Every developer using Claude Code, Cursor, or Copilot has seen it: the prose that comes back reads like a press release. “It’s not just a feature, it’s a catalyst for innovation.” “Let that sink in.” These patterns are so common they’ve been catalogued — by Wikipedia, of all places. blader/humanizer is an open-source agent skill that checks your AI-generated text against 26 documented patterns and rewrites them out. It’s sitting at 69.7K GitHub stars and v3.1.0 shipped in September.
Built on Wikipedia’s AI Writing Field Guide
What separates humanizer from detector tools and generic paraphrasers is its source material. The 26 patterns come directly from Wikipedia’s “Signs of AI Writing” — a list built by WikiProject AI Cleanup over thousands of hours reviewing AI-submitted articles. The tells aren’t ML-detected or inferred from embeddings. They’re named, documented, and human-readable: you can read each rule, understand it, and argue about it if you disagree.
That’s a meaningful distinction. AI detector services operate as black boxes — high confidence score, no explanation. Humanizer tells you exactly which pattern triggered and why. It rewrites surgically rather than paraphrasing wholesale, which means the factual content and your intended structure stay intact.
The Five Tells It Hits Hardest
According to the README, the five strongest AI writing tells are: “Not X but Y” constructions (“It’s not just a tool, it’s a paradigm shift”), one-line closers (“Let that sink in”), deep-sounding sayings (“At its core, what really matters is…”), staged run-ups (“Here’s the thing” / “Honestly?”), and arguing with no one (“I’m not saying X, but…”). The skill checks 21 more beyond those five.
The practical difference is visible in a paragraph. Here’s the kind of text humanizer is designed to catch:
Before: “This groundbreaking tool boasts a vibrant interface, offering a seamless experience that fosters collaboration and enhances productivity. It’s not just a tool — it’s a catalyst for innovation.”
After: “The app has a shared workspace where your team can edit the same document at the same time. It saves about 2 hours per week compared to emailing files back and forth.”
Specific over vague. Claims replaced by measurements. The second version trusts the reader to draw conclusions. In a blind test run by the project, judges preferred the humanized version over raw AI output 16 times out of 16.
How It Actually Works
Humanizer is a single Markdown file you install as an agent skill — no API calls, no external service, no cost beyond your existing coding assistant. It runs in Claude Code (2.1.142+), Cursor, Codex, Copilot, Gemini CLI, and Windsurf. Install takes one command:
# Claude Code (plugin method)
/plugin marketplace add blader/humanizer
# Or via skills CLI for any agent
npx skills add blader/humanizer --global --agent '*'
Once installed, run /humanizer. If you want it to match your voice rather than produce generic “natural” text, paste 2–3 paragraphs of your own writing first. The skill rewrites paragraph by paragraph, then runs a second audit pass for anything it missed in the first pass. MIT licensed, source on GitHub.
What v3.1.0 Added
The September release added Pattern #26, targeting a specific AI behavior: replies that re-explain context the reader already gave. If you ask a question and the AI response starts by summarizing what you just said back to you, that’s Pattern #26. The release also extended Pattern #25 to catch text that describes its own sourcing or layout (“In this article, I will cover…”), added “headings written for effect” to Pattern #20, and restructured the README to lead with the before/after example and five strongest tells. Version history was moved to a separate CHANGELOG.md.
The Anti-Slop Skill Layer
Humanizer isn’t alone. It’s part of a growing layer of “anti-slop” skills for coding agents. Hallmark by Together AI (46.8K stars) does the same job for UI design — it enforces that AI-generated interfaces look intentionally designed rather than template-stamped. A Chinese-language fork, Humanizer-zh by op7418, has 21.9K stars. The pattern is consistent: as developers use agents for more content, tools to audit and correct the AI patterns in that content become essential infrastructure.
The honest limitation: humanizer strips verbal tells, not shallow thinking. If the underlying content is vague, removing the hedges and tricolons exposes the vagueness rather than fixing it. Use it after you’ve verified the substance — not as a substitute for thinking through what you’re trying to say.

