed -webkit-background-clip or background-clip is text and whose background-image contains a gradient function.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; unslop-ui (Jonathon's AI Wiki); How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
Surface
minor · a tool can check it · #distinct-no-decorative-glow
No element has a chromatic box-shadow or text-shadow with zero offset and blur of 3 px or more, and no chromatic radial-gradient halo sits behind the hero or sections on a dark page, unless it encodes a state.
Why and how to check
Five independent source groups call glows and halos a default AI look. Impeccable's detector uses a minimum glow strength of 3 px and, for radial halos, a minimum effective alpha of 0.14 and contrast of 1.30 on the surface. One lineage's numbers; treat as indicative.
Check: Parse computed box-shadow and text-shadow lists: flag entries with 0 horizontal and vertical offset, blur at or above 3 px and a saturated colour. Flag background-image radial-gradient with a saturated centre fading to transparent over a dark surface.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); unslop-ui (Jonathon's AI Wiki); Escaping AI Slop: Fix the 4 Overused AI UI Patterns; How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
advisory · a tool finds it, a person confirms · #distinct-glass-with-purpose
Every element using backdrop-filter blur is listed in the direction file with its function (for example nav over media, overlay separation), has a solid-colour fallback, and glass is not the default style of cards.
Why and how to check
Impeccable and Capital & Compute flag decorative glass; the Reddit dataset indicates users rarely complain about glassmorphism, and trend articles disagree on whether it has peaked. No source gives a numeric limit, so this is a judgement call.
Check: Count elements with computed backdrop-filter containing blur. For each, confirm a recorded function and that text on it passes contrast against the worst-case content behind it (see a11y-contrast-text). Check a @supports fallback exists.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); How to Fix AI Slop in Web Design, Tell by Tell; JCarterJohnson/vibecoded-design-tells; backdrop-filter; Website design trends 2026
minor · a tool can check it · #distinct-no-side-stripe-border
No rounded element (radius 4 px or more) has a chromatic border on a single side wider than 1 px.
Why and how to check
Impeccable calls the thick coloured one-sided border on a card the most recognisable tell and sets 4 px radius as the point where it reads as a rounded card; Noqta makes the same claim. Square callouts with a severity rule are an older convention and are not flagged.
Check: For each element, compare computed border widths per side. Flag where exactly one side is above 1 px, its colour is chromatic, and the two corners away from that side have radius at or above 4 px.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; Escaping AI Slop: Fix the 4 Overused AI UI Patterns
minor · a tool can check it · #distinct-radius-elevation-scale
The CSS defines a named radius scale and a named elevation scale, each element role (button, input, card, badge, image) maps to a documented value, a single radius of 16 px or more is not applied to all roles, and no row of three or more cards pairs a hairline border with a shadow blur of 32 px or more.
Why and how to check
Noqta recommends a named scale (24 px everywhere versus a 6 px card example), Laith flags 16 to 24 px on everything and stacked border-shadow-tint, Impeccable flags rows of cards with a hairline border and a shadow of at least 32 px blur. The local design-taste skill asks for one consistent scale, so the compromise is a documented role map.
Check: Collect computed border-radius for buttons, inputs, cards, badges and images; fail when all share one value at or above 16 px. Collect box-shadow blur on card rows: fail when 3 or more adjacent cards have a visible 1 px border and blur at or above 32 px. Confirm radius/shadow tokens exist in the direction file.
Sources: Escaping AI Slop: Fix the 4 Overused AI UI Patterns; How to detect AI slop in your design (ux-skill wiki); pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); Why Your AI Keeps Building the Same Purple Gradient Website; redesign-existing-projects SKILL.md (local); design-taste-frontend SKILL.md (local)
advisory · a tool can check it · #distinct-customize-library-defaults
Component-library defaults (for example shadcn/ui tokens or Tailwind palette values) are customised: the shipped radius, colour and font variables differ from the library's init output.
Why and how to check
The Reddit dataset ranks the default shadcn/Tailwind look as its top specific tell; the local design-taste skill says never ship default state. Two groups, one of them an informal proxy of opinion.
Check: Generate the library's default theme in a scratch project; diff its CSS variables for radius, background, foreground, primary and font against the shipped CSS; flag when radius, primary and font all equal the defaults.
Sources: unslop-ui (Jonathon's AI Wiki); JCarterJohnson/vibecoded-design-tells; design-taste-frontend SKILL.md (local)
Template chrome
minor · a tool can check it · #distinct-eyebrow-restraint
Small uppercase or small-caps tracked labels placed directly above headings number at most ceil(sections / 3) on the page, none sits above the hero headline unless recorded as an exception, and numbered markers (01, 02, 03) appear only where the content is a real sequence.
Why and how to check
Impeccable bans a kicker above a heading outright and flags the hero chip; Anthropic's skill and Capital & Compute list tracked all-caps eyebrows and numbered markers; Noqta flags excess tracking on uppercase eyebrows; the local skill sets 'at most one per three sections'. The strict source and the lenient source disagree; this rule checks the lenient threshold.
Check: Find elements with text-transform uppercase (or small-caps), positive letter-spacing and length of 40 characters or fewer that are the previous sibling of a heading. Count them and compare with ceil(section count / 3). Separately flag numeric prefixes in these labels where the section list is not an ordered process.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); How to Fix AI Slop in Web Design, Tell by Tell; frontend-design SKILL.md; Escaping AI Slop: Fix the 4 Overused AI UI Patterns; design-taste-frontend SKILL.md (local)
minor · a tool can check it · #distinct-no-fake-status-badges
No pulsing status dot and no 'Live', 'New', 'Beta' or version pill appears unless it is bound to real changing state or a real release, and no pill chip sits above the hero heading.
Why and how to check
Noqta, Impeccable (pulsing-dot, hero-eyebrow-chip), the redesign skill and the design-taste skill all flag decorative liveness badges. Reserving pulse for live data is the stated fix.
Check: Find small elements (aspect ratio 1.5 or less) with an infinite pulse animation; find pill-shaped elements (radius at or above half the height) whose text is new, live, beta, alpha, or matches a version pattern, within the first viewport; flag unless linked to data attributes or a release record.
Sources: Escaping AI Slop: Fix the 4 Overused AI UI Patterns; pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); redesign-existing-projects SKILL.md (local); design-taste-frontend SKILL.md (local)
advisory · a tool can check it · #distinct-no-template-chrome
No more than two links or buttons end in an arrow glyph, no more than one meta string uses middle-dot separators, and em-dashes stay below saturation (fewer than 8 in total, or less than one per 500 characters of body text).
Why and how to check
Anthropic's skill and Laith list arrows on links and buttons; the Anthropic skill and the local skill ration middle-dot strings; Impeccable's em-dash rule is advisory and only fires at 8 or more em-dashes near one per 500 characters. The 'two' and 'one' limits are our own working thresholds. The local skill's total em-dash ban is not shared by any other source.
Check: Count link and button labels ending in an arrow character (arrow, chevron); count text nodes with 2 or more middle-dot separators; count em-dashes and body-text length.
Sources: frontend-design SKILL.md; How to detect AI slop in your design (ux-skill wiki); How to Fix AI Slop in Web Design, Tell by Tell; pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); design-taste-frontend SKILL.md (local)
Layout
minor · a tool finds it, a person confirms · #distinct-no-identical-card-grid
No section is built mainly from three or more same-size sibling cards that each hold an icon (or rounded icon tile), a heading and a short paragraph; the most important item gets more space than the others.
Why and how to check
Five independent source groups name the three-column icon-card feature row and the icon tile above a heading; Impeccable lists both as rules. Laith's tool lets you waive it when the product truly has three equal pillars.
Check: In the DOM find sibling groups of 3 or more elements with widths within 5% of each other that each contain an svg/img icon, a heading and a paragraph. Report groups; a reviewer accepts a group only when the content is genuinely parallel and recorded as an exception.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; Why Your AI Keeps Building the Same Purple Gradient Website; unslop-ui (Jonathon's AI Wiki); How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
minor · a tool can check it · #distinct-no-nested-cards
No bordered, shadowed or tinted card container sits inside another such container.
Why and how to check
Impeccable has a nested-cards rule and wystack lists cards inside cards; Laith flags stacking shadow, border and tint on one card; the local skill allows cards only where elevation signals hierarchy. Flattening with spacing and dividers is the common fix.
Check: For each element with a visible border, box-shadow, or non-transparent background different from its parent plus radius at or above 4 px, check whether an ancestor within the same section also qualifies. Report depth 2 or more.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local)
minor · a tool finds it, a person confirms · #distinct-hero-not-template
The first viewport shows a real, subject-specific element (the product UI, work, place or object) and does not match the stock skeleton: at most two of these five signals appear together (centred headline, badge or eyebrow above it, exactly two buttons, a logo row within the next section, a three-card feature row as the following section).
Why and how to check
Five independent groups describe the same skeleton; Capital & Compute lists it as 'stock page order'. A centred hero alone is common and conventional, so this rule scores the whole skeleton plus whether real subject content leads.
Check: At 1440 px, score the five signals from the DOM and computed styles. A reviewer confirms the hero's dominant visual is real subject content (not a gradient blob or a div mock-up).
Sources: How to Fix AI Slop in Web Design, Tell by Tell; Why Your AI Keeps Building the Same Purple Gradient Website; unslop-ui (Jonathon's AI Wiki); How to detect AI slop in your design (ux-skill wiki); pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); design-taste-frontend SKILL.md (local); Homepage design principles
advisory · a tool can check it · #distinct-section-rhythm-varied
Section paddings or heights are not all identical, no layout family is used for more than two consecutive sections, and a page of 8 sections uses at least 4 different layout families.
Why and how to check
Four independent groups flag uniform spacing and repeated section shapes (Impeccable's monotonous-spacing rule; Laith's uniform section rhythm). The numbers 4 of 8 and 2 in a row come from one local skill only.
Check: Measure each top-level section's computed padding-block and height; flag a coefficient of variation near zero. Classify each section's skeleton (full-bleed, split, grid of N, list, quote) and count runs and distinct families.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); wystack agent-kit anti-patterns reference v0.12.0; How to detect AI slop in your design (ux-skill wiki); Escaping AI Slop: Fix the 4 Overused AI UI Patterns; design-taste-frontend SKILL.md (local)
Motion
Content
minor · a tool can check it · #distinct-no-filler-copy
Shipped copy contains no lorem ipsum, no placeholder people or brands (John Doe, Jane Smith, Acme, Globex, Initech), none of the filler words streamline, empower, supercharge, unlock, elevate, seamless, unleash, world-class, enterprise-grade, next-generation, cutting-edge, game-changer, and every statistic is traceable to a source or removed.
Why and how to check
Five independent groups list invented round numbers, placeholder names and filler verbs; Impeccable has a buzzword rule and Laith names the stat pattern (for example 10x, 99.9%, 10,000+). The existing guideline set already covers plain tagline and objective language (ux-tagline-specific, ux-objective-language, ux-descriptive-labels); this rule adds the mechanical list check.
Check: Extract visible text and alt text. Regex for the list above (case-insensitive), lorem ipsum, and standalone round statistics (10x, 99.9%, \d{1,3},000+). For every statistic, require a data-source or footnote in the content brief; flag otherwise.
Sources: pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); How to Fix AI Slop in Web Design, Tell by Tell; Escaping AI Slop: Fix the 4 Overused AI UI Patterns; How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local); redesign-existing-projects SKILL.md (local)
Imagery
minor · a tool finds it, a person confirms · #distinct-real-imagery-and-icons
No img uses a placeholder or stock-demo host, the product is not shown as a mock-up built from empty div rectangles (live interface built from real text, controls and states is fine), no hero-sized inline SVG scene is assembled from primitive shapes, no emoji or Unicode pictograph stands in for an icon, and icons come from one family at one stroke weight.
Why and how to check
Five independent groups flag stock or machine-made imagery, emoji icons or default icon sets; Impeccable has a shape-assembled-illustration rule (advisory) and bans glyph icons in its craft floor; the local skill suggests picsum placeholders only while developing, so they must be replaced before shipping. Existing rule ux-informative-imagery already asks for imagery that informs. When no real screenshots or photos exist yet, show the product as live HTML/CSS with real text and states (not empty boxes), or leave the imagery out and mark the gap; never ship a placeholder host or stock demo image. For physical goods there is no interface to show, so the offer itself (price, seats, dates, a calculator or booking control built from real HTML), type set large as the image, or an omitted and marked gap are the honest options; real photographs have to be commissioned.
Check: List img and background-image sources and flag placeholder hosts (picsum, placehold, via.placeholder, unsplash source URLs used as demo); regex visible text and button labels for pictographic emoji used as icons; flag svg elements above half the viewport width built from at least a handful of primitive shapes; compare icon svg stroke widths. A reviewer confirms the hero image is real.
Sources: Escaping AI Slop: Fix the 4 Overused AI UI Patterns; pbakaus/impeccable (README, LICENSE, NOTICE, detector Rust source); unslop-ui (Jonathon's AI Wiki); How to detect AI slop in your design (ux-skill wiki); design-taste-frontend SKILL.md (local); redesign-existing-projects SKILL.md (local)
Concept
advisory · a person checks it · #distinct-swap-test
With the logo and name replaced by a competitor's, a reviewer can still tell which site is which from the hero, type, colour and imagery alone.
Why and how to check
Only Noqta states this test; Prg.sh's call for personality and a clear point of view is related. Treat as a judgement heuristic.
Check: Screenshot the first two viewports, swap the logo and brand name with a competitor's, and ask a reviewer who did not build it to identify the site and say what the design idea is. If they cannot, record a finding.
Sources: Escaping AI Slop: Fix the 4 Overused AI UI Patterns; Why Your AI Keeps Building the Same Purple Gradient Website
Process
advisory · a person checks it · #distinct-review-against-published-criteria
Before shipping, the page is reviewed against the four published Awwwards criteria (Design 40%, Usability 30%, Creativity 20%, Content 10%) with one piece of evidence recorded per criterion, and any criterion without evidence is treated as failing.
Why and how to check
Awwwards publishes these weights and states that at least 18 jurors score each nominee; Webby lists seven criteria (including accessibility and speed under Visual Design and Functionality) without weights; CSS Design Awards gives only thresholds. Using a public rubric makes the review explicit; the evidence requirement is our own.
Check: Score Design, Usability, Creativity and Content on 1 to 10 with a cited screenshot or report per score; compute 0.4*Design + 0.3*Usability + 0.2*Creativity + 0.1*Content; optionally also check the Webby list (content with a point of view, structure and navigation, visual design, functionality, interactivity, innovation, overall experience).
Sources: Awwwards Evaluation System; Webby Awards Judging Criteria: Websites and Mobile Sites (2026/2027); About CSS Design Awards; Lando Norris, Site of the Day