Text from Claude gives itself away in the first line: "in today's digital world", an em dash every other word and the eternal three-item lists. One free skill cleans that up automatically ✍️
What it is and who it's for
Humanizer (the blader/humanizer repo) is an open-source skill for Claude Code that removes the "AI handwriting" from your text. Inside are 33 named AI-writing patterns collected from Wikipedia's "Signs of AI writing" guide (maintained by WikiProject AI Cleanup on the basis of thousands of real AI-text examples). The skill doesn't do "rewrite it nicely" — it works like an editor: finds a specific pattern → fixes it → then in a second pass asks itself "what still reads as AI?" and cleans that too.
This is the tool from those viral posts about "how to make Claude always write like a human — free and automatic". Posts often mangle the name: "Blader Humanizer" is the author's handle (blader) plus the skill name; the skill itself is just Humanizer.
Who it's for: anyone shipping text that came out of an AI — bloggers, marketers, newsletter, docs and landing-page authors. Any level: it installs in a minute and then works on its own.
Key capabilities
- 33 patterns in 5 groups: content (significance inflation, vague attributions like "experts say"), language and grammar (AI vocabulary such as delve/underscore, avoiding a plain "is", synonym carousels, the "rule of three"), style (em-dash overdose, emoji, bold headers), communication (chatbot artifacts) and fillers/hedges (filler openers, excessive hedging, obligatory upbeat endings).
- Matches your voice: give it 2–3 paragraphs of your own real writing and the skill picks apart your sentence rhythm, pet words and habits, then rewrites the AI text in YOUR style rather than a "neutral clean" one.
- Two-pass cleanup: after fixing patterns it separately hunts for whatever still reads as AI and cleans it.
- Portability: the working artifact is a plain
SKILL.md(Markdown). The skill runs in any agent that supports skills, and you can even paste its text into ChatGPT as instructions — same trick. - MIT license, ~29k GitHub stars in six months (repo created in January 2026, actively updated).
Get started in 5 minutes
- In Claude Code install it as a plugin:
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
Or cross-agent: npx skills add blader/humanizer (with --agent '*' it goes into every configured agent at once). Or fully by hand: copy SKILL.md into your skills folder.
2. Call /humanizer and paste the text — or just ask: "humanize this text: …".
3. Want your own voice — give it 2–3 paragraphs of your writing first, then the AI text to rewrite.
When to use it, when not to
✅ Take it if Claude writes your posts, articles and emails and readers smell the AI from the first paragraph. ✅ Take it if you want a free open-source technique inside your own agent rather than yet another paid "AI humanizer" service. ❌ Don't take it as an "anti-detector" for turning in university papers: the skill cleans style, it doesn't guarantee getting past detectors — Wikipedia's own guide stresses that the patterns are signs, not proof. ❌ Skip it if the job is writing from scratch to a brief in ChatGPT: the Write For Me (GPT) card in this section is closer to that. ❌ Want to package YOUR style into a permanent skill rather than cleaning after the fact → build it with skill-creator.
The honest price
Completely free and open source (MIT) — no card required, commercial use allowed. You only pay for tokens on your own plan/API: the skill is a prompt instruction, no extra services or keys. Tag: free / oss.
Gotchas
- The patterns were collected from English material (Wikipedia's guide is about English text). In other languages vocabulary markers like "delve" won't apply, but the structural ones — em-dash overdose, the rule of three, bold header-lists, inflated bureaucratic significance — transfer beautifully. Check the result with your own eyes.
- Without a voice sample the skill produces a "neutral human" style and can over-dry the text: it will cut em dashes and emoji even if those are your signature. Fixed by giving it a sample of your writing.
- It's a prompt, not an algorithm: quality depends on the model and the result isn't deterministic. On important texts, do a final read-through yourself.
- Style ≠ substance: the skill doesn't fix hallucinations, emptiness or factual errors — only HOW something is written.
- The repo is very much alive (last push — June 2026, ~60 open issues): the pattern set and command shapes change between versions — check the README.
🤖 Prompt accelerator
"Here are 2–3 paragraphs of my own real writing — study the rhythm, sentence length and my pet words:
. Now apply the humanizer skill to the text below: find and remove AI-writing patterns (significance inflation, AI vocabulary, em-dash and bold overdose, the rule of three, filler openers, excessive hedging), and rewrite it in my voice, not in a "neutral clean" one. Then do a second pass: ask yourself what STILL reads as AI and clean it up. Don't change the meaning or the facts. Text: "