SlopScore

Find the patterns that make writing read as AI-generated, and fix them before you publish.

SlopScore scans text for 20+ habits of machine-written prose: "it's not X, it's Y" contrasts, dramatic one-line closers, "let's dive in" openers, inflated significance ("a pivotal moment"), AI vocabulary ("delve", "seamless", "leverage"), em-dash overuse, bold-label bullet lists, and chatbot leftovers like "I hope this helps!". It returns a 0-100 score and points at every match with a short fix.

It's a linter, not a detector. It doesn't guess who wrote a text. It shows you the specific sentences readers will notice.

• Web app (free, runs in your browser): https://slopscore-nine.vercel.app
• Pattern guides with examples: https://slopscore-nine.vercel.app/patterns/

CLI

npx github:kburrus64-max/slopscore README.md docs/

(Not on npm yet. The unscoped slopscore name on npm belongs to a different project, so install from GitHub for now: npm i -D github:kburrus64-max/slopscore.)

FAIL  71  Pure slop        docs/launch-post.md
           1: Chatbot residue: "Great question"
           3: Not X, but Y: "this isn't just a tool, it's"
           3: Inflated significance: "marks a pivotal moment"
ok     4  Reads human      docs/install.md

Options:

flag	default	what it does
--max <n>	40	exit 1 if any file scores above n
--format text|json|github	text	github prints annotations for Actions
--ext <list>	.md,.mdx,.txt,.html,.rst	extensions to scan inside directories
--ignore <ids>	none	skip patterns, e.g. dash,triad
--include-quoted	off	also flag text inside double quotes (quoted examples are skipped by default)
--min-words <n>	20	skip very short files

Code blocks, inline code, front matter, URLs and HTML tags are ignored, so READMEs with examples don't get flagged for their code. Read from stdin with -.

GitHub Action

name: SlopScore
on:
  pull_request:
    paths: ["**/*.md", "docs/**", "content/**"]
jobs:
  slop:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: kburrus64-max/slopscore@v1
        with:
          paths: "docs content README.md"
          max: "40"

Each match shows up as an annotation on the changed line. The check fails when a file goes over max.

Agent skill

SlopScore ships as an Agent Skill for Claude Code, Cursor, Codex and other agents that read SKILL.md files. The skill tells the agent to score its draft, rewrite only the flagged spans, and re-check:

npx skills add kburrus64-max/slopscore

As a Claude Code plugin (skill plus the hosted MCP server), from inside Claude Code:

/plugin marketplace add kburrus64-max/slopscore
/plugin install slopscore@slopscore

Cursor can load the same repo as a plugin through .cursor-plugin/plugin.json.

Prompt-only writing skills such as humanizer and no-ai-slop tell an agent what to avoid; SlopScore gives it a deterministic check to run afterwards.

Library

import { analyze, PATTERNS } from "slopscore";

const r = analyze("Great question! Let's dive in.", { ignoreQuoted: false });
r.score;    // 0-100
r.label;    // "Reads human" | "A little sloppy" | "Sloppy" | "Pure slop"
r.matches;  // [{ id, label, category, start, end, text, hint }]

No dependencies. Works in Node 18+ and the browser.

How scoring works

Each match has a strength. Strong tells (chatbot leftovers, not-X-but-Y, dramatic closers) count 3 points, medium tells 2, weak tells 1. Points are divided by text length (per 100 words, with a 100-word floor) and mapped to 0-100. Weak signals like a single em dash or one three-item list don't count unless stronger tells are also present, because careful human writers use them all the time.

The rules are regular expressions plus a sentence-shape check for rows of fragments. They are deterministic: the same text always gets the same score.

Limits

• English only for now.
• It flags patterns, not authorship. Plenty of human writing has a few of these, and a clean score doesn't prove a person wrote something.
• The CLI skips text inside double quotes, so style guides can quote bad examples. The web app and API count quoted text unless you pass ignoreQuoted.

Hosted API

The web app has a free JSON API (POST /api/check, up to 5,000 characters), a URL scorer, and pay-per-call endpoints for longer documents and batches. See https://slopscore-nine.vercel.app/llms.txt.

Credits

The pattern list builds on Wikipedia's Signs of AI writing (WikiProject AI Cleanup) and two MIT-licensed agent skills: blader/humanizer and petergyang/no-ai-slop. The detection code and scoring here are original.

License

MIT © Anansi Data. See LICENSE.