Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Anti-Slop Check: Why a Page Looks AI-Made

    1

    "It looks AI-made" is usually blamed on purple, dark mode or neon.

    Free

    Security scanned
    Anti-Slop Check: Why a Page Looks AI-Made

    Anti-Slop Check: Why a Page Looks AI-Made

    Example session with this skill installed

    이 랜딩 AI 티 나는지 봐줘: ./landing/index.html (Does this landing page look AI-made? Check it.)

    • Read your context and instructions
    • Compiled the anti-slop check

    Summary

    Ran on a test landing page built with effects stacked on purpose. Desktop lights 3 of 5 pixel axes (candidate), phone 1 of 5. 6 important and 11 normal findings, plus accessibility.

    check.mjs output (excerpt, real run)

    표면 효과 — 첫 화면 픽셀 (켜진 축에 *)
      PC 1366×768  fire 3/5 · ramp 0.341* · hue_travel 0.25 · sat_share 0.999* (cdetail 1.3) · hue_n 3* · glow 0
      폰 390×844  fire 1/5 · ramp 0.132 · hue_travel 0.18 · sat_share 1* (cdetail 2.3) · hue_n 2 · glow 0
      판정: 겹침 후보
    
    [겹침]
      중요  화면을 덮는 그라디언트 — PC 1 · 폰 1
      중요  한 상자에 장식 3겹 — PC 4 · 폰 4
            a.btn [그라디언트+그림자+큰 라운드] / div.card [그림자+굵은 테두리+블러+큰 라운드]
      보통  네온 글로우 — PC 5 · 폰 5
      보통  그라디언트 글자 — h1 "Revolutionize Your Workflow"
    [문구]
      중요  상투 문구 — 25개 (혁신 · 솔루션 · 시너지 · seamless · cutting-edge · supercharge ...)
    [움직임]
      중요  바운스 · 오버슈트 이징 — a.btn × 2
    [접근성]
      중요  대비 4.5:1 미만 — a.btn.alt "데모 예약" [2.77:1]
      보통  13px 미만 글자 — 6
    
    요약: fire 3/5 · 중요 6 · 보통 11
    

    (겹침 = stacking, 문구 = copy, 움직임 = motion, 접근성 = accessibility, 중요/보통 = important/normal.)

    The same check on a plain page gave fire 0/5 and 0 findings.

    Three fixes

    1. Background gradient -> one color; get the impact from heading size instead (body).
    2. Card shadow + thick border + blur + big radius -> one 1px line (div.card).
    3. "Revolutionize your workflow with cutting-edge AI" -> one sentence saying what changes for the reader (.hero p).

    Shall I apply them? After editing I re-run the check and show fire and finding counts before and after.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    "It looks AI-made" is usually blamed on purple, dark mode or neon. We measured 11,918 reference screenshots (award winners, curated galleries, component marketplaces, animation library demos and screens AI tools actually produced) and those rules did not hold: purple separated award winners from AI screens no better than a coin flip (AUC 0.51), and neon showed up more in award winners than in AI output. What did hold was restraint. A page reads as AI-made when three or more surface effects stack on one screen. So this skill counts the stacking instead of blaming a color.

    What it does

    1. Opens a URL or local HTML at 1366x768 and 390x844 with Playwright, scrolls once so late blocks load.
    2. Pixels (first screen): five axes plus one gate. ramp (full-screen gradient), hue_n (accent colors), sat_share (saturated area, only when color areas are smooth so photos do not count), hue_travel (color wheel sweep), glow (centre glow on dark). Axes on = fire; 3 is a candidate, 4 is strong.
    3. DOM: gradients covering 30%+ of the screen, boxes with three decorations at once (gradient + shadow + thick border + blur + big radius), neon glow, gradient text, card-in-card, several filled buttons on the first screen, three font families, off-grid spacing, emoji used as labels, icon-tile rows, bounce easing, endless decorative motion.
    4. Copy: stock phrases in Korean and English (seamless, cutting-edge, supercharge, 혁신, 솔루션, 시너지 ...), em dashes inside body sentences.
    5. Accessibility: contrast under 4.5:1, text under 13px, 44px touch targets, and a real Tab press to see whether focus is visible.
    6. Writes report.json and screenshots, then you get a short report and three fixes in priority order. It edits only if you say so, then re-measures and shows before / after numbers.

    Rules it follows

    • Never says a page "was" or "was not" made by AI. The measure catches one failure pattern, not authorship. In our data, screens AI tools actually made lit up 6.8% of the time, award winners 8.5%.
    • Keeps your brand colors. Suggests removing a layer, not a new palette.
    • Other people's sites are measured and reported only; screenshots stay on your machine.

    findings.md explains the measurement, the tie-handling bug that once made purple look predictive (AUC 0.859 before the fix), and the limits.

    Instructions are in Korean; the check works on pages in any language.

    Source: https://github.com/Stophyuk/solutionzip-skills

    Pixel axes read only the first screen; a cookie banner or popup covering it gets measured instead. The three-decorations-per-box line was tuned on hand-made screens, not yet on a large set of real AI output. Gradient or photo backgrounds are skipped in the contrast check. It cannot tell whether AI made a page. Instructions are in Korean.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 16 days ago

    • Free with an Agensi account

    Requires

    Node.js 18+
    playwright
    Node.js

    Permissions

    Browser
    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Listed16 days ago

    What's inside

    Frequently Asked Questions

    Not by itself. In 11,918 measured screenshots purple did not separate award winners from AI screens (AUC 0.51 after fixing a tie-handling bug), and dark pages are common among award winners. The skill counts how many effects stack, not which color you use.

    No. In the measured set, screens AI tools actually made lit up less often (6.8%) than award winners (8.5%). It catches one failure pattern: stacked surface effects plus template copy.

    No. It keeps your brand colors and proposes removing layers, three fixes at a time. It edits only when you ask, then re-measures so you see the numbers before and after.

    It can open any public URL and report. The screenshots are for your own review; do not republish other people's pages.

    New in Frontend & Web Apps