How the scan works

The four axes

A page can be slop in more than one way. slop-detect scores four independent axes; the two slop axes want a low score, the system and AEO axes want a high one.

Install

Nothing to set up for a one-off scan. npx fetches the CLI, opens the page in a real browser, and prints the score.

# one-off scan, zero install
npx slop-detect https://your-site.com

# or install the CLI globally
npm i -g slop-detect
slop-detect https://your-site.com

The CLI

Score one or many URLs. The flags compose: axes, presets, the system axis, CI gating.

slop-detect <url> [...urls]        # score one or many pages
  --copy              # also score the copy axis
  --axes all          # design + copy together
  --design-md auto    # check against <origin>/DESIGN.md
  --preset strict     # score a curated subset
  --remote            # scan via the API, no local browser
  --json              # machine-readable output
  --fail-on heavy     # exit non-zero to gate CI

AEO · agent-readable

The AEO axis asks a different question: can AI engines actually fetch, read, and cite this page? A page can look perfectly human-made and still be invisible to ChatGPT, Claude, or Perplexity because it blocks their crawlers or serves JS soup. 8 weighted checks, total weight 100. The four required checks are the fundamentals; the rest are recommended bonuses. Higher is better: AI-Ready (≥80), Partial (≥50), Invisible (<50).

The system axis · DESIGN.md

The absolute slop score asks "does this look AI-generated?" The system axis asks the durable question: does the page honor its own declared design system? Point it at a DESIGN.md (Google Labs' open spec: colors, typography, radii) and it reports drift: fonts in use that aren't declared, CTA or surface colors off the palette, radii off the scale. A bespoke site checked against its own system scores Aligned, never "slop". Every drift item is a named, contestable signal.

slop-detect https://your-site.com --design-md auto

  ◳ system   Drifting   62/100   (DESIGN.md)
  drift  Inter is in use but is not declared in the system
  drift  the CTA fill is off the declared palette
  drift  a 16px card radius is off the declared 4 / 8 / 12 scale

Web & REST API

One curl away. POST a URL, read back the grade, score, verdict, and definitions version.

# score a page on the design + copy axes
curl -s -X POST https://slop-detect.com/api/scan \
  -H 'Content-Type: application/json' \
  -d '{"url":"https://your-site.com","axes":["design","copy"]}' | jq

endpoints

Auth. No key needed for normal use. GET endpoints are public; POST is per-IP rate-limited (scan 6/min from the web, 3/min server-to-server). Pass X-API-Key or Authorization: Bearer for higher, per-key limits: free 10/min, pro 60/min, unlimited. Foreign browser origins are rejected without a key; the CLI and MCP are not.

Continuity & directory

Detection is free and stateless. The optional continuity layer remembers a domain, re-scans it daily, and emails you when it regresses to slop or drifts off its DESIGN.md between redesigns. list: true also opts the domain into the public, crawlable directory with a real backlink.

Programmatic · the core engine

@slop-detect/core is the pure, runtime-agnostic engine: it runs in Node, a Cloudflare Worker, or the browser, and knows nothing about fetching a page. You open the page, run each pattern's detect() in-page, then hand the triggered results to the scorer.

import { scorePatterns, PATTERNS, DEFINITIONS_VERSION } from '@slop-detect/core';

// core is pure: YOU open the page (Playwright/Puppeteer) and run each
// pattern's detect() in-page, then hand the triggered results to the scorer.
const summary = scorePatterns(results);
console.log(summary.score, summary.grade, summary.tier);
console.log(PATTERNS.length, DEFINITIONS_VERSION); // 27 2026.09

Presets & agents

Not every audience cares about the same tells. A preset scores a curated subset or weighting without forking the ruleset. Pass --preset on the CLI or { "preset": "strict" } to the API.

MCP · self-audit before shipping

slop-detect-mcp is a Model Context Protocol server that lets an agent scan a page before it ships and pull the fix prompt back into the loop. Four tools: scan_page, check_aeo, check_design_system, fix_prompt.

{
  "mcpServers": {
    "slop-detect": { "command": "npx", "args": ["-y", "slop-detect-mcp"] }
  }
}

CI gate · GitHub Action

The action scans a deploy-preview URL, posts a sticky PR comment with the grade and triggered patterns, and fails the check when slop creeps above your threshold. fail-under takes a number (fail if the score exceeds it) or a letter grade. Leave it empty for report-only mode.

- uses: ravidsrk/slop-detect/packages/[email protected]
  with:
    url: ${{ steps.preview.outputs.url }}   # your deploy-preview URL
    fail-under: 'B'                       # fail if the grade drops below B

Tiers & grades

The 0-100 score is the source of truth; the tier and letter grade are derived from it. Lower is cleaner. The grade bands are read from the engine, so they cannot disagree with the scorer.