Anti-AI Writing Style Skill for Claude

A Claude skill that strips the most common AI tells out of your writing. The kind of tells that make a reader stop mid-paragraph and think "a robot wrote this."

Why this exists

In 2025, researchers at UMD and UMass ran a study called "People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text" (Russell, Karpinska & Iyyer, arXiv:2501.15654). They paid five experienced LLM users to read 300 articles and label each one as human or AI. The group got 299 out of 300 right. Zero false positives. They beat almost every commercial AI detector on the market, including on text that had been paraphrased and "humanized" specifically to fool them.

The researchers also documented exactly what gave the AI text away. This skill turns those findings into rules Claude follows when it writes for you.

In July 2026, The Economist ran a second study with a different method. They asked ChatGPT, Claude, Gemini and Grok to rewrite their own articles, then compared 55,940 sentences and 1.2m words against their own journalism, against CNN, the New York Times and the Washington Post, and against bestselling fiction from 1950 to 2022. Version 2 of this skill folds in what they found.

The first 10 rules came from years of hands-on content work and pattern spotting. Rules 11 through 18 came from the 2025 study. Rules 19 through 25 came from the 2026 one. Every rule is tagged with its source and year.

Why this has a version number

The tells decay.

Between the two studies, twelve months apart, the banned-word list from version 1 largely expired. Models stopped delving. The tapestries went away. What replaced them is harder to spot, because the new offenders look ordinary: significant, increasingly, consequences, implications.

The em dash finding reversed outright. Version 1 banned em dashes because AI overused them, which was true when I wrote it. By 2026, ChatGPT was using markedly fewer em dashes than human writers, and only Claude still used more. The rule survives in version 2, but for an honest reason: readers have been trained to read the dash as a robot's fingerprint, so it costs you something with a human audience even though it no longer marks you for a detector.

Models learn from human writing and human feedback. They pick up what people find impressive and drop what people call out. So any static list of forbidden words has a shelf life, and the structural rules have held up far better than the vocabulary ones. That's why the word lists here are dated and the structure rules aren't.

Two rules reversed in a year. Expect more to. See CHANGELOG.md.

What it covers

Structural habits like the rule of three, formulaic transitions, and hypophora. Banned vocabulary, in two dated tiers, because the list moves. Attribution verbs, because humans just write "says" and AI can't stop writing "notes" and "emphasizes." Quote homogeneity, too-perfect grammar, safe and opinion-free content, relentlessly positive tone, generic intros, invented experts named Dr. Sarah Thompson, and yes, em dashes.

Version 2 adds the things the 2026 study measured that nobody was talking about:

  • Long words. Every model writes with far more eight-letter-plus words than journalists do. Orwell called it pretentious diction in 1946 and it's still the cleanest single separator.
  • Punctuation scarcity. This one runs against instinct. Models use fewer commas and semicolons than humans and hardly any parentheses, so the newer tell is a page with too little punctuation. Which means stripping out your em dashes without replacing them makes the problem worse, not better.
  • Sentence rhythm. Models write sentences of stubbornly similar length, at roughly half to three-quarters of human variation. The rhythm gives it away even when every sentence is fine on its own.
  • "And." The most overused word in AI writing, and the reason the sentences run long.
  • Nominalizations, scientific lingo in copy that isn't scientific, and "it's not X but Y," which Claude and ChatGPT use more than any human corpus including bestselling fiction.

Install in Claude.ai

You need a Pro plan or above with code execution turned on (Settings > Capabilities).

  1. Download anti-ai-writing-style.zip from this repo
  2. In Claude, go to Customize > Skills
  3. Click the + button and choose Upload a Skill
  4. Upload the zip and toggle it on

Claude will apply it automatically when you ask for written content. If it doesn't trigger, say "use my anti-ai-writing-style skill for this."

Install in Claude Code

Copy the anti-ai-writing-style folder into ~/.claude/skills/ for use across all projects, or into .claude/skills/ inside a specific repo to share it with everyone working in that codebase. No zip needed.

Repo contents

anti-ai-writing-style/
  SKILL.md                 The skill itself
anti-ai-writing-style.zip  Ready to upload to Claude.ai
CHANGELOG.md               What changed in v2, and what reversed
README.md                  You are here

Who made this

Hi, I'm AMA (Ann Marie Almariei). Partner & Creative Orchestrator at Sageworx, 20+ years in creative and content strategy, currently building with AI every day. I made this skill because I write for a living and refuse to ship copy that sounds like a robot.

Credit

Rules 11 through 18: Jenna Russell, Marzena Karpinska, and Mohit Iyyer (2025), People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text, arXiv:2501.15654. Their annotated dataset and detection guide are at github.com/jenna-russell/human_detectors.

Rules 19 through 25: How to spot AI writing, The Economist, July 30, 2026. Findings described here in my own words; go read theirs, the charts are worth it.

Rule 19 also owes a debt to George Orwell, Politics and the English Language (1946), who diagnosed the long-word problem eighty years before anyone had a language model to blame it on.

This skill isn't affiliated with or endorsed by any of the above.

Feel free to fork it, adapt it, or add your own rules. If you find a new AI tell in the wild, that's rule 26 waiting to happen.