Humanizer is an agent skill that rewrites AI-sounding text so it reads like a person wrote it, without changing what the text actually says. It ships as Markdown rather than as a program, so it plugs into whatever assistant you already run — the repository's own topics name Claude Code, Codex and Cursor. It is meant for anyone who drafts with a model and then publishes under their own name: developers writing documentation, changelogs and launch posts, and writers who want the speed of a model without its fingerprints on the page.
What it does
The problem it targets is familiar. AI prose has tells — em dashes everywhere, forced triads, hollow words like "testament" — and readers pick up on them even when they cannot name them. Humanizer takes that text and gives it back rewritten, with the claims intact.
A few specifics matter more than the pitch:
- It works from 35 patterns taken from Wikipedia's "Signs of AI writing", the page maintained by WikiProject AI Cleanup. The list is public and documented, not a private heuristic you have to trust blindly.
- It does not invent anything. A name, number, date, quote, citation or other factual detail has to come from your source text or from you.
- Pointed at a file, it touches only the prose. Code blocks, data, frontmatter and link targets are left alone.
- It shows its work. You see the first rewrite plus a short critique of whatever still sounds artificial, then the final version — so you can judge the edit rather than accept it.
How it works
The skill runs two passes. The first rewrite does not treat the original structure as fixed, which means sentences can be merged, split or reordered rather than merely reworded. The draft that comes out is then checked twice: once against the 35 patterns, and once against the claims in the original. Anything that still fails gets rewritten again.
Style is handled by context. Personal writing keeps the writer's voice; technical and reference prose is pushed toward neutral and plain. If you hand it a sample of your own writing, that sample overrides the built-in style rules and the rewrite follows your habits instead of the defaults.
Because the whole thing is Markdown instructions rather than a binary, there is nothing to install in the usual sense and nothing that locks you to one vendor. Any agent with skill support can load it. GitHub reports Python as the repository's main language, but the skill itself is prose telling a model what to look for and what to do about it.
Getting started
The simplest path is to call the skill and paste:
/humanizer
[paste your text here]
Plain language works as well — "Please humanize this text: …" gets the same behaviour. To rewrite a file, give it the path instead: "Humanize the prose in docs/launch-post.md", and it edits the prose in place while leaving everything else untouched.
For voice matching, include two or three paragraphs of your own writing in the same message before the text you want rewritten, labelled as a sample. The skill will follow that sample rather than its default style rules. The repository is MIT licensed and carries an install badge for skills.sh, so it can be dropped into a team's shared skill set without a licensing conversation.
When to use it / when not
It fits when a model wrote the first draft and a human has to own the result: release notes, README prose, a blog post, an internal memo, marketing copy that needs to sound like a person. It also fits editing at the file level, where you want the prose cleaned up and the code samples left exactly as they are.
It is the wrong tool in a few cases. It is not a fact-checker — it preserves your claims, including the wrong ones, and it will not fill a gap with a plausible statistic. It will not rescue a draft that has nothing to say; removing the tells from empty writing leaves shorter empty writing. And the documented flow is interactive, built around pasting text and reading the critique, so it is not the obvious choice for an automated pipeline that needs silent, deterministic output.
Humanizer is worth taking seriously if you publish text that a model helped write, which by now is most people who publish anything. Its appeal is the modesty of the design: no model to host, no service to sign up for, no scoring dashboard — a Markdown file, a public checklist of 35 patterns anyone can read, and a two-pass rewrite that argues with itself in front of you. At roughly 41,000 stars and 3,500 forks under an MIT licence, it has clearly struck a nerve, and the low cost of trying it is the point. Teams standardising on agent skills should look at it early, both for the tool itself and as a compact example of how much a well-written skill file can do without a line of application code.