Humanizer 调研：把维基百科「AI 写作特征」清单变成一个能装进任何 Agent 的 Skill

by Mycelium Protocol

---

项目地址：https://github.com/blader/humanizer 方法依据：维基百科 WikiProject AI Cleanup《Signs of AI writing》 https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing 授权：MIT

---

一句话结论

Humanizer 不是一个 App，是一份纯 Markdown 写的 Skill 文件（SKILL.md），装进任何支持 Skill 的 Agent 就能用——Claude Code、Claude Desktop 等。它做的事很单一：把读起来像 AI 生成的文字，改写得不像，但保留原作者的事实、含义和语气。 MIT 协议，36120 star，是今天写的几个项目里星标最高的一个。

不是拍脑袋定义”AI 味”，是抄维基百科的作业

这个项目最聪明的地方，不是它自己发明了一套”怎样才算 AI 味”的标准，而是直接用维基百科 WikiProject AI Cleanup 维护的《Signs of AI writing》清单——这是一群志愿者专门盯着维基百科词条里的 AI 生成痕迹、持续更新的一份权威参考。项目引用了这份文档的核心论点：

“大语言模型用统计算法猜下一个词该是什么。结果趋向于适用最广泛情况的、统计上最可能的那个答案。”

站在一份被持续维护、有社区共识的清单上做事，比自己拍一套规则要靠谱得多——这份清单会随着 AI 写作特征的演变而更新，Humanizer 也就跟着更新，不需要自己重新总结。

35 种模式，几个最典型的

清单分三类：内容层面（虚假重要性、堆砌形容词式的浅层分析、销售话术）、语言语法层面（“不是X而是Y”句式、被动语态、名称反复变换）、风格层面（破折号泛滥、加粗小标题、表情符号、标题用 Title Case）。挑几个最一眼能认出来的：

模式	改前	改后
夸大重要性	”marking a pivotal moment in the evolution of…"	"was established in 1989 as part of a wider decentralization”
浅层 -ing 分析	”symbolizing… reflecting… showcasing…”	只保留原文能支撑的内容
强行三件套	”innovation, inspiration, and insights”	该几个词就几个词，不硬凑三个
破折号泛滥	”institutions—not the people—yet this continues—“	改用句号、逗号、冒号或括号
加粗小标题列表	”Performance: Performance improved”	列表不产生信息量时直接改成散文
制造假的深刻	”At its core, what matters is…”	直接说重点，不绕这层壳
强行制造金句碎片	”It had no preference. No prior. No nostalgia.”	用自然的句长和具体的陈述

看这几个例子会发现一个共同点：AI 写作特征往往不是”写错了”，是”写得太用力”——每句话都想显得重要、每个列举都想凑够三个、每个转折都想加个破折号强调。Humanizer 干的事是把这层”用力过猛”卸掉。

流程：先改，再查，再定稿

三步走：先起草一版改写；检查这版改写里还有没有残留的 AI 写作模式，同时检查有没有在改写过程中悄悄篡改了原意；确认没问题再定稿。

明确不做的事：不发明事实、姓名、日期、引语、引用。任何改写里新增的细节，必须来自原文或者作者本人提供的信息——这条边界写得很硬，避免”为了让文字更自然”反而编造内容。

语气匹配：给个样本，照着你的调调改

如果你想让改写结果贴近自己的写作习惯，可以先粘贴 2-3 段自己的文字当样本，Humanizer 会匹配这段样本的节奏、用词、标点习惯和”故意的怪癖”，再去改写目标文本——不是套用一个通用的”人类文风模板”，是学你自己的。

三种安装方式

# Skills CLI，全局安装
npx skills add blader/humanizer --global

# Claude Code 插件市场
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

# 手动：把 SKILL.md 扔进任意 Agent 的技能目录
git clone https://github.com/blader/humanizer.git /path/to/your/skills/humanizer

谁该看这个

适合：日常大量用 AI 辅助写作、但不希望文字读起来一股统一的”AI 腔”的人；需要把内部文档、公开博客、社交文案过一道”去味”处理的场景。

不适合 / 需要注意：它改的是文风，不是事实核查工具——原文如果本身有错误，Humanizer 不会帮你纠正，它明确的边界就是”不改变事实，只改怎么说”。

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

TL;DR

Humanizer isn’t an app — it’s a skill written in plain Markdown (SKILL.md) that installs into any skill-compatible agent: Claude Code, Claude Desktop, and others. What it does is narrow and specific: rewrite text that sounds AI-generated so it doesn’t, while keeping the original writer’s facts, meaning, and voice. MIT licensed, 36,120 stars — the highest star count of anything covered today.

Not inventing “AI-ness” from scratch — borrowing Wikipedia’s homework

The cleverest part of this project isn’t that it invented its own definition of “sounds like AI.” It’s that it directly uses the “Signs of AI writing” catalog maintained by Wikipedia’s WikiProject AI Cleanup — a group of volunteers who continuously track AI-generation traces in Wikipedia articles and keep an authoritative, living reference updated. The project quotes the document’s core argument:

“LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases.”

Building on a continuously maintained list with community consensus is far more reliable than inventing your own rulebook — the list evolves as AI writing tells evolve, and Humanizer evolves with it without having to re-derive its own taxonomy.

35 patterns, a few of the most recognizable

The catalog splits into three categories: content-level (inflated importance, shallow -ing analysis stacked as adjectives, sales language), language/grammar-level (“not X but Y” constructions, passive voice, shifting names), and style-level (dash overuse, bold mini-headings, emoji, Title Case headings). A few of the most instantly recognizable:

Pattern	Before	After
Inflated importance	”marking a pivotal moment in the evolution of…"	"was established in 1989 as part of a wider decentralization”
Shallow -ing analysis	”symbolizing… reflecting… showcasing…”	Keep only what the source supports
Forced groups of three	”innovation, inspiration, and insights”	Use the number of items the meaning actually needs
Dash overuse	”institutions—not the people—yet this continues—“	Switch to periods, commas, colons, or parentheses
Bold mini-heading lists	”Performance: Performance improved”	Use prose when a list adds no value
Manufacturing false depth	”At its core, what matters is…”	State the point directly, skip the framing
Forced punchline fragments	”It had no preference. No prior. No nostalgia.”	Use natural sentence lengths and specific claims

The common thread across these examples: AI writing tells usually aren’t “wrong” — they’re “trying too hard.” Every sentence reaches to sound important, every list strains to hit exactly three items, every transition wants a dash for emphasis. What Humanizer does is take that strain out.

The process: rewrite, check, finalize

Three steps: draft a rewrite; check that draft for remaining AI writing patterns and, separately, for any claims that quietly changed meaning during the rewrite; finalize only once both checks pass.

Explicitly out of scope: it does not invent facts, names, dates, quotes, or citations. Any detail added during rewriting must come from the source text or from the writer directly — a hard boundary meant to prevent “making the prose sound more natural” from sliding into fabrication.

Voice matching: give it a sample, it follows your cadence

If you want the rewrite to match your own writing habits, paste 2-3 paragraphs of your own writing as a sample first. Humanizer matches that sample’s rhythm, word choice, punctuation habits, and deliberate quirks before rewriting the target text — not a generic “human writing template,” but learning from you specifically.

Three ways to install

# Skills CLI, global install
npx skills add blader/humanizer --global

# Claude Code plugin marketplace
/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

# Manual: drop SKILL.md into any agent's skills folder
git clone https://github.com/blader/humanizer.git /path/to/your/skills/humanizer

Who should look at this

Good fit: anyone who writes a lot with AI assistance day-to-day but doesn’t want everything to read with the same flattened “AI voice”; scenarios needing a pass over internal docs, public blog posts, or social copy to strip that tell.

Not a fit / worth noting: it changes prose style, not a fact-checker — if the source text has errors, Humanizer won’t correct them. Its explicit boundary is “change how it’s said, not what’s true.”

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.