A new open source plugin for Anthropic’s Claude Code has turned a Wikipedia cleanup guide into a practical set of instructions for making AI-generated text sound less like AI-generated text.
The plugin, called Humanizer, was released on Saturday by tech entrepreneur Siqi Chen. It uses a list of 24 language and formatting patterns that Wikipedia editors have identified as common signs of chatbot writing.
How Humanizer Works
Humanizer is built for Claude Code, Anthropic’s terminal-based coding assistant. The tool is described as a “skill file,” which means it is a Markdown-formatted file that adds written instructions to the prompt used by the large language model behind the assistant.
Those instructions tell Claude to avoid patterns that Wikipedia editors have associated with AI-written prose. Chen published the plugin on GitHub, where it had picked up over 1,600 stars as of Monday.
The idea is simple: if editors can describe the language habits that make AI output easy to recognize, a model can also be told to stop using those habits. Chen pointed to that reversal directly on X.
“It’s really handy that Wikipedia went and collated a detailed list of ‘signs of AI writing,’” Chen wrote on X. “So much so that you can just tell your LLM to… not do that.”
Unlike a normal system prompt, a Claude skill file is formatted in a standardized way that Claude models are fine-tuned to interpret with more precision than a plain system prompt. Custom skills require a paid Claude subscription with code execution turned on.
Where The Checklist Came From
The source material behind Humanizer is a guide from WikiProject AI Cleanup. The project is a group of Wikipedia editors who have been looking for AI-generated articles since late 2023.
French Wikipedia editor Ilyas Lebleu founded the project. Its volunteers have tagged over 500 articles for review and, in August 2025, published a formal list of the patterns they were repeatedly seeing.
That list was meant to help editors spot likely AI-assisted writing. Humanizer uses it for the opposite purpose: it turns those observations into writing instructions for Claude.
This does not mean the plugin makes AI output accurate, original, or trustworthy. In limited testing described in the source article, Chen’s skill file made the AI agent’s writing sound less precise and more casual. It also had possible drawbacks: it would not improve factuality and might hurt coding ability.
What Counts As An AI Tell
The Wikipedia guide includes many examples, but one of the clearest categories is inflated language. Some chatbot prose uses grand phrases where plain description would work better.
Examples cited in the source include phrases such as “marking a pivotal moment” and “stands as a testament to.” The guide also points to promotional-sounding descriptions, including views called “breathtaking” and towns described as “nestled within” scenic regions.
Another pattern is the use of “-ing” phrases at the end of sentences to create an analytical tone. The source gives the example “symbolizing the region’s commitment to innovation.”
Humanizer’s answer is to steer Claude toward plainer statements. The source gives this before-and-after example:
Before: “The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain.”
After: “The Statistical Institute of Catalonia was established in 1989 to collect and publish regional statistics.”
The edited version removes the ceremonial framing and keeps the factual point. That is useful in many kinds of writing, but it also shows how fragile phrase-based AI detection can be. If the signal is a writing habit, the model can be asked to avoid the habit.
Why Detection Remains Hard
The Humanizer plugin highlights a larger problem for AI writing detection. The source article notes that AI writing detectors do not work reliably because there is nothing inherently unique about human writing that consistently separates it from LLM writing.
Many language models tend toward certain styles, but they can also be prompted away from them. The Humanizer skill is an example of that. The source also notes that OpenAI had a yearslong struggle against the em dash, showing that removing a visible style marker can be difficult even when the target is known.
The issue also runs in the other direction: people can write in ways that look chatbot-like. The source article notes that professional writing may contain traits that trigger AI detectors, because large language models learned from examples of professional writing scraped from the web.
The Wikipedia guide itself includes an important caveat. Its list is based on observations, not absolute rules. A 2025 preprint cited on the page found that heavy users of large language models correctly spot AI-generated articles about 90 percent of the time. That still leaves 10 percent as false positives, which can mean wrongly dismissing quality writing while trying to filter out AI slop.
That makes the deeper lesson clear. Pattern checklists can help identify suspicious writing, especially unedited chatbot output, but they are not proof. Humanizer shows why: once a pattern is public, it can become both a warning sign and a recipe for avoiding the warning sign.
For editors, publishers, and anyone reviewing text, the more durable question may not be whether a sentence contains a familiar AI flourish. It may be whether the work is factually sound, useful, and accountable for what it claims.