Slop Scan

The question: does this rendered site visually have the common fingerprints of an AI-generated website?

In scope: what you can see. Typography quirks, gradient habits, mechanically perfect composition, repeated card systems, pill overload, accent words in headings, glow, bento, excessive rounding, icon treatment, spacing rhythm, em-dash-heavy visible copy, and similar tells.

Out of scope, always: whether photos, reviews, stats or content are real; whether the design fits the business or would work for another one; copy quality beyond em dashes; UX and conversion. Don't raise these even when you notice them.

1. Render and measure (primary evidence)

bash scripts/render.sh <url | folder | file.html> <outdir>

It loads the page in Chromium (serving a local folder itself), measures every fingerprint from computed styles and visible text, and saves desktop-top.png, desktop-full.png, mobile-full.png and render-scan.json. The first run installs Playwright into ~/.cache/slop-scan. If it fails twice, stop fighting tooling: use Claude in Chrome or ask for screenshots.

For a repo with a dev server, start it and pass the localhost URL.

2. Look at the screenshots (decides the verdict)

View all three screenshots. The measurements can over- or under-count; your eyes decide. Go through references/tells.md and mark each fingerprint Present, Partial, Absent, or Not checked (if you couldn't see it). Never mark something Absent because you didn't look.

3. Verdict

• Yes, strong AI fingerprint: 6+ fingerprints present, or 3+ of the big tells.
• Some AI fingerprints: 3–5 present, or 2 big tells.
• No, few AI fingerprints: 0–2 present.

One tell alone is a trend. The AI look is several landing together.

Big tells: em dashes in visible text, accent word in headings, gradient text or decorative gradients, glow blobs, pill/eyebrow overload, icons in rounded tiles, repeated 3-card rows, mechanical spacing rhythm, centered hero + two buttons.

Name any matching cluster from references/tells.md (v0/shadcn kit, Lovable/Bolt purple SaaS, Claude-style cream editorial, dark glow SaaS, bento product page).

4. Report

Short, terminal-friendly, no wide tables.

Does this visually have the fingerprints of an AI-generated site? YES (strong)
9 present, 6 big tells. Cluster: v0/shadcn kit.

PRESENT
- Em dashes: 14 in visible text, incl. h1 and both CTAs
- Accent word: "spotless" in the h1 set in Caveat
- Pills: 7, three sitting above headings
- Icon tiles: 6 Lucide icons in 44px rounded squares
- 3-card rows: services and pricing, identical 357x230 cards
- Rhythm: 96px padding on 7/7 sections, same 1120px column
PARTIAL
- Fade-up: 4 sections hidden until scroll
NOT CHECKED
- Hover states (static screenshots)

FIX ORDER
1. ...

5. Fix (only when asked, or after the report and the user says go)

To find the code for a fingerprint, python3 scripts/scan.py <repo> lists source matches with file and line.

1. Remove the fingerprint; don't restyle it into another one. Check every replacement against references/banned-replacements.md.
2. Typical fixes:
   • Em dashes → periods, commas or colons. Punctuation only; don't rewrite copy.
   • Accent word in a heading → same font, style and color as the rest of the heading.
   • Gradient text/backgrounds → flat color. Glow blobs and glass → delete.
   • Pills and eyebrow labels → delete; keep real controls (filters, form chips).
   • Icon tiles → delete the tile; use the icon plain or drop it.
   • Excessive rounding → fewer, smaller radii, varied by level; drop hairline-on-everything.
   • 3-card rows and bento → vary count and size; not every group needs to be cards.
   • Mechanical rhythm → vary section spacing by importance; let something go full-bleed or break the column.
   • Centered hero + two buttons → one action; consider a non-centered composition.
   • Oversized clean sans → a typeface with character, sized to the content.
   • Fade-up on everything, hover lift on every card → remove; keep at most one deliberate moment.
3. Touch only fingerprints. Leave content, photos and structure that aren't tells.
4. Re-run render.sh, view the new screenshots, and report before → after. Flag any banned replacement that slipped in.

Save the report as slop-scan.md next to the project when asked.