#Rules for writing Layr itself
Because this law will appear on the Layr website, its own copy should follow the same standard.
#Write for a specific reader
Human-facing prose should assume the reader is a designer, engineer, founder, or product person trying to evaluate an interface. Use product examples they can recognize. Avoid generic motivational prose.
#State the point early
Readers should understand the rule before reading the theory.
#Keep information density high
A long article can be valuable when each section adds a new tool, distinction, example, or exception. Length becomes slop when paragraphs repeat the same principle with different adjectives.
#Prefer literal verbs
Use words such as:
- remove;
- compare;
- group;
- label;
- show;
- hide;
- measure;
- preserve;
- retry;
- select;
- explain;
- verify.
These verbs describe interface behavior directly.
#Vary rhythm naturally
Avoid writing every section with identical paragraph length, identical lists, or the same sentence opening. Technical reference material can be structured while still sounding authored.
#Keep strong opinions tied to evidence
When a claim comes from community discussion, label it as community evidence. When a claim comes from established accessibility guidance, say so. When a category is a Layr-specific inference, name it that way.
#Preserve uncertainty
AI-slop aesthetics change quickly. The rule should make room for trend drift and intentional exceptions.
#Rule maintenance
The Intentionality Law should evolve without turning into a frozen style guide.
#Version the signature registry
Maintain time-sensitive signals separately from the stable law.
Example:
signature_registry:
version: "2026.09"
signals:
- default-shadcn-treatment
- ai-purple-primary
- purple-gradient-text
- cream-serif-sage-cluster
- repeated-rounded-card-cluster
- pseudo-technical-metadata
- generic-ai-copy-cluster
The stable rule remains intentionality. The signature list can change as trends change.
#Review signals periodically
A pattern that is highly recognizable in 2026 may become ordinary platform convention later. Another pattern may replace it.
Re-evaluate:
- prevalence;
- false-positive rate;
- community recognition;
- semantic legitimacy;
- accessibility impact;
- brand use;
- framework defaults.
#Keep community signals separate from usability violations
A site can look generic and still be easy to use. A site can look original and still be unusable.
Layr should report these dimensions independently when possible, so a team can tell the difference between a page that needs a design decision and a page that needs a usability fix. Collapsing the two lets a genericity complaint borrow the urgency of a real defect, which is how anti-slop review loses credibility.
Example:
impact:
genericity: high
usability: low
accessibility: none observed
A surface with that profile deserves a design conversation and no incident. A surface with low genericity and high accessibility impact deserves the opposite, and the separation makes both routable to the people who can act on them.
#Research foundation
This is an original Layr rule. The sources below inform its detection model and editorial guidance. Each entry states what the source measured or claimed, what this law took from it, and what it cannot support, so a reader can weigh a finding against the evidence behind it rather than against the reputation of the citation.
#Community and AI-slop research
#JCarterJohnson, `vibecoded-design-tells`
A self-published, reproducible Reddit-mining project focused on visual cues people associate with vibe-coded websites. Its dataset reports 3,214,533 posts scanned across 47 subreddits between 2020 and 2026, 46,971 on-topic posts, and 3,033 comments drawn from 125 canonical threads asking why AI-built sites look alike. A design-context subset of 32,822 posts carries most of the aesthetic discussion, and roughly 29.5% of that subset registers negative sentiment about how these sites look.
The method matters more than the totals. The project ranks tells by prevalence share rather than raw count, because the topic barely existed before 2024: normalized per 10,000 posts, discussion rises from 1.81 in 2023 to 277.91 in 2024 and 335.95 in 2025. Raw counts would have measured the growth of the conversation instead of the weight of each complaint. Comment share inside the 125 canonical threads is the primary signal, since that material is entirely on topic, and post share across the design-context subset is secondary. An independent agent then re-checked each finding for reality, visual validity, and false-positive risk, and 11 of the 12 leading tells survived that pass.
The ranked results this law draws on:
| Tell | Comment share | Post share |
|---|---|---|
| Default shadcn/Tailwind kit | 2.5% | 2.2% |
| Purple or violet primary | 2.3% | 0.7% |
| Gradients and gradient hero text | 2.0% | 1.2% |
| Excessive animation and template fades | 1.1% | 2.0% |
| Broad rounding and pill buttons | 0.8% | 0.5% |
| Dark mode with neon glow | 0.7% | 1.3% |
| Emoji used as icons | 0.5% | 2.7% |
| Generic sans faces such as Inter or Geist | 0.4% | 0.4% |
| Symmetric hero, three cards, CTA | 0.4% | 1.6% |
The negative results carry as much weight as the positive ones. Bento grids rank last and are contested by the complainers themselves. Glassmorphism registers between 0.2% and 0.3%. Complaints about mesh, blob, and aurora backgrounds were rejected as a keyword artifact of that run. Layr's list of patterns that do not count as slop by themselves exists because of those three results.
Read the numbers for their order rather than their magnitude. Every share is small in absolute terms, because most comments describe the overall feeling of sameness instead of naming a layer, and the author states that share metrics understate the visual impact of fonts and gradients. The ranking tells Layr which signals deserve weight. It does not tell Layr that 2.5% of interfaces have a problem.
- Repository: https://github.com/JCarterJohnson/vibecoded-design-tells
- UI skill: https://github.com/JCarterJohnson/vibecoded-design-tells/blob/main/skill/SKILL.md
- Data and methodology: https://github.com/JCarterJohnson/vibecoded-design-tells/blob/main/unslop-ai-ui/DATA_AND_GRAPHS.md
The companion unslop-ui skill is a second and separate input. It asks for four decisions before an interface is built: a named reference site or design direction, a specific brand color that is not the framework default, a chosen typeface with a stated reason, and a layout intent derived from real user goals. It carries the rule this law restates as the replacement test, that a fix must apply the project's own choices rather than swap one default for another, and it supports an unslop-ignore marker so deliberate choices survive scanning. Layr's correction model owes its sequence to that skill. Its 2026 addition of the cream, serif, and sage combination to the same list is the clearest available evidence that anti-slop recipes turn into recipes.
Limits to restate in any finding built on this source. The work is self-published rather than peer-reviewed. It measures what people say about interfaces, not what those interfaces do to task success, so nothing here supports a usability claim. Reddit skews toward developers, indie founders, and early-stage products, which over-represents the categories where generated sites are most common. Two of the largest subreddits needed backoff recovery during collection, which affects coverage. Treat the percentages as dataset-specific measurements of community recognition.
#Reddit discussions reviewed
Several 2026 discussions repeat themes of visual sameness, generic purple/blue palettes, rounded boxes, default component-library output, repetitive landing-page structures, and the need for project-specific design decisions.
Two things recur across them that the ranked dataset does not capture. Complaints cluster on marketing pages rather than on product surfaces, because landing pages are what people share, which means the corpus says more about first impressions than about interfaces people use daily. And commenters describe the effect before they can localize it: the page is identified as generated, then the reasons get assembled afterwards. That order is why this law reviews clusters and why it keeps authorship out of findings.
Representative threads:
- https://www.reddit.com/r/ClaudeCode/comments/1u7g0z5/i_scanned_3200000_posts_across_47_ai_and_saas/
- https://www.reddit.com/r/vibecoding/comments/1p5vfjr/i_analysed_500_vibe_coded_websites_heres_what_i/
- https://www.reddit.com/r/vibecoding/comments/1ubewk8/any_advice_to_escape_the_ai_slop_web_design_look/
- https://www.reddit.com/r/vibecoding/comments/1s37qlp/what_are_the_signs_of_an_ai_slop_ui/
- https://www.reddit.com/r/micro_saas/comments/1skltqv/stop_making_vibecoded_sites_that_look_like_ai_slop/
- https://www.reddit.com/r/webdesign/comments/1vfck61/it_feels_like_there_are_more_ai_websites_than/
Treat comments as observations and opinions, not controlled usability evidence. They establish that a pattern is recognizable to a technical audience in 2026. They establish nothing about comprehension, task success, error rates, or conversion, and a finding that cites them should claim only what they can carry.
#AI-writing pattern reference
#blader/humanizer
humanizer is an MIT-licensed agent skill for editing AI-sounding prose. It works from one premise: a language model writes what is most likely to come next, while a person chooses for one reader and one subject. Every pattern it documents is a statistically favored default rather than a decision, which makes it this law's argument applied to sentences instead of surfaces.
Its catalog holds 25 patterns in five groups, and it grades them instead of listing them flat. Five are strong enough to act on after a single sighting: not-X-but-Y contrasts, one-line closers and dramatic fragments, aphorisms that dress an ordinary point as a hidden truth, staged run-ups before the point, and arguing against a position nobody raised. The rhythm group is weak alone and needs companions: forced triads, repeated sentence openings, dashes used as a universal connector, stacked qualifiers, hyphenated pairs, and passive constructions that hide who acts. A third group covers inflation and borrowed authority, including overused vocabulary, inflated significance, vague connective verbs, shallow -ing riders, sales language, unnamed experts, and the avoidance of is, are, and has. A fourth covers formatting by rule: bold as decoration, decorative headings, curly quotation marks. The last covers leftovers from chat and draft, where chatbot residue is described as the most certain tell on the list and the easiest to miss, because the wrapper surrounds real content.
Three of its structural decisions carry into Layr unchanged. Patterns are graded by strength rather than treated as equal. Several tells appearing together are the safeguard against false positives. And the skill names when not to act, exempting quotations, titles, proper names, discussion of the phrase itself, and any text written before models could have produced it.
Layr adapts those ideas to product copy and changes what the review protects. Humanizer preserves the writer's voice. Layr preserves the reader's ability to act, so a consequence, a constraint, a recovery path, or an accessibility instruction outranks every pattern in the catalog. Where the two conflict, the information stays and the pattern survives. The examples and wording in this document are original to Layr.
#Emoji-as-formatting sources
Wikipedia's Signs of AI writing carries a dedicated Emoji as formatting entry among its style signals, and frames the whole list the way this law frames its own: the patterns are "only potential signs of a problem, not the problem itself." The sparkle's adoption as the visual shorthand for AI is documented separately, which is what makes it the most self-defeating glyph a product can open a section with. The Reddit-mined dataset already cited under community research measured emoji used as icons rather than emoji in prose; it is the closest quantitative evidence available and is reported in section 12 rather than restated here.
- Signs of AI writing: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
- Sparkles emoji: https://en.wikipedia.org/wiki/Sparkles_emoji
- Detecting AI text, including emoji tells: https://www.rte.ie/brainstorm/2025/1128/1545935-chatgpt-ai-writing-text-detection-words-phrases-emojis/
#Established UX and accessibility guidance
#Nielsen Norman Group: Aesthetic and Minimalist Design
Nielsen's eighth usability heuristic argues that irrelevant or rarely needed information competes with relevant information for attention, so every extra unit on a screen diminishes the visibility of the units that matter. The supporting article separates two ideas that are usually merged. Minimalist design means including everything necessary and nothing more, which makes removing too much a failure of the same heuristic as adding too much. Aesthetic design means applying scale, hierarchy, balance, contrast, and the Gestalt principles well. A visually spare interface can fail the heuristic by omitting what people need, and a visually rich interface can satisfy it completely.
That distinction is load-bearing here. It is the guidance that stops an anti-slop review from becoming an argument for flat, grey, empty pages, and it supplies the removal criterion this law uses: maximize the elements with informational value, reduce the ones without it, and judge both by what the reader needs rather than by how much is on the screen.
The article also describes the aesthetic-usability effect, where people judge attractive interfaces as easier to use and forgive real problems in them. Layr uses that in both directions. A polished generated surface can hide an incomplete product model, which is why the interaction catalog exists. A surface that reads as generic is not automatically harder to use, which is why genericity and usability are reported as separate dimensions.
#Nielsen Norman Group: Concise, Scannable, and Objective Web Writing
Morkes and Nielsen tested five versions of one travel site holding identical information: a promotional control, a version cut to about half the word count, a version reformatted with headings, lists, and bold text, a version with the promotional language stripped out, and a version combining the last three. Measured against the control, the concise version scored 58% better on the study's usability index, the scannable version 47%, the objective version 27%, and the combined version 124%.
Two results survive the study's age. Conciseness did the most work of any single edit, and the three edits compounded rather than traded against each other, which argues against treating scannable formatting as a substitute for cutting words. The finding that readers distrust promotional language also explains why generic benefit copy underperforms on product pages even when it reads smoothly.
State the caveats whenever the numbers are cited. This is a 1997 study, run on a single site, with a small participant pool, and the percentages describe that experiment rather than a constant to apply elsewhere. Layr cites the ranking and the direction, never a promised improvement.
#W3C WAI: Clear and Understandable Content
This cognitive accessibility objective collects thirteen design patterns under one goal, that people with language and processing impairments can understand a page. The patterns this law's copy catalog draws on most are clear words in place of complicated vocabulary, simple tense and voice, avoiding double negatives and nested clauses, literal rather than figurative language, succinct text, unambiguous formatting and punctuation, instructions separated rather than combined, implied content made explicit, and white space used as visual separation.
Two of them cut against a reflex an anti-slop review can develop. Explaining implied content asks writers to state unstated assumptions, and providing summaries asks for an overview of long material. Both add words. A review optimizing only for brevity will delete exactly the sentences this guidance requires, which is why the law's criterion is informed action rather than word count.
State the status precisely in a finding. These are supplemental W3C materials supporting WCAG 2, not success criteria, so they inform recommendations and do not establish conformance failures.
#W3C WAI: Keep Text Succinct
The pattern asks for content chunked into short blocks: one topic per paragraph, the main point first, one point per sentence, unordered lists for items and requirements, numbered lists for sequences of three or more steps, and concise descriptive headings. It suggests breaking paragraphs longer than about 50 words, and watching for repeated linking words such as and and but, which usually mark a sentence carrying more than one point.
It names who this serves: readers with learning disabilities, memory impairments, or difficulty sustaining attention, and anyone reading while doing something else. That framing changes the argument. Front-loading the main point and splitting a long sentence are accessibility work rather than style preference, which gives Layr a basis for copy findings that would otherwise reduce to taste.
Two entries in the copy catalog come directly from this pattern. Copy that hides the useful fact is the failure it describes, and the formatting-by-rule entry inherits its distinction between lists a reader jumps into and prose that happens to be arranged vertically.
#Apple Human Interface Guidelines: Design Principles
Apple's current design principles connect purpose, clarity, hierarchy, consistency, harmony, and craft, and three of them map onto this law closely. Clarity is described as a product of hierarchy, where order, spacing, and contrast make the most important item on a screen the most obvious one, which is the standard the centering and whitespace entries measure against. Consistency is stated behaviorally, that things which look the same should behave the same, which is the argument this law makes against decorative accents, fake controls, and status color without a mapping. Craft is framed as the attention to detail that builds trust, naming typography, responsive animation, performance, and iteration, and it supports the layout-quality entry against the assumption that anti-slop work is mostly color and typeface selection.
Cite this as platform design guidance from a vendor with a strong point of view, not as independent research. It is persuasive about intent and hierarchy. It is not evidence about what people recognize as generated, and it describes conventions that are strongest on Apple platforms.
#Source and evidence caveats
The term AI slop is informal. It does not have one accepted academic definition.
Community recognition changes as tools, templates, and model defaults change. A visual pattern that reads as generated today may become an ordinary convention later. Human designers also use the same patterns, and AI-assisted teams can produce highly intentional work.
For those reasons, Layr should use this law to evaluate decision quality and observable interface consequences. Authorship should remain outside the finding.
The most durable question remains simple:
Does this interface contain a choice because the product needs it, or because the generator needed something to put there?