Humanizer turns Wikipedia's AI-writing tells into a 40k-star agent skill

What happened

A project called blader/humanizer has reached roughly 40,000 GitHub stars, and according to a dev.to post its entire implementation is a single Markdown file. Built by Siqi Chen, the skill rewrites text to strip out the recognizable habits of LLM writing.

The origin is a Wikipedia problem. Editors kept finding machine-generated text dropped into articles, reading in a distinctly recognizable way, so WikiProject AI Cleanup wrote a page cataloguing the tells. Chen's move was simple in retrospect: hand that catalogue to Claude Code and ask it to build a skill from it.

Because the result is a SKILL.md file following the Agent Skills format, there is no package, model, API or dependency tree — just instructions. Anything that supports skills can run it, and anything that does not can have the file pasted into a system prompt.

How it works

Humanizer encodes 35 patterns drawn from Wikipedia's "Signs of AI writing" page. Each is a named failure mode with a before-and-after example, which dev.to argues matters more than it sounds: the model is told it wrote "serves as a testament to" and given a specific replacement, rather than the useless instruction to write naturally.

Processing runs in two passes. It rewrites first, without treating the original paragraph structure as fixed, then checks the draft against all 35 patterns and the original claims, and rewrites whatever still fails. Pasting text in shows both stages plus a short critique of what still sounds artificial.

The patterns fall into five groups:

• Content: inflated importance, name-dropping, vague attribution such as "experts believe," and the formulaic resilience arc.
• Language and grammar: overused vocabulary (testament, landscape, showcasing, delve), "serves as" and "boasts" instead of plain "is" and "has," the not-X-but-Y construction, forced groups of three, and fake "from X to Y" ranges.
• Style: em dashes, excessive bold, lists where prose would do, title-case headings, emojis, hyphen pileups, and manufactured profundity.
• Chatbot residue: "I hope this helps," knowledge-cutoff disclaimers, "Great question!"
• Filler: "in order to" instead of "to," stacked hedges, and endings that gesture vaguely at a bright future.

Two design decisions stand out. Since v2.9.0 there is a hard no-fabrication rule: names, numbers, dates, quotes and citations must come from the source or the user, and a missing detail should trigger a question rather than an invention — sensible guardrails for a tool whose job is making text sound more specific. Pasting a few paragraphs of your own writing also makes it follow your rhythm and quirks instead of a default style. Pointed at a file, it touches only prose; code blocks, data, frontmatter and link targets stay put.

Running it across agents

For Claude Code, install via the skills CLI with npx skills add blader/humanizer --global, as a plugin on version 2.1.142 or newer, or on Claude Desktop by uploading the repo ZIP as a skill. dev.to notes the Desktop path was broken for a while by a symlink problem in the repo, patched in v2.11.1 and fixed properly in v2.11.2 — worth retrying if it failed for you months ago.

The same CLI targets more than 40 agents, including Codex, Cursor, Gemini CLI, GitHub Copilot, OpenCode, Amp, Cline and Windsurf, each with its own global skills folder. For unlisted tools, copy SKILL.md into the agent's skill folder manually. For a plain chat interface with no skill support at all, paste the file's contents as a system prompt; you lose the slash command and file-editing mode, but the 35 patterns still apply.

The caveat on cross-agent use, per dev.to: the skill is a prompt, so results track the underlying model's instruction-following. Frontier models apply the full pattern list and self-critique properly, while smaller local models may catch the obvious vocabulary swaps and quietly skip the harder judgment calls.

What it is not

It is not an AI-detector defeat button, and dev.to frames that misuse firmly in both directions: detectors are unreliable enough that they regularly flag human writing, so passing one proves little, and the tool's job is removing bad writing habits rather than adversarially targeting a classifier. It is the wrong instrument for dodging academic integrity checks. Beyond that, some patterns are opinions rather than rules, it cannot add substance to thin text, and overcorrection is a real risk.

Why it matters

The mechanism is as interesting as the output. Volunteers on Wikipedia documented how AI writing fails, one person pointed a coding agent at that documentation, and the resulting prompt file spread to roughly 40,000 stars largely because it has essentially no distribution problem. It is also a clean demonstration of the Agent Skills format's portability: one artifact that installs into dozens of tools and degrades gracefully to a pasted system prompt elsewhere. For anyone publishing AI-assisted text, the 35 patterns double as a standalone editing checklist, and the MIT-licensed source is readable enough that dev.to suggests you learn something from it even if you never install it.