Humanizer rewrites 26 AI writing tells for human readers, and does not try to beat AI detectors

Built on Wikipedia's Signs of AI writing, not on detector scores

Humanizer takes its checklist from Wikipedia's Signs of AI writing, the guide Wikipedia editors use to catch AI-generated text. Its target is a human reader. The project says outright that getting past AI detectors is not a goal and that detectors still flag most of its output. That is the first thing to know before installing it: if a school, client or platform runs a detector on your text, Humanizer will not change the verdict in most cases. What the project does claim is reader preference. In a blind test recorded in issue #229, judges preferred Humanizer's rewrite over the original AI text 16 times out of 16. That is a single small test, filed in the project's own tracker, so treat it as an example rather than a measured accuracy figure.

Five tells trigger an edit on sight, weak ones need company

The skill checks 26 patterns, numbered by strength and frequency and grouped into families such as staging instead of stating and rhythm by rule. The first five justify an edit on a single sighting: Not X but Y constructions such as It's not just a feature, it's a shift; one-line closers such as Let that sink in; sayings that sound deep, such as At its core, what really matters is; a staged run-up such as Here's the thing or Honestly?; and arguing with no one, such as I'm not saying X, but. Lower in the list, forced triads and repeated sentence openings get rewritten to the number of items or the subject the meaning needs. Some patterns are marked weak alone, including dashes used as a universal connector, and count only when several tells share a passage, because a careful writer may use any one of them on purpose.

It asks for a missing fact instead of inventing one

The workflow runs in stages. Humanizer marks every tell it finds, strongest first, drafts a rewrite without treating the original structure as fixed, checks that draft against both the patterns and the original claims, and then writes the final version. When you paste text, it shows its work: the first rewrite, a short critique of anything that still sounds artificial, and the final version. The rule that matters most is about facts. Any name, number, date, quote or citation must come from the source text or the writer. If a sentence needs a detail that is missing, Humanizer asks rather than filling it in. The consequence is that a rewrite of vague marketing copy can come back with questions instead of a finished paragraph, and you have to supply the specifics yourself.

The Claude Code plugin needs version 2.1.142 or newer

In Claude Code, Humanizer installs from its own plugin marketplace:

/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

Installed as a plugin, it answers to /humanizer:humanizer rather than the plain /humanizer that other installs use. The plugin requires Claude Code 2.1.142 or newer. On older versions, install it through the Skills CLI instead:

npx skills add blader/humanizer --global --agent claude-code

Codex uses the same CLI with a different agent value:

npx skills add blader/humanizer --global --agent codex

For Claude.ai and Claude Desktop there is no command. You download the repository as a ZIP from the Code menu on GitHub and upload it as a skill in Settings, which also means updates do not arrive on their own; you repeat the download when a new release lands.

--global and the asterisk agent decide where the skill lands

One command installs Humanizer for every agent the Skills CLI supports, including Gemini CLI, GitHub Copilot and Windsurf:

npx skills add blader/humanizer --global --agent '*'

The --global flag is the switch between scopes. With it, the skill is available everywhere for that agent; leave it off in any of the commands and the skill installs only in the current project. That choice matters on a team: a project install travels with the repository context you are working in, while a global one changes every session on your machine. For an agent the Skills CLI does not know, the fallback is manual: copy SKILL.md into that agent's skill folder. Because the skill is plain Markdown, any harness that supports skill-style instructions can read it, and the repository also carries .claude-plugin/ and .cursor-plugin/ directories for those hosts.

Pointed at a file, it edits prose and leaves code and frontmatter alone

There are three ways to call it. Invoke the skill and paste text, ask in plain language, or give it a file path:

/humanizer

[paste your text here]

Please humanize this text: [your text]

Humanize the prose in docs/launch-post.md

In file mode it changes only the prose and leaves code, data, frontmatter and link targets untouched, which makes it usable on documentation and blog sources that mix Markdown with code samples. Register also stays put: personal writing keeps the writer's opinions and quirks, while technical and reference prose is kept neutral and plain. It will not turn a reference page into a chatty one, and it will not strip a personal essay of its opinions.

Voice matching copies your quirks, dashes included

To make the rewrite sound like you, include a sample of two to three paragraphs of your own writing before the AI text:

/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]

Humanizer then follows the sample's rhythm, word choice, punctuation and deliberate quirks, including dashes if you use them. That creates a trade-off worth knowing: the dash pattern is one of its own tells, so a sample full of dashes will keep them in the output. If your house style forbids a habit, remove it from the sample first, because the skill treats whatever the sample contains as intended.

Version 3 arrived within six weeks of 2.11.1

Releases have moved quickly: v2.11.1 on 2026-08-18, v3.0.0 on 2026-09-06 and v3.1.0 on 2026-09-28, with the last push on 2026-10-28. The repository is MIT licensed and not archived, and it keeps a CHANGELOG.md alongside SKILL.md and AGENTS.md. A major version bump in a skill means the instructions themselves changed, so output on the same input can differ between versions. Teams that review AI-edited copy should note which version produced a draft, and re-check a few known samples after upgrading, especially for ZIP installs on Claude.ai that stay on whatever version was uploaded.

Editorial conclusion

Humanizer suits writers and teams who want AI-drafted prose edited toward plain, specific human writing inside Claude Code, Codex or another skills-capable agent. It is the wrong tool for anyone whose goal is passing an AI detector, since detectors still flag most of its output. Before relying on it, check your Claude Code version is 2.1.142 or newer for the plugin, and review its critique step on a sample of your own text.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes