slop-radar

Find AI slop in English and German text. slop-radar flags the buzzwords, filler phrases and formatting habits that make writing read like a chatbot draft, and turns them into a score from 0 to 100.

Try the live demo in your browser. It runs the same engine and phrase database as the CLI.

slop-radar does not try to prove who wrote a text. It measures AI-style writing: the phrases and structures that make text generic, whether a model or a person typed them. Use it as a linter for prose.

What it checks

• 437 English and 210 German phrases: chatbot pleasantries ("I hope this helps"), stock openers ("in today's digital age"), hype ("game changer", "unlock the full potential"), filler connectors ("moreover"), significance inflation ("plays a pivotal role") and their German counterparts ("tauchen wir ein", "maßgeschneiderte Lösungen", "auf das nächste Level").
• 16 structural patterns: "Let me …" starters, "It's not X, it's Y" pivots, "not only … but also" contrasts, em-dash chains, emoji headers, bullet overload, bold numbered lists, passive-voice density and more.
• Buzzword density: a short text packed with buzzwords loses extra points.

Matching is built for everyday text, not just exact copies of the list:

Score

Install

The npm package is not published yet (#1). Until then, install straight from GitHub:

npm install -g github:renefichtmueller/slop-radar

Requires Node.js 18 or newer. Once the package is on npm, npm install -g slop-radar and npx slop-radar will work as well.

CLI

slop-radar essay.md                  # full report (same as: slop-radar check essay.md)
slop-radar score article.txt         # score and rating only
slop-radar json draft.md             # machine-readable JSON, e.g. for CI
cat text.md | slop-radar             # read from stdin
cat text.md | slop-radar score

--lang en|de|auto    Force the language (default: auto-detect)
--help               Show help
--version            Show version

Colors are used only when writing to a terminal. Set NO_COLOR=1 to turn them off, FORCE_COLOR=1 to force them.

Library

import { detect, score } from "slop-radar";

const detection = detect("This transformative journey leverages cutting-edge innovation.", "en");
const result = score(detection);

console.log(result.score);   // 92
console.log(result.rating);  // "HUMAN"
console.log(detection.phraseMatches.map((m) => m.phrase));
// [ "leverage", "cutting-edge", "transformative", "journey" ]

Every phrase match carries positions and lengths, so you can highlight hits in the original text. detectWith(database, text, language) runs the engine against your own phrase lists; it has no Node.js dependencies and works in the browser.

How scoring works

Start at 100, then:

List-style patterns (bold numbered items, emoji headers, bullet blocks) report every hit but count at most a few times, so one long list cannot sink a document. The score is clamped to 0-100.

Example

Input:

Let me dive deep into this transformative journey. Here's the thing -- in today's
fast-paced landscape, it's worth noting that leveraging cutting-edge solutions is
crucial. Moreover, this holistic approach empowers stakeholders to unlock
unprecedented synergy.

Output (slop-radar check, abbreviated):

  Score: 40/100   LIKELY AI

  Buzzwords found: 17
    "dive deep", "crucial", "landscape", "cutting-edge", "transformative",
    "unprecedented", "unlock", "empower", "synergy", "journey", "moreover",
    "it's worth noting", "in today's fast-paced", "stakeholder",
    "here's the thing", "holistic approach", "leveraging"

  Patterns detected: 2
    let-me-starter, heres-the-thing

  Score breakdown:
    Buzzwords:               -34
    Buzzword density:        -20
    Let me / Here's:         -6
    Final:                    40

Rewritten:

How do we make our product development faster? We found that using modern tools
cut our deployment time by 40%. The team now ships weekly instead of monthly,
and customer complaints dropped.

Score: 100/100 HUMAN. Specific, concrete, no filler.

German works the same way. This paragraph scores 64/100 SUSPICIOUS with 8 phrase hits:

In der heutigen schnelllebigen Welt ist es wichtig zu beachten, dass maßgeschneiderte
Lösungen einen echten Mehrwert schaffen. Tauchen wir ein: Dieser ganzheitliche Ansatz
spielt eine entscheidende Rolle und hebt Ihr Unternehmen auf das nächste Level.

Phrase database

The databases are plain JSON in src/database/:

• phrases-en.json: 437 English phrases
• phrases-de.json: 210 German phrases
• patterns.json: 16 structural patterns (regex, weight, optional maxCount)

New phrases are welcome. CONTRIBUTING.md explains what qualifies and how inflections are handled.

Browser demo

demo/ is a static page that imports the compiled engine from demo/lib/. After changing the engine or the database, regenerate it:

npm run build:demo

CI fails if demo/lib/ is out of date, so the demo cannot drift from the CLI again.

Claude Code skill

Copy skill/SKILL.md (or superpowers-skill/SKILL.md) into .claude/skills/slop-radar/ to use slop-radar from Claude Code.

Why this exists

Text written by language models has a recognizable style: filler words, hedges, forced enthusiasm and a predictable structure. Once you notice it, you see it everywhere, and readers do too. slop-radar makes those habits visible so you can cut them.

Use it to:

• Clean up your own drafts
• Check content before publishing
• Score AI drafts and revise until they read like a person wrote them
• Enforce a writing standard in CI

Contributors

• @renefichtmueller: creator and maintainer
• Terry Sweetser (@tcsweetser): first test suite and CI test runs, the bare-file CLI fix, the three counting bugs in #3 and their fixes (#2, #4)

Thank you! See CHANGELOG.md for who changed what.

Related projects

• claude-cortex: persistent memory for Claude Code sessions.
• claude-sync: multi-device sync for Claude Code.

License

MIT