slopbench: 30 default AI landing pages, audited before and after

A public benchmark for “does this look AI-made?”: 30 pages generated from one-line prompts, the same audit run on every page, and three redesign methods compared on the same pages with blinded judges.

How the corpus was made

Each page was generated by a fresh Claude Code subagent in its own context, with no skills loaded, so no generator saw any other page. Odd ids ran on claude-opus-5-5, even ids on claude-sonnet-5-5. Each subagent got exactly two things, both stored with the corpus.

The design prompt, always one line with no design guidance:

Build a landing page for a <niche> business.

And a fixed harness note that only sets the output format and isolation:

Write the page as a single self-contained file at {OUT}/index.html. You may load scripts, styles, fonts and images from public CDNs. Do not ask questions. Do not load or invoke any skills. Do not read any other files in the repository. When you are done, reply with the single word: done

Single self-contained files keep 120 or more page audits cheap and need no per-site install. The format constraint is in the harness note, not the design prompt. The prompts keep their grammar as generated (“a artisan bakery”), because changing them after the run would make the record wrong.

The 30 pages

Counted from the page sources on 2026-11-22: 27 use a warm cream background token, 3 are near-black, 21 put an italic phrase inside the h1, and 19 load Fraunces.

IdNicheModelBackground
s01dental clinicclaude-opus-5-5cream
s02artisan bakeryclaude-sonnet-5-5cream
s03yoga studioclaude-opus-5-5cream
s04emergency plumbingclaude-sonnet-5-5cream
s05family law firmclaude-opus-5-5cream
s06small-business accountingclaude-sonnet-5-5cream
s07specialty coffee roasterclaude-opus-5-5cream
s08mobile dog groomingclaude-sonnet-5-5cream
s09wedding photographyclaude-opus-5-5cream
s10residential real estate agencyclaude-sonnet-5-5cream
s11independent auto repairclaude-opus-5-5cream
s12floristclaude-sonnet-5-5cream
s1324-hour gymclaude-opus-5-5near-black
s14tattoo studioclaude-sonnet-5-5cream
s15landscaping and garden designclaude-opus-5-5cream
s16heating and air conditioning (HVAC)claude-sonnet-5-5cream
s17veterinary clinicclaude-opus-5-5cream
s18craft breweryclaude-sonnet-5-5near-black
s19project management softwareclaude-opus-5-5cream
s20AI writing assistantclaude-sonnet-5-5cream
s21personal finance appclaude-opus-5-5cream
s22cybersecurity consultingclaude-sonnet-5-5near-black
s23architecture firmclaude-opus-5-5cream
s24language schoolclaude-sonnet-5-5cream
s25home cleaning serviceclaude-opus-5-5cream
s26roofing contractorclaude-sonnet-5-5cream
s27food truckclaude-opus-5-5cream
s28boutique hotelclaude-sonnet-5-5cream
s29physiotherapy clinicclaude-opus-5-5cream
s30kids' coding schoolclaude-sonnet-5-5cream

Treatments and baselines

Every treatment edits its own copy of each corpus page, one fresh agent per page, so no run sees another page. All treatment agents ran on claude-opus-5-5, so the comparison is between methods, not models. Every treatment got the same task text:

This landing page looks generic and AI-made. Redesign it so it no longer does, keeping its business, content and purpose. Edit the file in place. It must remain a single self-contained HTML file (assets from public CDNs are allowed).
TreatmentMethod
beforeThe corpus page as generated
restyleThis project’s Claude Code plugin skill and CLI
impeccableImpeccable 4.10.0 (Apache-2.0), unmodified clone of pbakaus/impeccable at commit b49177d
tasteTaste Skill 2.5.1 (MIT), unmodified clone of Leonxlnx/taste-skill at commit f6b7657

What is measured

  • Tell count: findings from impeccable detect (62 rules) and restyle’s 17 rules, by severity.
  • Second-order looks: cream with serif and terracotta; near-black with one acid accent; the hairline broadsheet.
  • Accessibility: axe-core WCAG 2 A and AA, serious and critical violations; WCAG contrast of every visible text node.
  • Performance: Lighthouse mobile scores, LCP, CLS and TBT.
  • Distinctiveness: distance of the page’s screenshot from the corpus, using CLIP ViT-B/32 image embeddings and the nearest neighbours in the reference set, leave-one-out for corpus pages.
  • Prototypicality: the share of category conventions the page keeps, so “distinct” cannot mean “broken”.
  • Blinded pairwise judge: judges see two screenshots named A and B, in both orders, with a fixed rubric; the key that maps A and B back to treatments is kept apart until all verdicts are in.
  • Homogeneity: mean pairwise CLIP similarity within each set, and how often each shared feature appears.

Results

Provisional numbers. Corpus counts are from a grep of the 30 page sources on 2026-11-22; the slopbench audit will replace them. Results marked "pending" have not been measured yet.

Not measured yet. The table of per-page means for the four treatments (tells, second-order looks, distinctiveness, axe, Lighthouse) and the blinded judges’ win rates appear here when the run is scored. Nothing on this site states a result before then.

Where restyle loses

Not measured yet. When the run is scored, every metric where a baseline beats restyle is listed here, with the pages it happened on.

Limits of this benchmark

  • The corpus is Claude only for now. Outputs from v0, Lovable and Bolt for the same 30 prompts can be added under slopbench/corpus-external/<tool>/ and scored with the same commands.
  • Pages are single HTML files, not full apps with routing and data.
  • Judges are fresh model instances given only the two screenshots and the rubric, not human designers. Designers also disagree with each other often, so a win rate near 50% means “no clear preference”.
  • restyle’s own rules are written by the same project that is being measured. Impeccable’s rules, axe and Lighthouse are independent of it.

Per-page reports in the gallery