Source: blader/humanizer · Maintainer: community: Siqi Chen (blader) · License: MIT · Checked: 24 October 2026
What is the humanizer skill?
The humanizer skill rewrites AI-sounding text so it reads like a person wrote it, without changing what it says: it marks tells such as "it's not just X, it's Y" contrasts, one-line closers and forced triads, drafts a rewrite, checks it, and returns a final version. It works in Claude Code, Codex, Claude.ai and other agents that support skills.
It lives in blader/humanizer, a single-skill repo with about 60,000 GitHub stars as of 24 October 2026. The LICENSE names Siqi Chen as the copyright holder. The repo had commits as recently as 16 October 2026, and the current version is 3.1.0.
The skill's description opens: "Rewrite AI-sounding text so it reads like the writer without changing what it says." It continues with the tells it targets and ends: "Based on Wikipedia's "Signs of AI writing.""
As of 24 October 2026, its skills.sh page listed about 10,400 installs (snapshot, 24 October 2026). That counts skills CLI installs only, not plugin or ZIP installs.
What does it do?
The SKILL.md starts with a theory of why AI text sounds the way it does: a model picks the choice that fits the widest range of readers, while a person writes for one reader and one subject. Every pattern is a form of that default.
It then runs four steps:
- Mark the tells. Read the whole text and mark every pattern, strongest first, including paragraph-level shapes.
- Draft the rewrite. Keep every supported claim. Structure can change; facts cannot be added.
- Check the draft. Look for remaining tells and for any fact, number or claim that was added or dropped.
- Write the final version. Restate points naturally instead of patching flagged phrases one at a time.
The 26 patterns are grouped in six sections and numbered by strength:
| Section | Examples of patterns |
|---|---|
| A. Staging instead of stating | Not X but Y, one-line closers, sayings that sound deep, staged run-ups, arguing with no one |
| B. Rhythm by rule | Forced triads, repeated openings, dashes everywhere, stacked qualifiers |
| C. Inflation and borrowed authority | Overused AI words, inflated significance, sales language, "experts believe" |
| D. Formatting by rule | Decorative bold, decorative headings, curly quotes |
| E. Leftovers from the chat | "Great question!", knowledge-limit disclaimers, writing about the document |
| F. Writing for the wrong reader | Re-explaining what the reader already knows |
Patterns 1 to 5 justify an edit on one sighting. Patterns marked "weak alone", such as passive voice or curly quotes, only count when several tells share a passage.
It has three output modes. For pasted text it returns the draft, a short list of remaining patterns, and the final rewrite. Given a file, it writes only the final text and leaves code blocks, commands, paths, YAML metadata and link targets unchanged. Used inside another task, such as writing a PR description, it returns only the final text.
The repo holds the root SKILL.md, plugin manifests for Claude Code and Cursor, agents/openai.yaml for Codex, a changelog and a validation script for maintainers.
How do you install it?
All routes below are from the README.
Claude Code plugin:
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
The README says the plugin answers to /humanizer:humanizer and "needs Claude Code 2.1.142 or newer". On older versions it gives:
npx skills add blader/humanizer --global --agent claude-code
That installs to ~/.claude/skills/humanizer/, the personal skill location in the Claude Code skills docs. Drop --global to install for the current project only.
Codex:
npx skills add blader/humanizer --global --agent codex
Every agent the skills CLI supports:
npx skills add blader/humanizer --global --agent '*'
skills.sh shows the plain form, npx skills add https://github.com/blader/humanizer --skill humanizer. Codex and Cursor read the same SKILL.md format; see Codex skills and Cursor skills.
Claude.ai and Claude Desktop. Download the repo as a ZIP from GitHub (Code, then Download ZIP) and upload it as a skill. Per Claude's help article on skills, uploads go through Customize, then Skills, and need code execution turned on.
How do you use it?
Type /humanizer and paste text, ask in plain language, or give it a file path.
Example prompts (suggestions, not quotes from the repo):
/humanizerfollowed by a pasted launch announcement.Humanize the prose in docs/launch-post.md. Leave the code samples alone.Here are three paragraphs I wrote last year. Match that voice and humanize this draft: [draft]Rewrite this PR description with the humanizer skill. Keep every technical detail.Check this cover letter for AI tells and list them before rewriting.
For voice matching, the README suggests pasting 2 to 3 paragraphs of your own writing first. The sample overrides the skill's rules, so if you use dashes, it keeps them.
Requirements: none. It is a single Markdown file with no scripts, packages or API keys.
Tips and limits
From the files and pages linked, not from testing:
- Security audits on skills.sh. As of 24 October 2026, it showed Pass from Gen Agent Trust Hub, Socket and Snyk, audited 17 October 2026.
- Not a detector bypass. The README states: "Getting past AI detectors is not a goal, and detectors still flag most of its output."
- It won't invent detail. If a sentence needs a name or number the source doesn't have, the skill asks rather than making one up. Fiction is the exception.
- It knows when not to act. The SKILL.md leaves quotations, titles and proper names alone, keeps salutations and sign-offs, and notes that text written before November 30, 2022 is not AI-written.
- Context cost. The maintainers' notes say every word of SKILL.md is read on each use and cap it at 5,500 words; the current file is roughly 5,200.
- Results claim. The README cites a blind test in issue #229 where judges preferred its rewrite 16 times out of 16. A user ran that test and posted it as a GitHub issue; it is not a formal or peer-reviewed study.
- License. MIT: copy and change it as long as you keep the copyright notice.
Save it in your Kipt library
In Kipt, open Skills and paste the repo link into the import box. You can also drop the folder or a .zip. Your skills live in your own private GitHub repo and install into Claude Code or Codex with one command:
npx skills add https://github.com/<you>/<repo>.git --all -g
A saved copy is also a good base for a house style: add your team's banned words to it in your own repo. Kipt is not affiliated with blader/humanizer.
Similar skills
- Doc co-authoring skill: Anthropic's workflow for drafting documents with Claude, before you humanize them.
- Internal comms skill: templates for status updates and newsletters in your company's format.
- Brand guidelines skill: Anthropic's skill for applying brand colors and fonts, a model for encoding your own house style.