Humanizer SkillHumanizer skill: make AI text read like a person wrote it
Humanizer is a free skill for Claude Code, Codex and other AI agents. It finds 26 signs of AI writing in a draft and rewrites them, without adding facts or dropping any.
Quick install
Run this in your terminal. It finds the coding agents you have installed and adds the skill to each:
npx skills add https://github.com/blader/humanizer --skill humanizerUsing claude.ai or the Claude Code plugin instead? See the install page.
Before and after
This is the example from the project's README. The rewrite keeps the facts (live co-editing, anchored comments, free on every plan) and loses the announcement voice.
I'm thrilled to announce that shared drafts are finally here! 🚀 For months, our own team was drowning in files named final_v7.docx — and we knew there had to be a better way. Now two people can edit the same doc at once, with every change appearing live for both of them. It's not just a feature; it's a whole new way to collaborate. Comments stay anchored to the exact sentence they reference, even as the text around them evolves. And the best part? It's available today on every plan, completely free. Let that sink in.
Shared drafts are out today. For months our own team passed around files named final_v7.docx, so we built a way for two people to edit the same doc at once, with each person's changes showing up live for the other. Comments stay pinned to the sentence they're about, even when the text around them changes. It's free on every plan.
In a blind test run by the project, judges preferred Humanizer's rewrite over the original AI text 16 times out of 16 (issue #229). More pairs are on the examples page.
What the Humanizer skill is
A skill is a set of written instructions an AI agent loads when a task calls for it. Humanizer is one Markdown file, SKILL.md, written by Siqi Chen (@blader). It tells the agent how to spot AI writing habits and how to edit them out.
The pattern list comes from Signs of AI writing, the page Wikipedia editors use to catch machine-generated edits. Humanizer turns that page into editing rules, each with a watch list, the reason it reads as artificial, and a before/after pair.
The skill edits for human readers. It does not try to fool AI detectors, and the author says detectors still flag most of its output. If you want text that reads well to people, it fits. If you need a detector score, see how it compares with AI humanizer tools.
The five strongest tells
Humanizer rewrites any of these on a single sighting. The other 21 patterns need more evidence before it acts.
1. Not X but Y
"It's not just a feature, it's a shift." The negative half names something nobody claimed.
2. One-line closers
"Let that sink in." A short paragraph that repeats the one before it.
3. Sayings that sound deep
"At its core, what really matters is..." An ordinary point dressed as a hidden truth.
4. A staged run-up
"Here's the thing." "Honestly?" The writer announces the point instead of making it.
5. Arguing with no one
"I'm not saying X, but..." A defense against an objection that appears nowhere else.
+21 more
Forced triads, dashes, AI vocabulary, inflated significance, chatbot residue and more. See all 26 patterns.
How it works
A language model picks the most likely next word, so it drifts toward phrasing that suits any reader and any topic. A person writes for one reader about one thing. The skill treats every tell as a symptom of that default choice and edits in four passes:
- Mark the tells. Read the whole text and flag every pattern, strongest first, including paragraph-level shapes such as three parallel examples.
- Draft a rewrite. Keep every supported claim. Structure can change; facts cannot. If a sentence needs a detail the source lacks, ask for it.
- Check the draft. Look for anything that still sounds generated and for any fact that was added or lost.
- Write the final version. Restate each point plainly and vary sentence length.
With pasted text you get the draft, a short critique and the final version. Give it a file path and it edits prose only, leaving code blocks, commands, frontmatter and link targets alone. When another task calls it (a commit message or PR description, say), it returns only the final text.
The 26 patterns at a glance
| Group | Patterns |
|---|---|
| A. Staging | Not X but Y · one-line closers · sayings that sound deep · staged run-up · arguing with no one |
| B. Rhythm by rule | Forced triads · repeated openings · dashes everywhere · stacked qualifiers · hyphenated pairs · passive voice |
| C. Inflation | Overused AI words · inflated significance · vague association · -ing riders · sales language · borrowed authority · avoiding "is" |
| D. Formatting | Decorative bold · decorative headings · curly quotes |
| E. Leftovers | Chatbot residue · knowledge-limit disclaimers · heading repeated in first line · writing about the document |
| F. Wrong reader | Re-explaining what the reader already knows |
Read each pattern with its watch list and examples, or paste a draft into the free AI writing checker to see which ones it trips.
Matching your own voice
Removing tells gets text halfway. To make it sound like you, give the skill a sample first:
/humanizer
Here's a sample of my writing for voice matching:
[paste 2-3 paragraphs of your own writing]
Now humanize this text:
[paste AI text to humanize]The sample overrides the default rules. If you use dashes, the rewrite keeps them at about your rate.
Humanizer skill FAQ
What is the Humanizer skill?
An open-source agent skill that rewrites AI-written text so it sounds like a person wrote it. It checks for 26 patterns from Wikipedia's Signs of AI writing guide and keeps every fact from the original.
How do I install the Humanizer skill in Claude Code?
Start Claude Code with claude in a terminal, then type /plugin marketplace add blader/humanizer and /plugin install humanizer@humanizer at its prompt. From a regular shell, use claude plugin marketplace add blader/humanizer and claude plugin install humanizer@humanizer. Details are in the install guide.
How do I add it to claude.ai? The /plugin commands did nothing.
The /plugin commands only work in Claude Code. In claude.ai or the desktop chat, turn on code execution under Settings → Capabilities, then upload humanizer.zip under Customize → Skills (+ → Create skill → Upload a skill).
Is the Humanizer skill free?
Yes, it is MIT licensed. You pay only for the model usage of whatever agent runs it.
Does it bypass AI detectors like GPTZero or Turnitin?
No. It is built for human readers, and the author notes that detectors still flag most of its output. More on that here.
Which agents does it work with?
Claude Code, Codex, Claude.ai, Claude Desktop, and every agent the Skills CLI supports, including Gemini CLI, GitHub Copilot and Windsurf. Anything else that reads skill folders can use a copy of SKILL.md.
Will it change my facts or add details?
It is told not to. Names, numbers, dates, quotes and citations must come from your text. When a sentence needs a missing detail, it asks you. Fiction is the one exception.
Can it edit files in my repo?
Yes. Say "Humanize the prose in docs/launch-post.md" and it rewrites the prose in place while leaving code, commands, paths, YAML and link targets unchanged.
Does it work on languages other than English?
The rules are written in English, but the skill says the not-X-but-Y formula appears in every language and should be handled the same way. Results in other languages depend on the model you run.