What is Humanizer?
Humanizer is a skill for Claude Code, Codex and other agents that support the format. It reads the text you point it at and rewrites the patterns that mark writing as machine-made: the "not just X, it's Y" construction, one-line closers, forced triads, overused stock words, decorative bold and em dashes used as a universal connector.
It is built directly on Wikipedia's "Signs of AI writing", the page WikiProject AI Cleanup editors use to spot generated text. Siqi Chen asked Claude Code to read that page and turn it into a skill; the two projects are not formally connected but lean on the same list. The current version numbers 26 patterns, edits the five strongest on sight, and can match a voice sample you paste in. It is explicit that it writes for human readers and does not try to pass AI detectors.
Is Humanizer worth it?
Most "humanise this" tools paraphrase at random and hope. Humanizer works from a named list instead: each of its 26 patterns is a specific failure mode with a before and after, so the model is told exactly what to cut rather than the useless instruction to "write naturally". That design is the reason it reads better than a generic rewrite, and the skill is candid that the weaker tells, such as dashes or curly quotes, are only acted on when several appear together, because a careful writer may use any one on purpose.
The evidence is thin but pointed. The repository's own blind test had judges prefer its rewrite to the original AI text 16 times out of 16, which is the author's test, not an independent one. It leans on real editing rules rather than a model's taste, which is why it travels well across agents.
The honesty is the selling point. Humanizer states plainly that it edits for human readers and that detectors still flag most of its output, so anyone hoping to sneak past a plagiarism or AI checker is in the wrong place. It also only touches prose, leaving code, data, front matter and link targets alone, which makes it safe to run on a Markdown file in a repository.
The catch: it is a prompt skill, so results vary with the model behind it and with how heavily the text was generated, and a careless run can flatten a distinctive voice into plain newspaper prose unless you give it a sample to match. With 69,115 stars and commits into November 2026, it is popular and maintained, but the "16 of 16" figure is the only number it offers.
Our take: hand it a blog draft that reeks of the machine and it will read like a person wrote it; hand it to a student trying to fool Turnitin and it will not help, which is rather the point.
Where it earns its place
A developer advocate drafting launch posts with an agent
Make a generated announcement read like the team wrote it
Run `/humanizer` on the draft, or paste two or three paragraphs of your own writing first so it matches your rhythm and punctuation. It shows its first rewrite, a critique of what still sounds artificial, and the final version.
A technical writer cleaning up model-written documentation
Strip stock phrases and decorative formatting from a docs file without breaking it
Give it the file path; it edits only the prose and reports which sections it reviewed and any it could not cover, leaving code blocks, front matter and link targets untouched.
A small marketing team editing AI-assisted copy
Remove the tells that make copy feel generated before it ships
Install the skill once across your agents and run it on each draft. It costs only the model tokens of the run, and the named patterns make its edits easy to check against the original.
Before you commit
- Setup
- Claude Code: `/plugin marketplace add blader/humanizer` then `/plugin install humanizer@humanizer` (2.1.142+), or `npx skills add blader/humanizer --global --agent claude-code`; Codex and other agents use the same `npx skills add` command.
- Integrations
- Claude Code, Codex, pi, Claude.ai and Claude Desktop, and any agent the Skills CLI supports, including Gemini CLI, GitHub Copilot and Windsurf; ported into Nous Research's Hermes Agent.
- Data and privacy
- The skill is a set of instructions; it sends nothing itself. Your text goes to whichever model your agent uses, under that provider's terms.
- Limits
- It does not defeat AI detectors; output quality depends on the model running it; it edits prose only; the single published result is the project's own blind test.
What’s good
- Edits from a named 26-pattern list, not random paraphrase
- Honest that it is for readers, not beating detectors
- Touches only prose; leaves code, data and links alone
- Can match a pasted sample of your own voice
What’s not
- Only evidence is the author's own 16-of-16 blind test
- Results vary with the model running it
- Can flatten a distinctive voice without a sample
- Does nothing against AI detectors
Best for
Writers and developers who want obvious AI tells removed from drafts and documentation while keeping the meaning intact.
Humanizer alternatives
Humanizer: common questions
- Is Humanizer worth it?
- Humanizer is a sharp, honest editing skill that strips the obvious AI tells from your prose, as long as you remember it is editing for readers, not for detectors. We rate it 4.0 out of 5.
- How much does Humanizer cost?
- Free and open source (MIT). You pay only the model tokens your agent spends running it.
- What are the downsides of Humanizer?
- Only evidence is the author's own 16-of-16 blind test; results vary with the model running it; can flatten a distinctive voice without a sample; does nothing against AI detectors.
- Who is Humanizer best for?
- Writers and developers who want obvious AI tells removed from drafts and documentation while keeping the meaning intact.
- Is Humanizer safe for company data?
- The skill is a set of instructions; it sends nothing itself. Your text goes to whichever model your agent uses, under that provider's terms.
- What are the best Humanizer alternatives?
- The closest alternatives we have reviewed are Caveman and Agent Skills (Addy Osmani).
Answers checked Nov 14, 2026.
Make Humanizer? Show this rating on your site
