If you pay for Claude and write your own docs or posts, it is the best free humanizer I know. Its README says beating AI detectors is not a goal.
| Item | What the repository and docs show |
|---|---|
| What it is | An Agent Skill (a SKILL.md file of instructions) for Claude, Codex and other skill-aware agents |
| Maker | GitHub user blader; README lists cfo.ai as sponsor |
| License | MIT |
| Current version | 3.1.0 (CHANGELOG, read 29 October 2026) |
| Patterns checked | 26, in six groups |
| Source of the patterns | Wikipedia's Signs of AI writing, maintained by WikiProject AI Cleanup |
| Claude Code requirement | Version 2.1.142 or newer for the plugin install |
| Claude.ai plans with skills | Free, Pro, Max, Team and Enterprise, with code execution turned on (Claude Help Center) |
| Price | Free; you pay only for the Claude or Codex plan you already use |
| Detector goal | None; the README says detectors still flag most of its output |
What the humanizer skill is
A skill, in Anthropic's terms, is a folder of instructions, scripts and resources that Claude loads when a task calls for it. The Claude Help Center article What are skills? says skills work through progressive disclosure: Claude reviews the available skills, loads the relevant one and applies its instructions. Anthropic's Agent Skills overview adds that only a skill's name and description sit in context until it is triggered, so installing one costs almost nothing until you use it. The format is published as an open standard at agentskills.io, which is why the same skill also runs in Codex and other agents.
blader/humanizer is one such skill. Its job is narrow: take prose that reads as machine-written and rewrite it so it reads like a person, without changing what it says. The README is explicit about the scope: "Humanizer edits for human readers. Getting past AI detectors is not a goal, and detectors still flag most of its output."
I like that honesty. Most humanizer apps lead with a detector score; this project leads with what a reader experiences. That makes it easier to judge, because the output either reads better to you or it does not.
My take: the narrow scope is the feature. A tool that only promises better prose is one you can check yourself in a minute.
blader/humanizer on GitHub: what's inside
The repository is small. On 29 October 2026 it showed about 60,900 stars and 4,800 forks. On 29 October 2026 the main branch held:
- SKILL.md: the skill itself, about 400 lines of instructions with a name, description and version in its front matter.
- README.md: install steps, usage, the pattern tables and before-and-after examples.
- CHANGELOG.md: version history from 2.x to 3.1.0.
- .claude-plugin and .cursor-plugin: manifests that let Claude Code and Cursor install it as a plugin.
- agents and scripts folders, an AGENTS.md file and the MIT LICENSE.
How the skill works
SKILL.md sets out a four-step loop:
- Mark the tells. Read the whole text and mark every pattern, strongest first, including paragraph-level shapes.
- Draft the rewrite. Keep every supported claim. Do not add a fact, name, number, date, quote or citation unless the source or the user supplied it; if a detail is missing, ask.
- Check the draft. Read it back, ask what still sounds generated and whether anything was added or dropped.
- Write the final version, stating each point naturally rather than patching phrases one by one.
When you paste text, the skill shows its first rewrite, a short critique and the final version. When you point it at a file, it changes only prose and leaves code, data, front matter and link targets alone.
The 26 patterns
The README groups them in six sections and numbers them by strength. The first five justify an edit on one sighting; some later ones are marked "weak alone" and count only when several tells appear together.
| Group | Examples from the README |
|---|---|
| A. Staging instead of stating | "Not X but Y", one-line closers, sayings that sound deep, "Let's dive in", arguing with no one |
| B. Rhythm by rule | Forced triads, repeated sentence openings, dashes everywhere, stacked qualifiers |
| C. Inflation and borrowed authority | "delve", "testament", "pivotal moment", sales language, "experts believe", "serves as" instead of "is" |
| D. Formatting by rule | Bold as decoration, decorative headings, curly quotes |
| E. Leftovers from chat and drafting | "Great question!", knowledge-limit disclaimers, headings repeated in the first sentence |
| F. Writing for the wrong reader | Re-explaining what the reader already knows (added in 3.1.0) |
These map closely onto Wikipedia's Signs of AI writing, which the README cites as its source. The CHANGELOG shows the list was consolidated from 35 patterns to 25 in version 3.0.0 to stay aligned with Wikipedia, then grew to 26 in 3.1.0.
Voice matching
The README documents a voice mode: paste two or three paragraphs of your own writing before the text to rewrite, and the skill follows the sample's rhythm, word choice and punctuation, "including dashes if you use them." This is the most useful feature for anyone polishing their own drafts, because the output moves toward your style rather than a generic one.
What the maintainer claims
The README reports that in a blind test logged as issue #229, judges preferred the skill's rewrite over the original AI text 16 times out of 16. That is the project's own test, published in its issue tracker, not a peer-reviewed study.
My take: I'd read 16 out of 16 as a sign the maintainer cares about readers, not as proof. Sixteen judgements is a small sample, and the people judging were not chosen at random. The better evidence is the SKILL.md itself: I read it end to end, and the instructions are specific, sensible and close to what a human editor would tell you.
How to install it
Commands as listed in the README on 29 October 2026. Once installed, the skill answers to /humanizer.
Claude Code
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer
Installed as a plugin, it answers to /humanizer:humanizer. The README says this needs Claude Code 2.1.142 or newer; on older versions use:
npx skills add blader/humanizer --global --agent claude-code
Claude.ai and Claude Desktop
- On GitHub, choose Code, then Download ZIP.
- In Claude, make sure code execution is on. The Help Center article on using skills in Claude lists it under Settings, Capabilities as "Code execution and file creation".
- Go to Customize, then Skills, choose "+", then "Create skill", then "Upload a skill", and upload the ZIP.
- Toggle the skill on.
On Team and Enterprise plans an owner must allow skills in organization settings first, and Enterprise organizations can turn on skill scanning for uploaded skills.
Codex and other agents
npx skills add blader/humanizer --global --agent codex
npx skills add blader/humanizer --global --agent '*'
The second command installs it for every agent the Skills CLI supports. Leave off --global to install for the current project only.
Using it
/humanizer
[paste your text]
Or in plain language: "Please humanize this text", or "Humanize the prose in docs/launch-post.md". For voice matching, paste your sample first and label it.
A workflow for using it on your own docs
Here is the routine I'd follow for a launch post, a README or a long email you drafted with AI help. It leans on the parts of SKILL.md that protect meaning, and it keeps you in charge of every change.
- Write or draft first, humanize last. Get the facts, structure and claims right before you touch the style. The skill is built to keep claims, so it will faithfully keep a wrong one.
- Pick a voice sample. Choose two or three paragraphs you wrote without AI help, on a similar topic and for a similar reader. A product update needs a product-update sample, not a wedding speech.
- Run it on a copy. In Claude Code, point it at a branch or a duplicate file. In Claude.ai, paste the text and keep the original in another tab.
- Read the critique, not just the final text. The skill shows its first rewrite, a short critique and the final version. The critique tells you which patterns it found, which is the fastest way to learn your own habits.
- Diff the result. In a repository,
git diffshows exactly which sentences changed. Outside one, put the two versions side by side and tick off every name, number, date and quote. - Reject edits that change meaning. Watch for dropped qualifiers ("may" becoming "will"), narrowed or widened scope, and tone shifts that make a careful claim sound certain.
- Read it aloud once. If a sentence still sounds like a brochure, cut it. The skill gets you most of the way; your ear finishes the job.
My take: steps 4 and 5 are where the value is. If you skip the review and paste the final text straight into production, you have handed your judgement to a prompt.
Example
This is a sentence I wrote myself to show the kind of edit the skill aims for. It did not come from the tool.
Before (stiff, AI-sounding): "Our new dashboard serves as a pivotal tool that not only streamlines reporting but also empowers teams to unlock actionable insights, fostering a culture of data-driven decision-making."
After (plain rewrite): "The new dashboard puts last week's sales, refunds and support tickets on one screen. Most teams will spend less time building reports by hand."
The rewrite drops four patterns from the README's list: "serves as", "pivotal", the "not only X but also Y" contrast and the inflated "empowers teams to unlock" phrasing. It also replaces a vague promise with specifics. Notice that the "after" version makes a claim you must be able to back up. If you cannot, the honest edit is shorter still.
Troubleshooting
These are the problems I'd expect from reading the README, the CHANGELOG and Anthropic's skills documentation, with the fix each source points to.
| Problem | Likely cause | What to do |
|---|---|---|
/plugin install fails in Claude Code |
Claude Code older than 2.1.142 | Update Claude Code, or use npx skills add blader/humanizer --global --agent claude-code |
| Skill does not appear in Claude.ai | Code execution is off, or the skill toggle is off | Turn on "Code execution and file creation" under Settings, Capabilities, then enable the skill under Customize, Skills |
| Upload option missing on a Team or Enterprise plan | An owner has not allowed skills | Ask the organization owner to enable skills in organization settings |
| Output sounds generic, not like you | No voice sample supplied | Paste two or three paragraphs of your own writing first and label them as the sample |
| A number or name changed | Model error despite the instructions | Reject the edit, and diff every rewrite against the original |
| Code or front matter got edited | Text pasted in rather than pointed to as a file | Point the skill at the file path so it edits prose only |
| Skill fires when you did not want it | Its description matched your request | Ask for the task without words like "humanize", or toggle the skill off for that chat |
My take: almost every issue in this table comes down to setup or review, not the skill itself. Fix the setup once and keep the review habit.
Skill vs humanizer apps
| blader/humanizer skill | Web humanizer apps | |
|---|---|---|
| Cost | Free (MIT) on top of your Claude or Codex plan | Free tiers and monthly plans, which vary by vendor |
| Where your text goes | To the model provider you already use | To the app vendor, under its own privacy terms |
| Transparency | Full prompt readable on GitHub | Usually closed |
| Stated goal | Readability; detector evasion explicitly not a goal | Many market detector results; claims vary |
| Fact handling | Instructed never to invent facts and to ask when a detail is missing | Varies; a 2025 audit found weak tools adding fake citations |
| Works on files | Yes, edits prose only and leaves code and front matter | Usually paste-in only |
| Extras | None | Some include detectors, plagiarism checks, team plans or APIs |
The audit in that row is by Masrour and colleagues, who reviewed 19 humanizer and paraphrasing tools in a paper at the GenAIDetect 2025 workshop and found the weakest added hallucinated citations and nonsensical phrases. The authors work at Pangram Labs, which sells a detector. I rank dedicated apps on the same fields in my best AI humanizers ranking, and my guides to do AI humanizers work and whether AI humanizers are safe cover the research and the terms to check.
When the skill is the better choice: you already use Claude Code, Claude.ai or Codex, you want to see and edit the instructions, and you are polishing your own docs, posts or emails.
When an app may fit better: you need many languages, team word pools, a detector in the same account or an API with published tiers, and you have read the vendor's privacy and refund terms. My AI tool privacy checklist covers what to check.
When neither fits: coursework where AI use is not allowed. Schools set their own rules, and a rewrite does not change what the rule covers. My guide on humanizing vs paraphrasing explains why rewriting to beat a detector is a bad bet, and the AI humanizer cheating guide walks through real university policies.
Should you install it?
| Use it if you | Skip it if you |
|---|---|
| Already pay for Claude, Claude Code or Codex | Do not use any skill-aware agent and do not want to start |
| Polish your own posts, docs, emails or release notes | Need output in many languages with one click |
| Want to read and edit the exact instructions | Want a detector score in the same window |
| Work in a repository and want prose-only edits that leave code alone | Need a team word pool, billing and an admin console |
| Care more about readers than about any score | Hope to get coursework past an AI rule |
My pick by situation:
- Solo writer on Claude.ai: install it. Ten minutes of setup, then it is there whenever you ask.
- Developer documenting a project: install it in Claude Code and run it on docs before each release. The prose-only file handling is the standout feature for this case.
- Marketing team on a schedule: start with the skill to learn the patterns, then decide whether you need an app with shared plans. One option with a built-in detector and 80+ languages is Walter Writes AI, which I cover in the Walter Writes review.
- Student: check your course policy first. If AI help is allowed with disclosure, the skill is a fine editor for your own draft, and you still disclose.
Skill, prompt or app: how they compare
| Option | Setup time | Cost | Best for | Main limit |
|---|---|---|---|---|
| blader/humanizer skill | About 10 minutes | Free on top of your Claude or Codex plan | Your own prose inside Claude | No detector, no team billing |
| A pasted prompt in any chatbot | None | Free tiers exist | One-off emails and posts | You re-paste the rules every time |
| Editing by hand | None | Free | Anything carrying your name | Time |
| A dedicated humanizer app | Sign-up | Free tiers and paid plans that vary | Volume, languages, shared plans | Closed prompts; privacy terms vary by vendor |
My take: the skill sits in a sweet spot between a throwaway prompt and a paid app. It is as transparent as a prompt and as repeatable as an app, and for most Claude users that is enough.
How we checked this
- What I read on 29 October 2026: the blader/humanizer repository page, README, SKILL.md and CHANGELOG on GitHub; Claude Help Center articles "What are skills?" and "Use skills in Claude"; Anthropic's Agent Skills overview; Wikipedia's Signs of AI writing; Masrour et al. 2025 on ACL Anthology.
- What this page covers: the documented contents and install steps, and my judgement on where the skill fits. Star and fork counts are as GitHub displayed them on that date and change daily.
For prompt-only alternatives, see my ChatGPT humanizer prompts and the manual method in how to humanize AI text by hand. Model choice is covered in the Claude review and the Anthropic provider profile; more guides are in the HumanRank guides hub.
Frequently asked questions
What is the Claude humanizer skill? Usually it means blader/humanizer, an open-source agent skill on GitHub that rewrites AI-sounding text against 26 patterns from Wikipedia's Signs of AI writing guide. It runs in Claude Code, Claude.ai, Claude Desktop, Codex and other skill-aware agents.
Is blader/humanizer free? Yes. It is MIT-licensed and free to install. You still need access to Claude, Codex or another agent that runs skills.
How do I install the humanizer skill in Claude Code?
Add the marketplace with /plugin marketplace add blader/humanizer, then run /plugin install humanizer@humanizer (Claude Code 2.1.142 or newer). On older versions, use npx skills add blader/humanizer --global --agent claude-code.
Can I use the humanizer skill in Claude.ai? Yes. Download the repository as a ZIP, turn on code execution, then upload it under Customize, Skills. Anthropic lists skills on Free, Pro, Max, Team and Enterprise plans.
Does the humanizer skill bypass AI detectors? No. The README is explicit that detectors still catch most of its output and that its aim is clearer text for human readers.
Who made blader/humanizer? GitHub user blader maintains it, and the README lists cfo.ai as a sponsor. The pattern list comes from Wikipedia editors' Signs of AI writing guide.
Does the skill change my facts? SKILL.md tells the model not to add any fact, name, number, date, quote or citation that was not in the source, and to ask when a detail is missing. Check the output against your original anyway.
REVIEW NOTEChecked 29 October 2026 against the blader/humanizer repository on GitHub, Anthropic's help center and developer docs on skills, and Wikipedia's Signs of AI writing guide.