Zed 的 Humanizer 技能:识别与重写 24 类 AI 写作痕迹的完整方法论

本文基于 Zed 仓库中的 humanizer 技能定义,系统讲解如何用一份 24 项模式清单识别 AI 生成文本的典型痕迹,并给出可执行的九步重写流程与标准化输出格式。读完后,你可以把这套方法直接用于任何由 LLM 起草的文案、博客、文档或 PR 描述,使文本在保留原意的前提下听起来更像真人撰写,而不是"被 AI 写过的"。

Humanizer 是什么:一个可被 Agent 直接调用的编辑技能

在 Zed 仓库的 .factory/ 目录下,团队为 Agent 会话维护了一批可复用的写作与调查技能。humanizer 就是其中之一,它的定位在技能描述中写得很明确:

Remove signs of AI-generated writing from text. Use after drafting to make copy sound more natural and human-written.

也就是说,它是一个起草后(drafting 之后)使用的润色/审查工具:输入一段已有文本,输出"去 AI 味"后的版本。技能定义文件 SKILL.md 的 frontmatter 遵循 Agent 技能的标准字段:

name: humanizer
description: Remove signs of AI-generated writing from text. Use after drafting
  to make copy sound more natural and human-written. Based on Wikipedia's
  "Signs of AI writing" guide.
allowed-tools: Read, Write, Edit, Glob, Grep, AskUserQuestion
user-invocable: true

其中 allowed-tools 声明了执行该技能时 Agent 可用的工具集(读文件、写文件、编辑、Glob/Grep 检索、向用户提问),user-invocable: true 表示用户可以主动触发它。触发方式在文档中给出(SKILL.md):

/humanizer                    # 审查当前文本中的 AI 痕迹
/humanizer "paste text here"  # 对指定文本做去 AI 化处理

这套方法的理论依据来自维基百科由 WikiProject AI Cleanup 社区维护的 "Signs of AI writing" 页面。其核心洞察只有一句话,但解释了为什么 AI 文本"一眼假":

"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."

LLM 用统计算法猜测下一个 token,结果会倾向于对最广泛场景都成立的、统计上最可能的表达。于是文本呈现出高度可预测的套话结构——这正是 24 项模式清单要捕捉的对象。

五步任务流程

技能文档把"humanize 一段文本"定义为五个步骤(SKILL.md):

1. 识别 AI 模式 —— 扫描后文列出的 24 种模式;
2. 重写问题段落 —— 用自然的表达替换 AI 腔;
3. 保留原意 —— 核心信息不能丢;
4. 注入"灵魂" —— 不只是删掉坏模式,还要让文本有个性;
5. 终审一遍 —— 问自己"哪里还能看出这是 AI 写的?",再改一轮。

值得注意的是第 4 步:文档强调"Avoiding AI patterns is only half the job. Sterile, voiceless writing is just as obvious as slop."——干净但毫无声口的写作,和烂 AI 文本一样容易辨认。下一节专门展开这部分。

人格与灵魂:去 AI 味的另一半工作

文档用一组"无灵魂写作的征兆"来定义"技术上干净、但依然假"的文本(SKILL.md):

• 每句话长度和结构都一样;
• 没有观点,只有中立陈述;
• 不承认不确定性或矛盾心理;
• 该用第一人称时用第三人称;
• 没有幽默感、没有棱角、没有个性;
• 读起来像维基百科条目或新闻通稿。

针对这些征兆,文档给出六条"加声口"的具体手法(SKILL.md):

手法	说明
有观点	不只是陈述事实,而是对事实做出反应。"I genuinely don't know how to feel about this"(我实在不知道该怎么看这件事)比中立罗列利弊更像人话
变化节奏	短促有力的句子,接上从容不迫的长句,混着用
承认复杂性	真人会有矛盾心理。"This is impressive but also kind of unsettling" 优于简单的 "This is impressive"
合适时用 "I"	第一人称不等于不专业,而是诚实。"I keep coming back to..." 表明是真人思考
允许一些混乱	完美的结构反而显得算法化。旁白、插入语、未成型的想法都是人味
感受要具体	不写 "this is concerning",而写 "there's something unsettling about agents churning away at 3am while nobody's watching"

文档给出的对照示例,展示了"干净但无灵魂"与"有脉搏"两种版本的差别:

Before(干净但无灵魂): The experiment produced interesting results. The agents generated 3 million lines of code. Some developers were impressed while others were skeptical. The implications remain unclear.

After(有脉搏): I genuinely don't know how to feel about this one. 3 million lines of code, generated while the humans presumably slept. Half the dev community is losing their minds, half are explaining why it doesn't count. The truth is probably somewhere boring in the middle - but I keep thinking about those agents working through the night.

24 类 AI 写作模式:完整清单

这是整个技能的核心资产。文档按五类组织 24 个模式:内容模式(1-6)、语言模式(7-12)、风格模式(13-18)、交流模式(19-21)、填充词与模糊限定(22-24)。每个模式都给出"要盯住的信号词"和前后对照示例。以下按原文档结构完整保留。

一、内容模式(Content Patterns)

1. 意义膨胀(Significance Inflation)

要盯住: stands/serves as、is a testament/reminder、a vital/significant/crucial/pivotal/key role/moment、underscores/highlights importance、reflects broader、symbolizing ongoing/enduring/lasting、marking/shaping the、represents a shift、key turning point、evolving landscape。

Before: The Statistical Institute was officially established in 1989, marking a pivotal moment in the evolution of regional statistics.

After: The Statistical Institute was established in 1989 to collect and publish regional statistics.

把"标志着地区统计演进的关键时刻"压缩回朴素事实:成立、为了什么。

2. 名望蹭名(Notability Name-Dropping)

要盯住: cited in NYT/BBC/FT、independent coverage、active social media presence、written by a leading expert。

Before: Her views have been cited in The New York Times, BBC, Financial Times, and The Hindu.

After: In a 2024 New York Times interview, she argued that AI regulation should focus on outcomes rather than methods.

修法:把"被多家媒体引用"的名声罗列,换成具体的引用来源加她实际说了什么。

3. 浅层 -ing 分析(Superficial -ing Analyses)

要盯住: highlighting/underscoring/emphasizing...、ensuring...、reflecting/symbolizing...、contributing to...、cultivating/fostering...、showcasing...。

Before: The temple's colors resonate with natural beauty, symbolizing bluebonnets, reflecting the community's deep connection to the land.

After: The temple uses blue and gold colors. The architect said these were chosen to reference local bluebonnets.

一连串悬垂分词构成的"象征/反映"链条,换成可核实的事实加直接引语。

4. 宣传腔(Promotional Language)

要盯住: boasts a、vibrant、rich(比喻义)、profound、showcasing、exemplifies、commitment to、natural beauty、nestled、in the heart of、groundbreaking、renowned、breathtaking、must-visit、stunning。

Before: Nestled within the breathtaking region, Alamata stands as a vibrant town with rich cultural heritage and stunning natural beauty.

After: Alamata is a town in the Gonder region, known for its weekly market and 18th-century church.

5. 模糊归因(Vague Attributions)

要盯住: Industry reports、Observers have cited、Experts argue、Some critics argue、several sources/publications。

Before: Experts believe it plays a crucial role in the regional ecosystem.

After: The river supports several endemic fish species, according to a 2019 survey by the Chinese Academy of Sciences.

"专家认为"必须落到具体的报告、年份和机构。

6. 公式化"挑战"章节(Formulaic "Challenges" Sections)

要盯住: Despite its... faces several challenges...、Despite these challenges、Challenges and Legacy、Future Outlook。

Before: Despite challenges typical of urban areas, the city continues to thrive as an integral part of growth.

After: Traffic congestion increased after 2015 when three new IT parks opened. The municipal corporation began a drainage project in 2022.

模板化的"尽管有挑战,依然蓬勃发展",换成有时间点和具体事件的事实。

二、语言模式(Language Patterns)

7. AI 高频词汇(AI Vocabulary Words)

高频词表: Additionally、align with、crucial、delve、emphasizing、enduring、enhance、fostering、garner、highlight(动词)、interplay、intricate/intricacies、key(形容词)、landscape(抽象义)、pivotal、showcase、tapestry(抽象义)、testament、underscore(动词)、valuable、vibrant。

Before: Additionally, a distinctive feature showcases how these dishes have integrated into the traditional culinary landscape.

After: Pasta dishes, introduced during Italian colonization, remain common, especially in the south.

这一条是纯词汇层面的"信号词表",做批量审查时可以直接作为 Grep 清单使用——这也是 frontmatter 中 allowed-tools 包含 Grep 的实用场景。

8. 回避系动词(Copula Avoidance)

要盯住: serves as/stands as/marks/represents [a]、boasts/features/offers [a]。

Before: Gallery 825 serves as the exhibition space. The gallery features four spaces and boasts over 3,000 square feet.

After: Gallery 825 is the exhibition space. The gallery has four rooms totaling 3,000 square feet.

AI 文本很少直接说"is/has",偏爱"serves as""boasts"这类替代系动词;修法是老老实实用回 is/are/has。

9. 否定式排比(Negative Parallelisms)

要盯住: "Not only...but..."、"It's not just about..., it's..."。

Before: It's not just about the beat; it's part of the aggression. It's not merely a song, it's a statement.

After: The heavy beat adds to the aggressive tone.

10. 三段论滥用(Rule of Three Overuse)

Before: The event features keynote sessions, panel discussions, and networking opportunities. Attendees can expect innovation, inspiration, and industry insights.

After: The event includes talks and panels. There's also time for informal networking.

"X、Y、Z"三项并列是 LLM 统计上最顺手的句式之一,连续两组三段并列基本可以断定。

11. 同义词轮换(Synonym Cycling)

Before: The protagonist faces challenges. The main character must overcome obstacles. The central figure eventually triumphs. The hero returns home.

After: The protagonist faces many challenges but eventually triumphs and returns home.

同一对象在相邻句子里被轮流换成 protagonist/main character/central figure/hero,是典型的"怕重复"算法痕迹。

12. 虚假区间(False Ranges)

要盯住: "from X to Y" 但 X 和 Y 根本不在同一个可比标度上。

Before: Our journey has taken us from the singularity of the Big Bang to the cosmic web, from the birth of stars to the dance of dark matter.

After: The book covers the Big Bang, star formation, and current theories about dark matter.

三、风格模式(Style Patterns)

13. 破折号滥用(Em Dash Overuse)

Before: The term is promoted by institutions—not the people themselves—yet this continues—even in documents.

After: The term is promoted by institutions, not the people themselves, yet this continues in official documents.

14. 粗体滥用(Boldface Overuse)

Before: It blends OKRs, KPIs, and tools such as the Business Model Canvas and Balanced Scorecard.

After: It blends OKRs, KPIs, and visual strategy tools like the Business Model Canvas and Balanced Scorecard.

术语首次出现时加粗可以,通篇每个专有名词都加粗则是 AI 排版习惯。

15. 行内标题列表(Inline-Header Lists)

Before:

• Performance: Performance has been enhanced through optimized algorithms.
• Security: Security has been strengthened with encryption.

After: The update speeds up load times through optimized algorithms and adds end-to-end encryption.

"- X: X has been ..."这种"加粗标签 + 重复标签解释"的列表,能合并回一句话的应尽量合并。

16. 标题大小写(Title Case Headings)

Before: ## Strategic Negotiations And Global Partnerships

After: ## Strategic negotiations and global partnerships

17. 专业写作中的 Emoji

Before: 🚀 Launch Phase: The product launches in Q3 / 💡 Key Insight: Users prefer simplicity

After: The product launches in Q3. User research showed a preference for simplicity.

18. 弯引号(Curly Quotation Marks)

Before: He said “the project is on track” but others disagreed.

After: He said "the project is on track" but others disagreed.

原文 Before 使用弯引号(U+201C/U+201D)、After 使用直引号;在纯文本/Markdown 语境中统一用直引号,避免排版软件自动替换造成的风格不一致。

四、交流模式(Communication Patterns)

19. 聊天机器人残留(Chatbot Artifacts)

要盯住: I hope this helps、Of course!、Certainly!、You're absolutely right!、Would you like...、let me know、here is a...。

Before: Here is an overview of the French Revolution. I hope this helps! Let me know if you'd like me to expand on any section.

After: The French Revolution began in 1789 when financial crisis and food shortages led to widespread unrest.

20. 知识截止日免责声明(Knowledge-Cutoff Disclaimers)

要盯住: as of [date]、Up to my last training update、While specific details are limited/scarce...、based on available information...。

Before: While specific details about the company's founding are not extensively documented in readily available sources, it appears to have been established sometime in the 1990s.

After: The company was founded in 1994, according to its registration documents.

21. 谄媚语气(Sycophantic Tone)

Before: Great question! You're absolutely right that this is complex topic. That's an excellent point!

After: The economic factors you mentioned are relevant here.

五、填充词与模糊限定(Filler and Hedging)

22. 填充短语(Filler Phrases)

Before	After
"In order to achieve this"	"To achieve this"
"Due to the fact that"	"Because"
"At this point in time"	"Now"
"It is important to note that"	(直接删掉)
"has the ability to"	"can"

23. 过度模糊限定(Excessive Hedging)

Before: It could potentially possibly be argued that the policy might have some effect on outcomes.

After: The policy may affect outcomes.

一层限定("may")足够了,could + potentially + possibly + might + some 五重叠加是 LLM 的免责式写法。

24. 泛泛的正面结尾(Generic Positive Conclusions)

Before: The future looks bright for the company. Exciting times lie ahead as they continue their journey toward excellence.

After: The company plans to open two more locations next year.

结尾必须是可核实的具体计划或事实,而不是"未来一片光明"式的口号。

九步重写流程

模式清单解决"识别",流程解决"操作"。文档的 Process 一节(SKILL.md)定义了完整循环:

1. 仔细读输入文本;
2. 找出 24 种模式的所有实例;
3. 重写问题段落;
4. 确保改后文本满足四条标准:
   • 读出声来自然;
   • 句式有变化;
   • 用具体细节代替模糊断言;
   • 在合适处使用简单结构(is/are/has);
5. 写出 humanized 初稿(Draft);
6. 自问:"哪里还明显是 AI 写的?";
7. 列出剩余的 AI 痕迹(如果有);
8. 根据审计结果再修一轮;
9. 呈现最终版本。

这套流程的关键设计是第 5-8 步的"自审计"回路:初稿往往只是把模式词替换掉了,句式节奏、结构对称性这些二阶痕迹还要在审计轮才暴露出来。

标准化输出格式

技能的产出不是"直接给你改好的文本",而是四段式报告(SKILL.md):

## Draft Rewrite
[第一遍 humanized 文本]

## AI Audit
What still reads as AI-generated:
- [剩余 AI 痕迹的要点列表]

## Final Version
[审计后修订的文本]

## Changes Made
- [所做转换的简要总结]

这个格式把"改了什么"和"为什么改"都显式化了:审计部分是给读者/审核者看的质检记录,Changes Made 部分是可复查的改动清单,适合作为 PR 或文档评审的附注。

完整示例:一段 AI 味十足的文本如何被处理

文档结尾给了一条端到端示例(SKILL.md),值得完整看一遍:

输入(AI 味极重):

Great question! Here is an essay on this topic. I hope this helps!

AI-assisted coding serves as an enduring testament to the transformative potential of large language models, marking a pivotal moment in the evolution of software development. In today's rapidly evolving technological landscape, these groundbreaking tools—nestled at the intersection of research and practice—are reshaping how engineers ideate, iterate, and deliver, underscoring their vital role in modern workflows.

第一遍重写(Draft Rewrite):

AI coding assistants speed up some tasks. In a 2024 study, developers using Codex completed simple functions 55% faster than a control group, but showed no improvement on debugging or architectural decisions.

The tools are good at boilerplate: config files, test scaffolding, repetitive refactors. They are bad at knowing when they are wrong.

自审计(AI Audit)——"仍然像 AI 的地方":

• 节奏仍然太工整(对照太干净、句长均匀);
• 结尾略带口号味。

最终版(Final Version):

AI coding assistants can make you faster at the boring parts. Not everything. Definitely not architecture.

They're great at boilerplate: config files, test scaffolding, repetitive refactors. They're also great at sounding right while being wrong. I've accepted suggestions that compiled, passed lint, and still missed the point because I stopped paying attention.

If you don't have tests, you're basically guessing.

改动清单(Changes Made):

• 删除聊天机器人残留("Great question!"、"I hope this helps!");
• 删除意义膨胀词("testament"、"pivotal moment"、"evolving landscape");
• 删除宣传腔("groundbreaking"、"nestled");
• 删除破折号;
• 用直陈句替代系动词回避("serves as");
• 加入第一人称视角和观点;
• 打破均匀的节奏。

对照初稿和终稿可以看到:审计轮做的不是替换词汇,而是把"工整的对照句"打碎,加入第一人称的真实经历("I've accepted suggestions that compiled, passed lint, and still missed the point")——这正是"加灵魂"而非"减模式"的差别。

这个技能在 Zed 仓库中的位置

从仓库结构看,.factory/ 目录是 Zed 团队为 Agent 工作流维护的提示词与技能库,目前包含两块内容:

• .factory/prompts/crash/ —— 崩溃调查系列提示词(调查、修复、关联 issue 三个文件),在 AGENTS.md 中被明确引用:当 Agent 打开或更新 PR、调查 Sentry 崩溃报告时使用这些提示词;
• .factory/skills/ —— 用户可调用的写作技能,目前有两个:brand-writer 和 humanizer。

两个写作技能形成互补关系。brand-writer 技能负责"按 Zed 品牌声口从零写文案"——它定义品牌语气(流畅、平静、直接)、核心信息与写作原则,并在"避免"清单里同样列出了 AI/营销陈词(破折号、对镜像句式、"it's not X, it's Y" 结构、buzzwords);而 humanizer 负责"对已有草稿做去 AI 化审查",其描述里 "Use after drafting" 的措辞正说明它处于流水线的后段。可以推断,团队内部的写作 Agent 会话大致是:用 brand-writer 类技能起草,再交给 humanizer 做模式审查与终审,配合 allowed-tools 中的 Grep 对全文做信号词扫描、Edit 落盘修改。

对想在自己的项目里复制这套做法的团队,可复用的要素按 SKILL.md 原文完整列出:24 项模式的信号词表(可直接转成 Grep 正则清单)、"人格与灵魂"六条手法、九步流程中的自审计回路,以及四段式输出格式。

小结

humanizer 技能把"文本像不像 AI 写的"这个模糊问题拆成了一个可执行的工程问题:24 类有明确信号词的模式负责识别,九步流程(含自审计回路)负责操作,四段式输出负责留痕。它解决的问题不是事实性错误,而是 LLM 统计倾向带来的表达指纹——意义膨胀、三段排比、系动词回避、五重限定、聊天机器人口癖。无论是审查博客文章、产品文档还是 Agent 生成的 PR 描述,这份清单加配套流程都能直接落地。