Humanizer: Detects 25 types of AI-generated writing and rewrites AI-sounding text into "human language".Humanizer:检测 25 种 AI 写作痕迹、把 AI 腔改写成「人话」的技能
Text written by AI can be recognised at a glance – "It is worth noting", "In summary", three consecutive parallel structures, excessive antithesis. Humanizer (71,000 stars) is an Agent skill that focuses on text: it detects 25 known AI-generated writing patterns and then rewrites the text without changing its original meaning, making it read like it was written by a human.
AI 写的东西一眼就能认出来——「值得注意的是」「总而言之」、排比三连、过度对仗。Humanizer(7.1 万星)是一个专注于文字的 Agent 技能:检测 25 种已知的 AI 生成写作模式,然后在不改变原意的前提下重写,让文字读起来像人写的。

How are 25 types of AI traces defined?25 种 AI 痕迹是怎么定义的
The author breaks down the flaws in AI writing into 25 identifiable patterns: empty summary sentences, forced uplifting endings, overuse of tricolons, excessive use of dashes, false expressions of empathy, piling up of nominalisations, and so on – each with clear criteria for identification, rather than simply saying "it feels a bit like AI".
The core constraint of the rewriting stage is "do not change the meaning": it is not a polisher, and does not add opinions to the text, but only removes traces, so the rewritten content remains faithful to the facts and structure of the original.
The current version, v4.0.0, has undergone several iterations, with 71,000 stars and a daily increase of hundreds of stars. It is one of the most popular projects in the field of text-based Agent skills.
作者把 AI 写作的破绽拆成了 25 种可判定的模式:空洞的总结句、强行升华的结尾、三段式排比滥用、过度使用破折号、虚假的共情表达、名词化堆砌等等——每一种都有明确的判别标准,而不是凭感觉说「有点 AI 味」。
重写阶段的核心约束是「不改变意思」:它不是润色器,不会往文里加观点,只做痕迹清除,所以改完的内容仍然忠实于原稿的事实与结构。
当前版本 v4.0.0 已迭代多轮,7.1 万星、日均涨星数百,是文字类 Agent 技能里最热门的项目之一。
Installation & Use安装与使用
Click the button above to download humanizer-4.0.zip (only 0.03MB), extract it to get the skills folder, and place it in ~/.claude/skills/ (Claude Code) or the skills directory of the corresponding Agent. Restart the session to load it.
When using it, simply paste the text to the Agent and say "Rewrite this passage using Humanizer": It will first list the detected AI patterns, then output the rewritten version with a before-and-after comparison, allowing you to confirm which traces have been removed item by item.
Suitable scenarios: Before publishing an AI-generated first draft of a blog post, marketing emails, product copy, or introductory text at the beginning of a document. Less suitable: Paragraphs containing precise technical terminology in technical documents – forcibly removing the "AI feel" may actually compromise accuracy.
点上方按钮下载 humanizer-4.0.zip(仅 0.03MB),解压得到技能文件夹,放入 ~/.claude/skills/(Claude Code)或对应 Agent 的技能目录,重启会话即可加载。
使用时直接把文本粘给 Agent 并说「用 humanizer 改写这段」:它会先列出检测到的 AI 模式清单,再输出改写稿和前后对照,你可以逐条确认哪些痕迹被移除。
适合的场景:AI 初稿发博客前、营销邮件、产品文案、文档开头的导语;不太适合的:技术文档中的精确术语段落——强行「去 AI 味」反而可能损伤准确性。
Position in the spectrum of AI-removal tools在去 AI 味工具谱系里的位置
We have previously added two similar skills to our repository: sepia-deai-skill, which focuses on reducing traces in academic texts, and unslop-skill, which focuses on avoiding clichés in design outputs. Humanizer stands out due to its clear list of modes, which can be audited item by item, and its extremely small size and zero dependencies, making it almost burden-free once installed.
Suggested combined approach: First, have the model write a draft → use Humanizer to remove traces → then have a human review the facts and tone. After these three steps, it's generally hard to detect any machine-like elements.
我们之前入库过两个同类技能:sepia-deai-skill 偏学术文本降痕,unslop-skill 偏设计输出的反套话。Humanizer 的差异化在于模式清单明确、逐条可审计,且体积极小、零依赖,装完几乎无负担。
建议的组合用法:先让模型写初稿 → humanizer 清痕迹 → 人再过一遍事实与语气,三层下来基本听不出机器腔了。
Note小提示
Removing the "AI feel" does not equate to falsification: Using AI to draft and then finalising the text yourself is a perfectly legitimate workflow. However, when it comes to academic submissions, media bylines, and similar contexts, please adhere to the disclosure requirements of the respective platforms. The skills pack that accompanies this article can be found via the download button above.
去 AI 味不等于造假:用 AI 起草然后自己定稿是完全正当的工作流,但涉及学术提交、媒体署名等场景时请遵守对应平台的披露要求。本文配套技能包见上方下载按钮。
⬇ Download · 点击下载:Humanizer v4.0 技能包(GitHub 仓库 zip)
MIT · 极小体量纯 Markdown 技能 · 解压后放入技能目录即可
阅读与点赞数据保存在你的浏览器本地,欢迎留下你的想法。
今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告
评论 文明发言,让讨论更有价值