A Claude Code plugin

A red pen for robot prose.

It reads your text the way an editor would. It finds the phrases that give away a machine wrote them, names each one, and hands back a rewrite that sounds like a person.

Install the plugin Read the source ↗

live markup

In today's fast-paced world,stock opener leveraging using data is no longer just a luxury — it's a necessity.antithesis + pivot em-dash Let's delve into how teams can unlock their full potential.two AI verbs Here's what the teams that act on their numbers do differently.

VERDICT: Heavy AI fingerprint — 5 strong tells across openers, vocabulary, and constructions.
AI-FEEL SCORE: 71 / 100 — strong AI signature · COUNTS: Strong 4 · Medium 1 · Weak 1

Eleven kinds of tell

Every scan reads the full catalog, drawn from Wikipedia's "Signs of AI writing," GPTZero, Grammarly, Pangram, and the Kobak et al. study on excess vocabulary in post-ChatGPT abstracts.

Telltale vocabulary

Delve, leverage, robust, landscape, tapestry: the words machines reach for.

II.

Stock openers

"In today's fast-paced world," "Let's dive in," and their endless cousins.

III.

Signature constructions

Antithesis, fake-depth formulas, false agency, self-answered questions.

IV.

Hedges & throat-clearing

"It's worth noting that," "Make no mistake": emphasis without content.

Meta-commentary

"Plot twist:", "a feature, not a bug," and other narrator asides.

VI.

Closers & conclusions

The grand "this is the future of work" sign-off that says nothing.

VII.

GPT response artifacts

The residue of a chat assistant left behind in finished prose.

VIII.

Punctuation & rhythm

Em-dash density, metronome sentences, formulaic fragments.

IX.

Voice & content

Vague declaratives, narration from a distance, agentless passive.

Structural tells

Uniform paragraphs and lists that betray a template underneath.

XI.

Engagement bait

The "let that sink in" rhetoric built to farm a reaction.

→

Every instance, flagged

Not one example per category. Every span it finds, scored and rewritten.

One number to watch fall

Findings roll into a single 0–100 AI-feel score, so you can edit a draft and watch the evidence drop in real time.

0204065100

0–20 reads human

21–40 mostly human, minor tells

41–65 AI-assisted feel

66–100 strong AI signature

It's a density of evidence, not a verdict on who wrote it. Detectors are unreliable; GPTZero once rated the Bible 88% AI. The score exists to make drafts comparable across edits, never to accuse. When a tell might be intentional, it's marked as such.

Clone it and go

Pull the repo and point Claude Code at it. Then just ask.

From GitHub

Straight from the repo

Clone the skill, load it into Claude Code, then ask Claude to "de-AI this draft." That's the whole setup.

# clone the repo
git clone github.com/
  ananas-agency/ai-pattern-detector

# load it into Claude Code
claude --plugin-dir
  ./ai-pattern-detector

Built on the research

The pattern catalog synthesizes published detection work, not vibes.

Wikipedia · Signs of AI writing Pangram Labs GPTZero · AI Vocabulary Grammarly Originality.ai Kobak et al. 2024 · arXiv:2406.07016 Walter Writes AI

Questions

The honest answers, including the ones about what it can't do.

Does it tell me whether a text was written by AI?

No. That's deliberate. It measures the density of AI-writing patterns and rolls it into a 0–100 "AI-feel" score, as a guide for editing. It's not a verdict on authorship. Detectors are unreliable (GPTZero once rated the Bible 88% AI), so it never accuses. It just shows you what reads like a machine.

Will it flag writing I wrote myself?

It can. Many tells show up in human writing too, which is why it scores severity rather than issuing pass/fail, and skips deliberate choices: domain jargon, parenthetical em-dashes, passive voice in scientific or legal registers. When a finding might be intentional, it says so in the verdict.

Where does my text go? Is it private?

It runs inside Claude Code, in your own session, on your machine. The plugin adds no external service and sends your text nowhere extra. It reads what you give it and replies in the same conversation.

What can it scan?

Pasted text, a file, a URL, or a draft already in your conversation. When a source mixes prose with code, config, or markup, it scans only the human-facing prose and leaves the rest alone.

Which languages does it support?

English. The skill is built on English phrases and patterns. The vocabulary and openers it looks for are all English AI tells, so it's designed for scanning English text.

Does it only run in Claude Code?

The install steps on this page are for Claude Code, where it ships as a plugin. Underneath, it's a standard Agent Skill (a single SKILL.md with its pattern catalog), so the same skill runs on other Claude surfaces that support skills, including the Claude apps and the Agent SDK. There you add it as a skill rather than installing the plugin. The detector isn't locked to Claude Code; only this packaging is.

Does it rewrite for me, or just point at problems?

Both. Every finding comes with a concrete before → after rewrite, plus a deep-dive that suggests a couple of ways to fix each one rather than a single imposed edit. At the end it offers to apply all the fixes and hand back the clean full text.

Is it free?

Yes. It's open source under the MIT license. Clone it from GitHub and load it into Claude Code.

An open invitation

We've all had enough of AI slop.

If you want to make the internet a better place to read, help build this skill. You're welcome here. Together, we can do it.

Contribute on GitHub →