@indiekitai/llm-trope-check

Detect AI writing patterns in your text before you publish.

A fast, zero-dependency CLI (and Node.js library) that scans text for LLM writing tropes — those tell-tale phrases and patterns that scream "AI wrote this."

Patterns sourced from tropes.fyi, HN community observations, and PNAS research (arXiv:2410.16107) identifying statistically overused LLM vocabulary.

Quick Start

npx @indiekitai/llm-trope-check blog-post.md
# or
cat essay.md | npx @indiekitai/llm-trope-check

Install Globally

npm install -g @indiekitai/llm-trope-check

# Then use:
ltc blog-post.md
ltc --help

Usage

ltc [file]             Scan a file
cat essay.md | ltc    Scan from stdin
ltc --list             List all known tropes
ltc --json [file]      JSON output (for scripts/CI)
ltc --help             Show help

Example Output

LLM Trope Check  44 words
──────────────────────────────────────────────────
Score  100/100  F  Almost certainly AI ⛔
       ██████████████████████████████████████████████████

Found 12 pattern(s) across 5 categories:

Openers & Affirmations
  L   1  ●●●  "Certainly!" opener
         Certainly! It's worth noting that this is a fascinating...
         💡 Just answer. Skip the affirmation.

False Profundity
  L   1  ●●●  "It's not X — it's Y" em-dash reframe
         💡 This pattern creates fake insight. State the idea plainly.

...

Score & Grades

| Grade | Score | Label | |-------|-------|-------| | A | 0–5 | Very human 🟢 | | B | 6–15 | Mostly human 🟡 | | C | 16–30 | Somewhat AI-flavoured 🟠 | | D | 31–50 | Heavily AI-flavoured 🔴 | | F | 51+ | Almost certainly AI ⛔ |

Score is normalised per 1000 words, so longer documents aren't unfairly penalised.

Exit code 1 for D or F grade — useful in CI pipelines.

Trope Categories

• Openers & Affirmations — "Certainly!", "Great question!", "I'll shoot straight with you"
• False Profundity — "It's not X — it's Y" em-dash reframe (the #1 AI tell)
• LLM Vocabulary — delve, tapestry, camaraderie, nuanced, comprehensive, robust, embark, navigate the complexities, unlock, realm, landscape, foster
• Hedging & Padding — "It's worth noting", "It is important to note", "Moreover/Furthermore", "fascinating aspect"
• Structural Tells — Present-participle openers, "As an AI language model", "Here are N ways..."
• Enthusiasm Overclaim — game-changer, revolutionary, cutting-edge, state-of-the-art

Run ltc --list to see all 30+ patterns with tips.

Use as a Library

import { scan } from '@indiekitai/llm-trope-check';

const result = scan('Certainly! Let\'s delve into the tapestry of this problem.');
console.log(result.score);   // 0–100
console.log(result.grade);   // 'A'–'F'
console.log(result.findings); // detailed per-trope results

scan(text) returns:

{
  wordCount: number,
  score: number,        // 0–100 LLM-ness score
  grade: string,        // 'A' | 'B' | 'C' | 'D' | 'F'
  label: string,        // human-readable label
  findingCount: number,
  findings: Finding[],
  byCategory: Record<string, Finding[]>
}

CI Integration

# .github/workflows/check-content.yml
- name: Check for AI tropes
  run: |
    npx @indiekitai/llm-trope-check --json blog-post.md | jq '.score'
    npx @indiekitai/llm-trope-check blog-post.md

Returns exit code 1 if score is D or F — perfect for blocking AI-heavy content in CI.

Why?

LLMs have recognisable writing tics baked in by instruction-tuning and RLHF. Research shows GPT-4o's top overused words include camaraderie and tapestry. Claude loves delve. All of them overuse the "It's not X — it's Y" em-dash reframe.

This tool helps you:

• Catch AI-generated content before publishing
• Audit your own AI-assisted writing for over-reliance on tropes
• Run content quality checks in CI

License

MIT © indiekitai