Is the Humanizer skill official from Anthropic?

No. It is a community skill by Siqi Chen (GitHub user blader), released under the MIT license. Anthropic's own anthropics/skills repository does not include a humanizer.

Will it make text pass AI detectors like GPTZero or Turnitin?

The repo makes no such claim and publishes no detector benchmarks. It removes 35 stylistic tells drawn from Wikipedia's "Signs of AI writing" list; whether that moves any given detector's score is untested by the project itself.

Does it work outside Claude Code?

Yes. It is one Markdown file, so any agent that supports skills can load it. The Skills CLI (`npx skills add blader/humanizer`) can target other agents via `--agent`, and Claude Desktop accepts the repo ZIP as a skill upload.

Does it change what the text says?

It is instructed not to. Its first rules are to keep every claim and invent no facts: names, numbers, dates, and quotes must come from the source or the writer, and it should ask for missing details rather than fabricate them. Verify the output anyway; enforcement is prompt-level, not code-level.

Skill review · by Siqi Chen (GitHub: blader)

Claude Humanizer Skill

The Claude humanizer skill this search points to is blader/humanizer, a single-file agent skill published by Siqi Chen in January 2026 and now at roughly 39,500 GitHub stars. It rewrites AI-sounding prose using 35 patterns taken from Wikipedia's "Signs of AI writing" page: inflated importance, sales language, em-dash overuse, forced groups of three, chatbot leftovers, filler, and the rest. It keeps every factual claim and is instructed to invent nothing. The one thing to know before installing: it is a style editor, not a detector-evasion tool. The repo never claims its output will pass GPTZero or Turnitin and publishes no detection benchmarks; separate, much smaller projects make that pitch. It is MIT-licensed, pure Markdown with zero dependencies, works in Claude Code, Claude Desktop, and any agent that supports skills, and was last pushed 2026-08-19. Several other humanizer skills exist; this is the one with the traction and a maintained release history.

View the sourceMIT licensed · 39,519★ · facts checked September 1, 2026

By Robert MacKelfresh

Founder, Automation Squad · September 1, 2026

Install

1Skills CLI (any supported agent): run `npx skills add blader/humanizer --global`. Drop `--global` to install only in the current project; add `--agent '*'` to install for every supported agent, then reload skills.
2Claude Code 2.1.142 or newer, plugin route: run `/plugin marketplace add blader/humanizer`, then `/plugin install humanizer@humanizer`. The command becomes `/humanizer:humanizer`.
3Claude Desktop / Claude.ai: download the repository as a ZIP from GitHub and upload it as a skill under Settings, Capabilities, Skills. This works as of v2.11.2, which removed the plugin symlink that previously broke the ZIP.
4Manual install: copy SKILL.md into your agent's skills directory, for example ~/.claude/skills/humanizer/SKILL.md for Claude Code.
5Invoke it with `/humanizer` followed by pasted text, ask in plain language ("humanize this text"), or give it a file path.

What it does

Pattern removal: Checks text against 35 documented patterns from Wikipedia's "Signs of AI writing" (maintained by WikiProject AI Cleanup): inflated importance, sales language, vague sources, stock AI words, em dashes, forced groups of three, bold mini-headings, chatbot artifacts, filler, and hedging.
Two-pass rewrite: Makes a first rewrite without treating the original structure as fixed, then critiques its own draft against the pattern list and the original claims, and rewrites whatever still sounds artificial. It shows both the first pass and the critique before the final version.
Fact preservation: Hard rule against invented facts: no name, number, date, quote, or citation may be added unless it comes from the source or the writer. If a sentence needs a missing detail, the skill is told to ask for it rather than fabricate it.
Voice matching: Paste two or three paragraphs of your own writing and it follows that sample's rhythm, word choice, and punctuation. The sample overrides the default style rules, including the em-dash ban.
File-aware editing: Point it at a file path and it edits only the prose, leaving code, data, frontmatter, and link targets untouched.

Common questions

Is the Humanizer skill official from Anthropic?: No. It is a community skill by Siqi Chen (GitHub user blader), released under the MIT license. Anthropic's own anthropics/skills repository does not include a humanizer.
Will it make text pass AI detectors like GPTZero or Turnitin?: The repo makes no such claim and publishes no detector benchmarks. It removes 35 stylistic tells drawn from Wikipedia's "Signs of AI writing" list; whether that moves any given detector's score is untested by the project itself.
Does it work outside Claude Code?: Yes. It is one Markdown file, so any agent that supports skills can load it. The Skills CLI (`npx skills add blader/humanizer`) can target other agents via `--agent`, and Claude Desktop accepts the repo ZIP as a skill upload.
Does it change what the text says?: It is instructed not to. Its first rules are to keep every claim and invent no facts: names, numbers, dates, and quotes must come from the source or the writer, and it should ask for missing details rather than fabricate them. Verify the output anyway; enforcement is prompt-level, not code-level.

Before you install

Unofficial. This is a community skill by one author; it is not from Anthropic and is not in the anthropics/skills repo.
No detector-evasion claim, no benchmarks. The docs describe style cleanup based on Wikipedia's pattern list; the project publishes no evidence that output scores lower on GPTZero, Originality.ai, or Turnitin. If beating a detector is the goal, this skill does not promise that.
It is one Markdown prompt with no code, so results depend entirely on the model following roughly 35 rules; output varies by model and by run.
Opinionated style rules: it strips em dashes, curly quotes, emojis, bold mini-headings, and title-case headings. Several of these collide with common house styles, and overriding them requires supplying a writing sample.
The spec churns fast: the pattern count went from 24 to 35 across seven months of releases, so behavior shifts between versions.
The Claude Desktop ZIP install was broken by a plugin symlink until v2.11.2 (repo issues #224 and #202).
The plugin install path requires Claude Code 2.1.142 or newer; older versions need the CLI or manual copy.
Its source list is a volunteer-maintained Wikipedia project page, not a peer-reviewed corpus, and that page keeps changing.

Facts come from the skill's own repository and docs on the date above, via the GitHub API — not from memory, and never from a sponsor.

Worth comparing

Aboudjem/humanizer-skill — 209 stars; 55 patterns, 5 selectable voices, and a 0-100 AI-tell score; pure Markdown; last pushed 2026-09-01.
jpeggdev/humanize-writing — 56 stars; a simpler Claude Code rewrite skill; last pushed 2026-03-14.
humanizerai/agent-skills — 42 stars; explicitly markets AI-detector evasion (GPTZero, Turnitin), a claim blader/humanizer deliberately does not make; last pushed 2026-02-05.

← All MCP servers & skills