Models fall back on the same defaults: Inter everywhere, a badge above a centered hero, cards with a colored edge, gradients and glows. Adrian Krebs scored 1,590 Show HN landing pages in April 2026 against 16 such patterns; 22% showed four or more and only 46% showed one or none. Give the agent your own tokens, components and written rules, and it stops reaching for the defaults.
What the Show HN study measured#
Krebs collected the patterns from designers, wrote each as a deterministic CSS or DOM check, and ran them in a headless browser over the latest Show HN submissions. He estimates 5 to 10% false positives from a manual check, and says one pattern alone does not make a page AI-generated.
| Tier | Patterns matched (of 16) | Share of 1,590 pages | Pages |
|---|---|---|---|
| High | 4 or more | 22% | 347 |
| Medium | 2 to 3 | 32% | 508 |
| Low | 0 to 1 | 46% | 735 |
He also notes that Show HN submissions rose so fast with Claude Code that the moderators restricted Show HN for new accounts.
The tells#
Krebs groups the patterns his designer friends named into four kinds:
- Fonts. Inter for everything, especially centered hero headlines; one serif or italic accent word in an otherwise Inter hero.
- Colors. "VibeCode Purple"; permanent dark mode with medium-grey body text; barely passing contrast; gradients everywhere; large colored glows and shadows.
- Layout. A badge above the H1; colored borders on the top or left edge of cards; identical feature cards with an icon on top; numbered 1, 2, 3 steps; stat banner rows; emoji in the nav; all-caps section labels.
- CSS. Stock shadcn/ui and glassmorphism.
Inside an AI product the same habits show up in the agent's own UI: a sparkle icon as the AI mark, a spinner for every wait, raw JSON printed into the chat, an approve button with no risk or receipt.
Why models repeat them#
A model writes the most likely next token, and the most likely page is the average of the pages it learned from. Without a constraint, every prompt pulls toward the same font, the same card and the same gradient. Nutlope's Hallmark skill, one of the better-known answers, works by naming the tells and refusing them: its README lists "the eyebrow above every heading, tracked-capital labels, one coloured word in a headline" among the tells it lints for in a real browser, and its GitHub repository showed 47,358 stars on 4 November 2026.
What to change#
| Lever | What it does | Where it lives |
|---|---|---|
| Tokens | Fixes color, type, radius and spacing, so the model picks from your values, not its averages | globals.css or a theme file |
| Components you own | The hard parts (states, motion, keyboard) come finished, so the model composes instead of inventing | components/ in your repo |
| A rules file | Says what not to do in words the agent reads every session: no eyebrows, no gradients, sentence case | CLAUDE.md, AGENTS.md, .cursor/rules |
| A skill | A packaged set of rules and references the agent loads for design work | .claude/skills/ or npx skills add |
| A lint pass | Catches the tells after the fact | A script, a browser check or a skill's audit mode |
The first two do the most. A rules file saying "never use a gradient" works better when the model has a token set with no gradient in it and a card component with no colored edge.
Where PixilUI fits#
PixilUI covers the agent moments (approval card, run timeline, tool call row, streaming text, thinking state) as source you own, on one set of tokens in two design styles, with copy rules that ban the tells above: no eyebrows, no subtitles under titles, no sparkle as the AI mark, no glows. For your coding agent, the Claude Code and Cursor setup adds the registry, the MCP server and a rules file in one step. It does not design your landing page; Hallmark and similar skills are aimed at that.
Questions#
What makes a website look AI-generated?
In Krebs' list: Inter for everything, a badge above a centered hero, colored borders on the edge of cards, identical icon-topped feature cards, purple and gradient color, big glows, all-caps labels, emoji navigation and glassmorphism. One alone proves nothing; four or more together is the pattern.
How many Show HN pages look AI-generated?
Of 1,590 Show HN landing pages Krebs scored in April 2026, 22% matched four or more of his 16 patterns, 32% matched two or three, and 46% matched one or none.
Does a design rules file stop AI slop?
It helps when it names the tells and gives the agent something better to use instead. Rules work best next to your own tokens and components, so the model has fewer defaults to fall back on.
What is the Hallmark skill?
An open-source (MIT) design skill for Claude Code, Codex, Cursor and OpenCode, made by Together AI, that builds pages from a catalog of themes and lints the result against tells that read as machine-made.
Related#
- Is AI-generated UI accessible? What the studies found: the defaults that fail people, not just taste.
- React component libraries for AI agent interfaces: what to install instead of prompting each moment.
- How to show tool calls and their results in an AI chat: one moment done without raw JSON.
Sources#
- Adrian Krebs: Scoring Show HN submissions for AI design patterns (20 April 2026), read 4 November 2026
- Nutlope/hallmark on GitHub (README and star count), read 4 November 2026
- Hallmark, read 4 November 2026