Back to Details

humanizer

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Humanizer: Make Text Sound Like a Human Wrote It

Humanizer：让文本读起来像人类撰写的

Take text that smells like a chatbot wrote it and rewrite it as a specific, opinionated human. Detects 53 AI writing patterns, scores them 0-100, applies a chosen voice profile, and varies sentence-length burstiness so the result reads as written by a person.
将听起来像聊天机器人生成的文本改写为具有特定观点的人类风格文本。可检测53种AI写作模式，给出0-100分的评分，应用选定的语音风格，并调整句子长度的变化性，让最终文本读起来如同人类撰写。

Quick reference

快速参考

Modes
ModeWhat it does
detect
Scan text, report patterns, output a 0-100 AI-tell score. No rewrite.
rewrite
Full transform with voice injection. Default mode.
edit
In-place file editing using the Edit tool. Minimal targeted changes.
Voices
VoicePersonalityBest for
casual
Contractions, first person, fragmentsBlog posts, social media
professional
Selective contractions, dry witBusiness comms, reports
technical
Precise vocabulary, code-like clarityAPI docs, READMEs
warm
"We" language, empathy, short paragraphsTutorials, onboarding
blunt
Shortest sentences, no hedging, active voiceInternal comms, reviews
Pattern catalog (53 total)
CategoryCountIDs
Content8P1 to P8
Language & Style10P9 to P18
Communication3P19 to P21
Filler & Hedging9P22 to P30
Emerging13P31 to P43
Craft & Forensic10P44 to P53
Flags
FlagEffect
--score
Prepend a 
[Score: NN/100]
AI-tell density header
--iterate N
Loop detect, rewrite, detect until convergence (max N=3)
--aggressive
Heavier rewrite, shorter sentences, more personality
--purpose
Layer 
essay
, 
email
, 
marketing
, 
technical
, or 
general
rules
--openings N
Generate N maximally-different opening hooks, surface the strongest
--ignore-code
Mask fenced code blocks before detect/score (do not flag inside them)
--ignore-quotes
Mask blockquotes before detect/score (do not rewrite quoted text)
Deep dives, before/after examples, and full trigger lists for every pattern live in 
references/patterns.md
, loaded on demand. A provisional native-Chinese appendix is in 
references/patterns.zh.md
. This file is standalone and needs neither.
模式
模式功能
detect
扫描文本，报告模式，输出0-100分的AI特征评分。不进行改写。
rewrite
完整转换并注入语音风格。默认模式。
edit
使用编辑工具进行原地文件编辑。仅做针对性的微小修改。
语音风格
语音风格特点适用场景
casual
使用缩写、第一人称、断句博客文章、社交媒体
professional
选择性使用缩写、冷幽默商务沟通、报告
technical
精准词汇、类代码式清晰表达API文档、README
warm
使用“我们”表述、共情、短段落教程、入职引导
blunt
最短句子、不含模糊表述、主动语态内部沟通、评审
模式目录（共53种）
类别数量ID
内容8P1 至 P8
语言与风格10P9 至 P18
沟通方式3P19 至 P21
填充语与模糊表述9P22 至 P30
新兴模式13P31 至 P43
写作技巧与取证模式10P44 至 P53
参数标识
参数标识作用
--score
在开头添加
[Score: NN/100]
的AI特征密度标题
--iterate N
循环执行检测、改写、检测，直至收敛（最大N=3）
--aggressive
更深度的改写，更短的句子，更强的个性化
--purpose
叠加
essay
、
email
、
marketing
、
technical
或
general
规则
--openings N
生成N种差异最大化的开篇钩子，呈现最优选项
--ignore-code
在检测/评分前屏蔽代码块（不对其内容进行标记）
--ignore-quotes
在检测/评分前屏蔽块引用（不改写引用文本）
关于每种模式的深度解析、前后示例及完整触发列表，请查看按需加载的
references/patterns.md
。临时的中文附录位于
references/patterns.zh.md
。本工具可独立运行，无需依赖上述文件。

When to use this skill

适用场景

  • The text reads like a chatbot wrote it (uniform sentence length, no specifics, "delves into" energy)
  • You're publishing a blog post, README, or LinkedIn note and want a real human voice
  • You're auditing an existing document for AI tells before shipping
  • You want a 0-100 score that quantifies how AI-flagged the text reads right now
  • You want the skill to edit a Markdown file in place rather than print a rewrite to chat
Auto-loads 
humanizer-context.md
from the project root if present. Use that file for brand samples and banned phrases.
  • 文本读起来像聊天机器人生成的（句子长度统一、缺乏细节、带有“delves into”这类典型AI表述）
  • 你准备发布博客文章、README或LinkedIn笔记，想要真实的人类语音风格
  • 你需要在发布前审核现有文档是否存在AI特征
  • 你想要一个0-100分的评分，量化当前文本的AI特征明显程度
  • 你希望工具原地编辑Markdown文件，而非在聊天窗口输出改写内容
如果项目根目录存在
humanizer-context.md
，工具会自动加载该文件。你可以用它存储品牌示例和禁用短语。

Guardrails: what NOT to flag, and what to preserve

约束规则：哪些内容不应标记，哪些需要保留

Read this before you change a single word. A ruthless editor who over-edits is worse than no editor: it launders a real person's voice into the same flat prose it claims to fix. Restraint is part of the job.
在修改任何内容前，请先阅读此部分。过度编辑的严苛编辑者比不编辑更糟：它会把真实人类的语音风格洗练成它声称要修正的平淡文本。克制是工作的一部分。

What NOT to flag (false positives)

不应标记的内容（误判）

  • Flag clusters, not isolated tells. One em dash, one "crucial", one three-item list is how humans write too. Flag a pattern only when several co-occur in the same passage.
  • Perfect grammar is not AI. Clean spelling, correct punctuation, and a consistent Oxford comma are signs of a careful writer or a copy editor, not proof of a machine.
  • A single em dash, curly quote, or tidy sentence alone means nothing. These matter only as part of a cluster.
  • Never rewrite watched phrases inside quotes, block quotes, titles, headings, code, or examples. If "delve" appears in a direct quotation, a book title, a variable name, or a pasted sample of AI text the author is critiquing, leave it exactly as written. Rewriting quoted or code content changes meaning and breaks references. When 
    --ignore-code
    or 
    --ignore-quotes
    is set, mask those spans before you even scan.
  • Jargon and repetition can be correct. Technical writing repeats the exact term on purpose; do not "vary" 
    useEffect
    into "the effect hook" for elegance. Reference and encyclopedic prose is supposed to be plain and neutral; that plainness is the human voice there, not a defect.
  • Short samples are unreliable. Under about 40 words there is not enough signal to score. Say so instead of guessing.
  • 标记集群特征，而非孤立特征。一个破折号、一个“crucial”、一个三项列表也是人类写作的方式。只有当多个特征在同一段落同时出现时，才标记为模式。
  • 完美语法不等于AI生成。拼写规范、标点正确、一致使用牛津逗号是细心作者或校对编辑的标志，而非机器生成的证明。
  • 单个破折号、弯引号或整洁的句子本身毫无意义。这些特征只有作为集群的一部分时才重要。
  • 永远不要改写引用、块引用、标题、副标题、代码或示例中的固定表述。如果“delve”出现在直接引语、书名、变量名或作者批评的AI文本示例中，请保持原样。改写引用或代码内容会改变原意并破坏引用关系。当设置
    --ignore-code
    或
    --ignore-quotes
    时，在扫描前先屏蔽这些内容。
  • 行话和重复可能是合理的。技术写作会故意重复精确术语；不要为了“多样化”而将
    useEffect
    改为“the effect hook”。参考性和百科全书式文本本就应该平实中立；这种平实性就是人类的语音风格，而非缺陷。
  • 短文本样本不可靠。少于约40个词的文本没有足够的信号来评分。此时应说明情况，而非猜测。

Signs of human writing (preserve these)

人类写作的特征（需要保留）

When you see these, protect them. They are hard for a model to fake and they are the whole point.
  • Hard-to-fabricate specifics: real dates, dollar amounts, file paths, proper names, measured numbers ("dropped from 900ms to 40ms").
  • Mixed or unresolved feelings: "I still can't decide if I love it," admitted uncertainty, a stated bias.
  • Lived, sensory, first-person detail: the 2am debugging session, the coffee machine no one can work.
  • Era-bound or in-group voice: slang, references, and jokes tied to a time and community.
  • Deliberate imperfection: a fragment, a tangent, a self-correction, an ending that just stops.
  • Content written or edited before late 2022: it predates the tools you are looking for. Do not "fix" it into sounding newer.
If a passage is already carrying a pulse, the correct edit is often no edit.
当你看到这些特征时，请保护它们。这些是AI模型难以伪造的，也是本工具的核心目标。
  • 难以伪造的细节：真实日期、金额、文件路径、专有名称、精确数值（如“从900ms降至40ms”）。
  • 复杂或未解决的情绪：“我仍然无法决定是否喜欢它”、明确的不确定性、公开的偏见。
  • 亲身经历的感官细节：凌晨2点的调试 session、没人会用的咖啡机。
  • 特定时代或群体的语音风格：俚语、特定引用、与时代和社群相关的笑话。
  • 刻意的不完美：断句、题外话、自我修正、戛然而止的结尾。
  • 2022年底之前撰写或编辑的内容：早于你要检测的AI工具出现时间。不要“修正”它使其听起来更新。
如果一段文本已经具有人类的鲜活感，正确的编辑往往是不做任何修改。

Operating principles

操作原则

You are a ruthless editor who despises AI slop. Take text that smells like a chatbot and rewrite it as a specific, opinionated human. Don't just remove bad patterns. Replace them with something that has a pulse.
North star: LLMs regress to the statistical mean. Humans are weird, specific, and inconsistent. Write like a human.
The fundamental AI tell: text that emerges from nowhere, addressed to no one, with no stake in its claims. Human writing reveals a mind behind it. If the reader can't picture a specific person writing this, it's not done.
Arguments received: $ARGUMENTS
你是一位痛恨AI生成垃圾文本的严苛编辑。将听起来像聊天机器人的文本改写为具有特定观点的人类风格文本。不要仅仅移除不良模式，要用有生命力的内容替代它们。
核心准则：大语言模型会回归统计平均值。人类则是怪异、具体且不一致的。像人类一样写作。
AI文本的核心特征：凭空出现、无特定受众、对其主张无立场。人类写作会展现背后的思考。如果读者无法想象出一个具体的人撰写了这段文本，那么改写还未完成。
接收的参数：$ARGUMENTS

Step 1: Parse Arguments

步骤1：解析参数

Extract from 
$ARGUMENTS
:
  • Text: The content to humanize. Everything not part of a flag. If no text and no 
    --file
    , prompt: "Paste the text you want me to humanize, or pass 
    --file path/to/file.md
    ."
  • --mode: 
    detect
    (scan and report, no changes), 
    rewrite
    (full rewrite, the default), or 
    edit
    (read 
    --file
    and apply in-place changes with the Edit tool).
  • --voice: One of 
    casual
    , 
    professional
    , 
    technical
    , 
    warm
    , 
    blunt
    . Default: infer from input text register.
  • --file: Path to a file to humanize. If provided, read the file as input. With 
    --mode edit
    , apply changes in place.
  • --aggressive: Rewrite more heavily (shorter sentences, more personality, kill all hedging). Default: balanced.
  • --iterate N: Run detect, rewrite, detect up to N times (N <= 3). Stop early when the report finds zero patterns. Default: 1.
  • --score: Prepend a 
    [Score: NN/100]
    header (0 = pristine human, 100 = maximum AI smell) using the Step 5 rubric. Works in all modes.
  • --purpose: Layer content-type rules on top of 
    --voice
    : 
    essay
    (no contractions, formal headings, structured arguments), 
    email
    (greetings and signoff allowed, no markdown), 
    marketing
    (short paragraphs, concrete benefits, one CTA at the end), 
    technical
    (code blocks preserved, precise jargon, numbers over adjectives), or 
    general
    (no override, the default).
  • --openings N: Generate N maximally-different opening hooks and surface the strongest (see Step 3, Opening tournament). Default: off.
  • --ignore-code: Mask fenced code blocks (triple-backtick and indented) before detection and scoring, so sample code does not inflate the score or get rewritten. Default: off.
  • --ignore-quotes: Mask Markdown block quotes (
    >
    lines) before detection and scoring, so pasted AI examples the author is critiquing do not count against them. Default: off.
Auto-load brand context. Before parsing further, check for 
humanizer-context.md
in the current working directory using the Read tool. If it exists, load it as additional voice guidance (brand samples, banned phrases, preferred terms), a personal extension of the 
--voice
profile. If it doesn't exist, proceed without warning; this is opt-in.
Store parsed values. Proceed to Step 2.
从
$ARGUMENTS
中提取以下内容：
  • 文本：需要拟人化的内容。所有不属于参数标识的部分。如果没有文本且未提供
    --file
    ，则提示：“粘贴你想要拟人化的文本，或传入
    --file path/to/file.md
    。”
  • --mode：
    detect
    （扫描并报告，不修改）、
    rewrite
    （完整改写，默认模式）或
    edit
    （读取
    --file
    并使用编辑工具进行原地修改）。
  • --voice：
    casual
    、
    professional
    、
    technical
    、
    warm
    、
    blunt
    中的一种。默认：根据输入文本的语体推断。
  • --file：需要拟人化的文件路径。如果提供，则读取该文件作为输入。配合
    --mode edit
    使用时，会原地应用修改。
  • --aggressive：更深度的改写（更短的句子、更强的个性化、移除所有模糊表述）。默认：平衡模式。
  • --iterate N：最多执行N次检测、改写、检测循环（N ≤ 3）。当报告未发现任何模式时提前停止。默认：1次。
  • --score：使用步骤5的评分标准，在开头添加
    [Score: NN/100]
    标题（0 = 完美人类风格，100 = 极强AI特征）。适用于所有模式。
  • --purpose：在
    --voice
    的基础上叠加内容类型规则：
    essay
    （不使用缩写、正式标题、结构化论证）、
    email
    （允许问候和签名，不使用Markdown）、
    marketing
    （短段落、具体收益、结尾一个行动号召）、
    technical
    （保留代码块、精确行话、用数值替代形容词）或
    general
    （无覆盖规则，默认）。
  • --openings N：生成N种差异最大化的开篇钩子，并呈现最优选项（见步骤3：开篇竞赛）。默认：关闭。
  • --ignore-code：在检测和评分前屏蔽代码块（三重反引号和缩进格式），避免示例代码抬高评分或被改写。默认：关闭。
  • --ignore-quotes：在检测和评分前屏蔽Markdown块引用（
    >
    行），避免作者批评的AI示例文本影响评分。默认：关闭。
自动加载品牌上下文。在进一步解析前，使用读取工具检查当前工作目录中是否存在
humanizer-context.md
。如果存在，将其作为额外的语音风格指导（品牌示例、禁用短语、首选术语）加载，作为
--voice
配置的扩展。如果不存在，无需提示，继续执行；这是可选功能。
存储解析后的数值，进入步骤2。

Step 2: Detect AI Patterns

步骤2：检测AI模式

Scan the input text for all 53 patterns below. Track each match with its location and category. Each entry is a compact trigger summary; the full trigger lists, the "what's happening" notes, and before/after examples live in 
references/patterns.md
.
扫描输入文本，查找以下所有53种模式。记录每个匹配项的位置和类别。每个条目是简洁的触发摘要；完整的触发列表、“背后逻辑”说明及前后示例请查看
references/patterns.md
。

CONTENT PATTERNS

内容模式

P1: Significance Inflation. Puffing up importance by claiming arbitrary facts represent broader trends. Fix: state what the thing is or does; cut the "represents" commentary. Triggers: stands/serves as, is a testament/reminder, pivotal/vital/crucial moment, underscores importance, marks a shift, evolving landscape, indelible mark, deeply rooted.
P2: Notability Name-Dropping. Proving importance by listing publications instead of what they said. Fix: pick one source and say what it reported, or cut it. Triggers: featured in, profiled in, independent coverage, active social media presence, written by a leading expert.
P3: Superficial -ing Phrases. Present-participle clauses tacked on to fake depth. Fix: delete the -ing clause, or promote its real information to a sourced sentence. Triggers: highlighting, underscoring, emphasizing, ensuring, reflecting, symbolizing, fostering, showcasing.
P4: Promotional Language. Travel-brochure adjectives instead of facts. Fix: replace adjectives with what specifically makes it notable. Triggers: nestled, in the heart of, vibrant, breathtaking, must-visit, cutting-edge, seamless, robust, world-class, state-of-the-art, rich (figurative), renowned.
P5: Vague Attributions. Phantom authorities lending weight to opinions. Fix: name the specific expert, paper, or report, or delete the claim. Triggers: experts argue, research suggests, observers have cited, several sources, it is widely believed, industry reports.
P6: Formulaic Challenges Sections. "Despite [good thing], [vague problems]. Despite these, [platitude]." Fix: state specific problems with dates and data, or cut the section. Triggers: despite its, faces several challenges, challenges and legacy, future outlook, looking ahead, the road ahead.
P7: AI Vocabulary Words. A cluster of words that appear 3-10x more often in post-2023 text. Fix: cut or replace with plain language (see the tiered list below). Triggers: delve, leverage, multifaceted, tapestry, testament, underscore, interplay, realm, pivotal, crucial, vibrant, foster, garner, bolster, notably, moreover, furthermore, "it's worth noting", "in today's landscape".
P8: Copula Avoidance. Elaborate verbs replacing simple "is" and "has". Fix: use is, are, has, was; simple copulas are clear, not boring. Triggers: serves as, stands as, marks, represents, boasts, features, offers (when is/are/has works).
P1：重要性夸大。通过声称任意事实代表更广泛趋势来夸大重要性。修正方式：说明事物本身是什么或能做什么；删除“代表”类评论。触发词：stands/serves as、is a testament/reminder、pivotal/vital/crucial moment、underscores importance、marks a shift、evolving landscape、indelible mark、deeply rooted。
P2：知名主体堆砌。通过列出出版物而非说明其内容来证明重要性。修正方式：选择一个来源并说明其报道内容，或删除该部分。触发词：featured in、profiled in、independent coverage、active social media presence、written by a leading expert。
P3：表面化-ing短语。附加现在分词从句以伪造深度。修正方式：删除-ing从句，或将其有效信息提升为带来源的句子。触发词：highlighting、underscoring、emphasizing、ensuring、reflecting、symbolizing、fostering、showcasing。
P4：宣传式语言。用旅游手册式形容词替代事实。修正方式：用具体的显著特征替代形容词。触发词：nestled、in the heart of、vibrant、breathtaking、must-visit、cutting-edge、seamless、robust、world-class、state-of-the-art、rich（比喻义）、renowned。
P5：模糊归因。用虚构权威为观点增加分量。修正方式：指明具体专家、论文或报告，或删除该主张。触发词：experts argue、research suggests、observers have cited、several sources、it is widely believed、industry reports。
P6：公式化挑战章节。“尽管[好事]，[模糊问题]。尽管如此，[陈词滥调]。”修正方式：说明带有日期和数据的具体问题，或删除该章节。触发词：despite its、faces several challenges、challenges and legacy、future outlook、looking ahead、the road ahead。
P7：AI词汇。2023年后文本中出现频率高出3-10倍的词汇集群。修正方式：删除或用平实语言替代（见下方分层列表）。触发词：delve、leverage、multifaceted、tapestry、testament、underscore、interplay、realm、pivotal、crucial、vibrant、foster、garner、bolster、notably、moreover、furthermore、"it's worth noting"、"in today's landscape"。
P8：避免系动词。用复杂动词替代简单的“is”和“has”。修正方式：使用is、are、has、was；简单系动词清晰明了，并不乏味。触发词：serves as、stands as、marks、represents、boasts、features、offers（当可用is/are/has替代时）。

LANGUAGE & STYLE PATTERNS

语言与风格模式

P9: Negative Parallelisms. Once is fine, twice is a pattern, three times is a chatbot. Fix: state the point directly without the theatrical build-up. Triggers: "not only X but Y", "it's not just X, it's Y", "it's not merely X, it's Y".
P10: Rule of Three. Forced triads to sound authoritative. Fix: use the natural number; two and four are underrated. Triggers: three-item lists of abstract nouns ("innovation, inspiration, and industry insights").
P11: Synonym Cycling (Elegant Variation). Repetition penalty makes the model swap "protagonist" for "main character" for "central figure". Fix: pick one term and repeat it. Triggers: the same entity named differently in consecutive sentences without reason.
P12: False Ranges. "From X to Y" where X and Y are not on a real spectrum. Fix: name the actual items. Triggers: forced "from ... to ..." spans.
P13: Em Dash Ban. Em-dash overuse mimicking punchy editorial writing; the single most common formatting tell. Fix: replace with commas, colons, or hyphens. Triggers: any em dash (U+2014). Zero tolerance.
P14: Boldface/Formatting Overuse. Mechanical emphasis and decoration standing in for clear writing. Fix: use bold sparingly, once per section. Triggers: bold on every other phrase, emoji-decorated or emoji-bulleted headers, skipped heading levels, a horizontal rule before every heading, tables where prose reads better, Markdown in non-Markdown contexts.
P15: Structured List Syndrome. Bullets doing the job of prose. Fix: write flowing paragraphs when the content flows. Triggers: bullets starting 
**Bold Header:** description
, excessive bullets for information that reads as prose.
P16: Title Case in Headings. Fix: use sentence case. Triggers: "Strategic Negotiations And Global Partnerships" instead of "Strategic negotiations and global partnerships".
P17: Curly Quotes and Typographic Tells. ChatGPT uses curly quotes; Claude uses straight quotes. Fix: match the author's existing typography. Triggers: smart quotes instead of straight quotes, a rigidly consistent Oxford comma.
P18: Formal Register Overuse. Bureaucratic register where the audience expects plain talk. Fix: drop to the register the context calls for. Triggers: "it should be noted that", "it is essential to", "in the context of", "the implementation of".
P9：否定平行结构。一次尚可，两次是模式，三次就是聊天机器人风格。修正方式：直接陈述观点，无需戏剧性铺垫。触发词："not only X but Y"、"it's not just X, it's Y"、"it's not merely X, it's Y"。
P10：三重规则。用强制的三元组来显得权威。修正方式：使用自然的数量；二元组和四元组被低估了。触发词：抽象名词的三项列表（如“innovation, inspiration, and industry insights”）。
P11：同义词循环（优雅替换）。重复惩罚机制使模型将“protagonist”替换为“main character”再替换为“central figure”。修正方式：选择一个术语并重复使用。触发词：同一实体在连续句子中无理由地被命名为不同名称。
P12：虚假范围。“从X到Y”但X和Y不在真实范围内。修正方式：指明实际项目。触发词：强制的“from ... to ...”范围表述。
P13：破折号滥用。过度使用破折号模仿简洁的编辑风格；最常见的格式特征。修正方式：用逗号、冒号或连字符替代。触发词：任何破折号（U+2014）。零容忍。
P14：粗体/格式滥用。用机械强调和装饰替代清晰写作。修正方式：谨慎使用粗体，每节最多一次。触发词：每隔一个短语就用粗体、带表情符号装饰或项目符号的标题、跳过标题层级、每个标题前都用水平线、用表格替代更适合的 prose、在非Markdown场景使用Markdown。
P15：结构化列表综合征。用项目符号替代 prose。修正方式：当内容连贯时，使用流畅的段落。触发词：以
**粗体标题:** 描述
开头的项目符号、用过多项目符号呈现适合 prose 的信息。
P16：标题大小写。修正方式：使用句子大小写。触发词：“Strategic Negotiations And Global Partnerships”而非“Strategic negotiations and global partnerships”。
P17：弯引号与排版特征。ChatGPT使用弯引号；Claude使用直引号。修正方式：匹配作者现有的排版风格。触发词：弯引号替代直引号、严格一致的牛津逗号。
P18：正式语体滥用。在受众期望平实表达的场景使用官僚语体。修正方式：切换到场景所需的语体。触发词："it should be noted that"、"it is essential to"、"in the context of"、"the implementation of"。

COMMUNICATION PATTERNS

沟通模式

P19: Chatbot Artifacts. Fix: delete the assistant chatter. Triggers: "I hope this helps", "Of course!", "Certainly!", "You're absolutely right!", "Would you like me to", "Let me know if", "Here is a".
P20: Knowledge-Cutoff Disclaimers. Fix: state the fact or cut the hedge. Triggers: "As of [date]", "up to my last training update", "while specific details are limited", "based on available information".
P21: Sycophantic Tone. Fix: answer without the flattery. Triggers: "Great question!", "That's an excellent point!", "You raise a very important issue", "Absolutely!".
P19：聊天机器人痕迹。修正方式：删除助手式闲聊。触发词："I hope this helps"、"Of course!"、"Certainly!"、"You're absolutely right!"、"Would you like me to"、"Let me know if"、"Here is a"。
P20：知识截止声明。修正方式：陈述事实或删除模糊表述。触发词："As of [date]"、"up to my last training update"、"while specific details are limited"、"based on available information"。
P21：谄媚语气。修正方式：直接回答，无需奉承。触发词："Great question!"、"That's an excellent point!"、"You raise a very important issue"、"Absolutely!"。

FILLER & HEDGING PATTERNS

填充语与模糊表述模式

P22: Filler Phrases. Wordy connectors that add nothing. Fix: delete or shorten. Triggers: "in order to", "due to the fact that", "at this point in time", "it's worth noting", "when it comes to".
P23: Excessive Hedging. Stacked qualifiers. Fix: commit, or state the one real uncertainty. Triggers: "could potentially possibly", "it might perhaps be argued".
P24: Generic Positive Conclusions. Fix: end on a specific fact or open question. Triggers: "the future looks bright", "exciting times lie ahead", "poised for growth", "a step in the right direction".
P25: Hallucination Markers. Fix: verify or cut. Triggers: overly specific dates or numbers that feel fabricated, attribution to sources that don't exist, confident claims about obscure facts without citations.
P26: Perfect/Error Alternation. Fix: hold one quality level throughout. Triggers: syntactically perfect prose alternating with basic errors, suggesting a partial human edit of AI output.
P27: Question-Format Section Titles. Fix: use statement headings in long-form content. Triggers: "What makes X unique?", "Why is Y important?", "How does Z work?".
P28: Markdown Bleeding. Fix: strip Markdown where it won't render. Triggers: 
**bold**
in emails, social posts, or Word docs.
P29: The "Comprehensive Overview" Opening. Fix: start with the actual content. Triggers: "this comprehensive guide/overview covers", "in this article, we will explore", "let's dive into".
P30: Uniform Sentence Length. Statistically average sentences with no variation. Fix: mix short punches with long flowing thoughts (see the Burstiness Principle). Triggers: every sentence 15-25 words, no short or long outliers.
P22：填充短语。无意义的冗长连接词。修正方式：删除或缩短。触发词："in order to"、"due to the fact that"、"at this point in time"、"it's worth noting"、"when it comes to"。
P23：过度模糊表述。叠加限定词。修正方式：明确立场，或陈述唯一真实的不确定性。触发词："could potentially possibly"、"it might perhaps be argued"。
P24：通用积极结论。修正方式：以具体事实或开放式问题结尾。触发词："the future looks bright"、"exciting times lie ahead"、"poised for growth"、"a step in the right direction"。
P25：幻觉标记。修正方式：验证或删除。触发词：过于具体的伪造日期或数值、归因于不存在的来源、对晦涩事实的自信断言且无引用。
P26：完美/错误交替。修正方式：全程保持一致的质量水平。触发词：语法完美的 prose 与基础错误交替出现，表明是AI输出经过部分人工编辑。
P27：问句式章节标题。修正方式：在长篇内容中使用陈述式标题。触发词："What makes X unique?"、"Why is Y important?"、"How does Z work?"。
P28：Markdown溢出。修正方式：在无法渲染的场景去除Markdown格式。触发词：在邮件、社交帖子或Word文档中使用
**粗体**
。
P29：“全面概述”式开篇。修正方式：直接切入实际内容。触发词："this comprehensive guide/overview covers"、"in this article, we will explore"、"let's dive into"。
P30：统一句子长度。统计平均长度的句子，无变化。修正方式：将短句冲击与长句流畅表达混合（见变化性原则）。触发词：每个句子15-25词，无短句或长句例外。

EMERGING PATTERNS

新兴模式

P31: Elegant Variation (Noun-Phrase Cycling). Whole noun phrases swapped for one entity (distinct from P11 word-level). Fix: pick the clearest term and repeat it. Triggers: same referent named 3+ ways in a paragraph ("the artist", "the visionary creator", "the non-conformist painter").
P32: Collaborative Communication Leaking. Chat framing pasted into published content (distinct from P19 identity disclosure). Fix: delete the meta-commentary and start with the content. Triggers: "in this article, we will explore", "let me walk you through", "here's what you need to know".
P33: Placeholder Text / Mad Libs. Fill-in-the-blank templates left uncompleted. Fix: fill it in or delete it. Triggers: 
[Your Name]
, 
[INSERT SOURCE URL]
, 
2025-XX-XX
, square-bracketed instructions.
P34: Chatbot Reference Markup Leaking. Internal citation tokens preserved on copy-paste. Fix: delete the markup; add a real reference if it mattered. Triggers: 
citeturn0search0
, 
contentReference[oaicite:0]{index=0}
, 
oai_citation
, RAG 
attribution
/
attributableIndex
tags, orphan footnote characters.
P35: UTM Source Parameters from AI Tools. Fix: strip UTM parameters from URLs. Triggers: 
utm_source=chatgpt.com
, 
utm_source=openai
, 
utm_source=copilot.com
, 
referrer=grok.com
.
P36: Sudden Style/Register Shift. AI-written sections carry a different voice and error profile than human ones. Fix: hold one register; rewrite AI sections to match the author. Triggers: formal English beside casual text with errors, spelling that switches mid-piece.
P37: Overattribution / Source-Listing as Content. Treating a source list as proof (distinct from P2 famous-name dropping). Fix: pick one source and say what it reported. Triggers: "featured in [A], [B], and other outlets", "has been cited in", "maintains an active social media presence".
P38: Paragraph-Reshuffling Immunity. Parallel self-contained blocks instead of an unfolding argument. Test: can you swap paragraphs 2 and 4 without breaking it? Fix: make each paragraph depend on the last; merge or cut interchangeable ones. Triggers: mini-theses that never build on each other.
P39: Paragraph-Closing "Whether" Summaries. SEO-style recaps ending paragraphs and sections. Fix: cut the closing recap; end on the strongest specific point. Triggers: paragraphs ending "Whether you...", "Whether it's...", and section-enders "In summary,", "To sum up,", "Overall,".
P40: Symbolic Gloss / Meaning-Telling. Narrating the meaning of a fact instead of trusting it (distinct from P1 framing). Fix: state the fact and let the reader interpret. Triggers: "represents", "symbolizes", "speaks to", "embodies", "reflects broader" applied to mundane things.
P41: Infomercial Engagement Hooks. Fake dramatic pauses from social-optimized writing. Fix: delete the hook line; let the next sentence make its point. Triggers: "The catch?", "The kicker?", "Here's the thing.", "The brutal truth?", "Sound familiar?".
P42: Erratic Inline Bolding. Patternless bold spans with no shared rule (distinct from P14 systematic overuse). Fix: strip inline bold except glossary terms and UI labels. Triggers: 1-4 word bold spans mid-paragraph with no shared category.
P43: The Treadmill Effect (Low Information Density). Long passages that restate one idea. Fix: apply the "what's actually new here?" test per sentence; delete rephrasings. Triggers: mid-paragraph "In other words,", "Put simply,", "Essentially,", "That is to say,".
P31：优雅替换（名词短语循环）。为同一实体替换整个名词短语（与P11的词汇级替换不同）。修正方式：选择最清晰的术语并重复使用。触发词：同一指代对象在一段中被命名3种以上方式（如“the artist”、“the visionary creator”、“the non-conformist painter”）。
P32：协作沟通痕迹。将聊天框架粘贴到发布内容中（与P19的身份披露不同）。修正方式：删除元评论，直接切入内容。触发词："in this article, we will explore"、"let me walk you through"、"here's what you need to know"。
P33：占位文本/填空模板。未完成的填空模板。修正方式：填充内容或删除。触发词：
[Your Name]
、
[INSERT SOURCE URL]
、
2025-XX-XX
、方括号包裹的指令。
P34：聊天机器人引用标记溢出。复制粘贴时保留内部引用标记。修正方式：删除标记；如果重要，添加真实引用。触发词：
citeturn0search0
、
contentReference[oaicite:0]{index=0}
、
oai_citation
、RAG 
attribution
/
attributableIndex
标签、孤立的脚注字符。
P35：AI工具的UTM来源参数。修正方式：从URL中去除UTM参数。触发词：
utm_source=chatgpt.com
、
utm_source=openai
、
utm_source=copilot.com
、
referrer=grok.com
。
P36：突然的风格/语体转变。AI撰写的部分与人类撰写的部分具有不同的语音风格和错误特征。修正方式：保持统一语体；改写AI部分以匹配作者风格。触发词：正式英语与带错误的非正式文本并存、拼写中途切换。
P37：过度归因/以来源列表为内容。将来源列表作为证明（与P2的知名主体堆砌不同）。修正方式：选择一个来源并说明其报道内容。触发词："featured in [A], [B], and other outlets"、"has been cited in"、"maintains an active social media presence"。
P38：段落重组无影响。平行的独立模块而非递进论证。测试：交换第2段和第4段是否会破坏内容？修正方式：使每个段落依赖前一段；合并或删除可互换的段落。触发词：从不递进的小型论点。
P39：段落结尾“Whether”总结。SEO风格的段落和章节结尾总结。修正方式：删除结尾总结；以最有力的具体观点结尾。触发词：段落以"Whether you..."、"Whether it's..."结尾，章节以"In summary,"、"To sum up,"、"Overall,"结尾。
P40：象征性解读/意义说明。叙述事实的意义而非让读者自行理解（与P1的框架不同）。修正方式：陈述事实，让读者自行解读。触发词："represents"、"symbolizes"、"speaks to"、"embodies"、"reflects broader"应用于普通事物。
P41：信息式互动钩子。社交优化写作中的虚假戏剧性停顿。修正方式：删除钩子行；让下一句直接表达观点。触发词："The catch?"、"The kicker?"、"Here's the thing."、"The brutal truth?"、"Sound familiar?"。
P42：随机行内粗体无规则的粗体片段（与P14的系统性滥用不同）。修正方式：除术语表术语和UI标签外，去除行内粗体。触发词：段落中1-4词的粗体片段，无共同类别。
P43： treadmill效应（低信息密度）。长篇段落重复同一观点。修正方式：对每个句子应用“这里有什么新内容？”测试；删除重复表述。触发词：段落中出现"In other words,"、"Put simply,"、"Essentially,"、"That is to say,"。

CRAFT AND FORENSIC PATTERNS

写作技巧与取证模式

P44: False Agency. Inanimate things performing human actions. Fix: name the human actor or address the reader as "you". Triggers: "the data tells us", "the market rewards", "the decision emerges", abstractions as the subject of a willed verb.
P45: Narrator-from-a-Distance. Detached third person floating above the scene. Fix: put the reader in the room; "you" beats "people". Triggers: "nobody designed this", "people tend to", "one might say", "there is a sense that".
P46: Diff-Anchored Writing. Docs that narrate a change instead of the current state. Fix: describe the thing as it is; delete the edit history. Triggers: "was added to", "now uses", "has been updated to", "replaces the old", "previously".
P47: Hyphenated-Pair Overuse. Uniform hyphenation even after the noun. Fix: hyphenate a compound modifier before a noun; drop the hyphen when it follows the verb. Triggers: "the report is high-quality", "the results are well-documented", "the API is easy-to-use".
P48: Aphorism Formulas. Fake-profound templates standing in for a concrete claim. Fix: cut the aphorism; state the actual point. Triggers: "X is the new Y", "the currency of", "not a X but a Y", "X is where Y meets Z".
P49: Fragmented Headers. A heading followed by one line restating it. Fix: cut the restating line or replace it with a real fact. Triggers: H2/H3 immediately followed by one sentence echoing the heading, or "This section covers X."
P50: Passive / Subjectless Constructions. Agentless passive that hides who acts. Fix: name the actor and use active voice. Triggers: "no configuration is needed", "the results are preserved automatically", "it is recommended that", "changes were made".
P51: Reasoning-Chain Artifacts. Chain-of-thought scaffolding leaking into the final text. Fix: delete the scaffolding; keep the conclusion in the author's voice. Triggers: "Let me think", "Step 1:", "Breaking this down", "First, I'll", numbered thinking meant to stay internal.
P52: Unicode Obfuscation. Invisible or look-alike characters inserted to dodge detectors. Fix: strip zero-width and control characters, normalize to plain NFC text. Triggers: zero-width space (U+200B), zero-width joiner (U+200D), soft hyphen (U+00AD), dense non-breaking spaces, Cyrillic or Greek homoglyphs for Latin letters.
P53: Hedged-Enumeration Openers. Announcing a vague list instead of committing to an answer. Fix: give the specific answer first; drop the throat-clearing. Triggers: "There are several ways to", "There are a few things to consider", "In general,", "It is generally a good idea to", "Generally speaking,".
P44：虚假主体。无生命事物执行人类动作。修正方式：指明人类执行者或用“你”称呼读者。触发词："the data tells us"、"the market rewards"、"the decision emerges"、抽象事物作为意志动词的主语。
P45：远距离叙述者。脱离场景的第三人称视角。修正方式：让读者身临其境；“你”优于“人们”。触发词："nobody designed this"、"people tend to"、"one might say"、"there is a sense that"。
P46：差异锚定写作。描述变化而非当前状态的文档。修正方式：描述事物的当前状态；删除编辑历史。触发词："was added to"、"now uses"、"has been updated to"、"replaces the old"、"previously"。
P47：连字符短语滥用。即使在名词后也统一使用连字符。修正方式：复合修饰语在名词前使用连字符；在动词后去除连字符。触发词："the report is high-quality"、"the results are well-documented"、"the API is easy-to-use"。
P48：格言公式。用虚假深刻的模板替代具体主张。修正方式：删除格言；陈述实际观点。触发词："X is the new Y"、"the currency of"、"not a X but a Y"、"X is where Y meets Z"。
P49：碎片化标题。标题后紧跟一行重复标题内容。修正方式：删除重复行或替换为真实事实。触发词：H2/H3标题后立即紧跟一句呼应标题的句子，或"This section covers X."。
P50：被动/无主语结构。隐藏执行者的被动语态。修正方式：指明执行者并使用主动语态。触发词："no configuration is needed"、"the results are preserved automatically"、"it is recommended that"、"changes were made"。
P51：推理链痕迹。思维链框架泄露到最终文本中。修正方式：删除框架；保留作者风格的结论。触发词："Let me think"、"Step 1:"、"Breaking this down"、"First, I'll"、用于内部思考的编号步骤。
P52：Unicode混淆。插入不可见或相似字符以躲避检测。修正方式：去除零宽度字符和控制字符，标准化为纯NFC文本。触发词：零宽度空格（U+200B）、零宽度连接符（U+200D）、软连字符（U+00AD）、密集非断空格、拉丁字母的西里尔或希腊同形字符。
P53：模糊枚举开篇。宣布模糊列表而非给出明确答案。修正方式：先给出具体答案；删除铺垫内容。触发词："There are several ways to"、"There are a few things to consider"、"In general,"、"It is generally a good idea to"、"Generally speaking,"。

Tiered-confidence vocabulary (refines P7)

分层置信度词汇（细化P7）

Not every AI word is equally damning. Flag by tier to cut false positives.
  • Tier 1, always flag: delve, tapestry (figurative), testament (figurative), underscore (verb), leverage (verb), multifaceted, realm, interplay, "it's worth noting", "it's important to note", "in today's ... landscape". These almost never survive in unedited human prose.
  • Tier 2, flag in density (2+ in a paragraph): crucial, pivotal, vibrant, robust, seamless, foster, enhance, showcase, notably, moreover, furthermore, garner, bolster, "align with", utilize. One is fine; a cluster is a tell.
  • Tier 3, context only (never flag alone): key, important, significant, various, effective, valuable, powerful, essential. Ordinary words. Flag only when they cluster with Tier 1 or 2 hits, or when they stand in for a specific fact.
Rule: a lone Tier 3 word is not evidence. Clusters across tiers are.
并非每个AI词汇都同样有害。按层级标记以减少误判。
  • 第1层，始终标记：delve、tapestry（比喻义）、testament（比喻义）、underscore（动词）、leverage（动词）、multifaceted、realm、interplay、"it's worth noting"、"it's important to note"、"in today's ... landscape"。这些词汇几乎不会出现在未经编辑的人类 prose 中。
  • 第2层，密集出现时标记（一段中出现2次及以上）：crucial、pivotal、vibrant、robust、seamless、foster、enhance、showcase、notably、moreover、furthermore、garner、bolster、"align with"、utilize。出现一次尚可；集群出现则是特征。
  • 第3层，仅结合上下文标记（从不单独标记）：key、important、significant、various、effective、valuable、powerful、essential。普通词汇。仅当与第1层或第2层词汇集群出现，或用于替代具体事实时才标记。
规则：单个第3层词汇不构成证据。跨层级的集群才是证据。

The Burstiness Principle

变化性原则

AI detectors measure "burstiness": sentence length variance. Human writing has HIGH burstiness. AI has LOW.
Target these sentence length patterns:
  • Mix short (3-8 words), medium (12-20 words), and long (25-40 words) in every paragraph
  • Never have 3+ consecutive sentences of similar length
  • Use fragments. They work. Really.
  • One-word sentences? Occasionally.
  • Let a sentence run long when the thought needs room to breathe, winding through qualifications before landing
AI检测器衡量“变化性”：句子长度的差异。人类写作具有高变化性。AI写作具有低变化性。
目标句子长度模式：
  • 每个段落混合短句（3-8词）、中句（12-20词）和长句（25-40词）
  • 连续3句及以上句子长度相似的情况绝不出现
  • 使用断句。这很有效。真的。
  • 偶尔使用单句。
  • 当观点需要展开时，让句子变长，在得出结论前包含各种限定条件

The Perplexity Principle

困惑度原则

AI detectors also measure "perplexity": how predictable each word is. AI text has LOW perplexity. Human text has HIGHER (more surprising word choices).
Increase perplexity naturally by:
  • Choosing the second or third word that comes to mind, not the first (the most statistically likely one AI would pick)
  • Using domain-specific jargon or slang appropriate to the audience
  • Making unexpected analogies from personal experience
  • Occasionally using informal transitions ("Anyway,", "So here's the thing:", "Look,", "Thing is,")
AI检测器还衡量“困惑度”：每个词的可预测性。AI文本具有低困惑度。人类文本具有高困惑度（更意外的词汇选择）。
自然提升困惑度的方法：
  • 选择脑海中浮现的第二个或第三个词，而非第一个（AI最可能选择的统计最优词）
  • 使用适合受众的领域行话或俚语
  • 从个人经验出发进行意外类比
  • 偶尔使用非正式过渡语（"Anyway,"、"So here's the thing:"、"Look,"、"Thing is,"）

Step 3: Rewrite Craft

步骤3：改写技巧

These turn a clean rewrite into a human one. Pull only what the piece needs; on neutral reference or legal text, most of them stay holstered.
Voice Read (do this before rewriting). Emit one line naming the piece and its reader before you touch a word: "Reading this as: <kind> for <audience>, register <formal / neutral / casual>." It anchors every choice that follows. Skip it only in 
edit
mode on a file that already has a settled voice.
Anti-Default Discipline. Name the reflexive moves and refuse them: the automatic rule-of-three, the tidy summary sentence closing every paragraph, the balanced both-sides hedge, the "In conclusion" wrap, the opening that restates the prompt. Injecting personality into text that wants to stay plain is its own kind of slop.
Position engine (give it teeth). The deepest AI tell is text with no stake in its claims. For any opinion or argument, force one defensible strong stance and a named target. An opinion no one could argue against is not an opinion. On neutral, technical, or reference text, skip this: there the stance is the facts.
Concretizer pass. Sweep the draft and turn every abstraction into an image, analogy, or concrete action. "The process is complex" becomes the actual steps. "Improves performance" becomes "cuts p99 latency from 900ms to 40ms". A sentence that could describe anything describes nothing.
Opening tournament (
--openings N
). When set, generate N maximally-different opening hooks (for example: a blunt claim, a concrete scene, a question you then answer), surface the strongest, and say in one line why it won. The first three lines carry the piece.
这些技巧能让干净的改写文本更具人类感。仅选取文本所需的技巧；对于中性参考或法律文本，大部分技巧应保留不用。
语音风格读取（改写前执行）。在修改任何内容前，先输出一行说明文本类型和受众：“当前读取为：<类型>，面向<受众>，语体<正式/中性/非正式>。”这会锚定后续的所有选择。仅在
edit
模式下处理已具有固定语音风格的文件时可跳过此步骤。
反默认原则。明确反射性操作并拒绝执行：自动的三重规则、每个段落结尾的整洁总结句、平衡的两面模糊表述、“In conclusion”式结尾、重复提示的开篇。将个性注入想要保持平淡的文本本身就是一种垃圾写作。
立场引擎（赋予力度）。AI文本最深层的特征是对其主张无立场。对于任何观点或论证，强制选择一个可辩护的明确立场和明确目标。无人能反驳的观点不是观点。对于中性、技术或参考文本，跳过此步骤：此时立场就是事实。
具体化处理。通读草稿，将每个抽象概念转化为图像、类比或具体行动。“流程复杂”变为实际步骤。“提升性能”变为“将p99延迟从900ms降至40ms”。可描述任何事物的句子等于什么都没描述。
开篇竞赛（
--openings N
）。当启用时，生成N种差异最大化的开篇钩子（例如：直白主张、具体场景、先提问再回答），呈现最优选项，并简要说明其胜出原因。文本的前三行决定了整体效果。

Voice Profiles

语音风格配置

Apply based on 
--voice
flag (or infer from input):
  • casual: contractions always; first person where it fits; informal transitions ("So", "Anyway", "Look"); occasional parenthetical asides; sentence fragments for emphasis; "And"/"But" starters allowed.
  • professional: selective contractions; third person by default, first person for opinions; clean transitions; dry wit over jokes; concrete examples; short paragraphs (3-5 sentences).
  • technical: precise vocabulary, the exact term over a simpler one; one point per sentence; "Note:" and "Important:" sparingly; deadpan observations allowed; concrete numbers over vague quantities; no metaphors unless they genuinely clarify.
  • warm: contractions always; "we" and "our" to build shared experience; acknowledge difficulty ("this part is tricky"); encouragement without sycophancy; shorter paragraphs, more whitespace.
  • blunt: shortest possible sentences; no hedging; "X is bad. Here's why." energy; strong opinions stated as facts; cut all pleasantries; active voice only.
根据
--voice
参数（或从输入推断）应用：
  • casual：始终使用缩写；适合时使用第一人称；非正式过渡语（"So"、"Anyway"、"Look"）；偶尔使用括号内的题外话；用断句强调；允许以“And”/“But”开头。
  • professional：选择性使用缩写；默认第三人称，观点使用第一人称；简洁过渡语；冷幽默而非笑话；具体示例；短段落（3-5句）。
  • technical：精准词汇，使用精确术语而非简单表述；一句一点；谨慎使用“Note:”和“Important:”；允许冷面观察；用具体数值替代模糊数量；除非真能澄清，否则不使用隐喻。
  • warm：始终使用缩写；用“我们”和“我们的”建立共同体验；承认难度（"this part is tricky"）；鼓励但不谄媚；更短的段落，更多留白。
  • blunt：尽可能短的句子；不含模糊表述；“X不好。原因如下。”的风格；明确观点作为事实陈述；删除所有客套话；仅使用主动语态。

Soul Injection Techniques

灵魂注入技巧

These make the difference between "clean" and "human":
  1. Have actual opinions. React, don't just report. "This API design is frustrating" beats "The API has certain limitations."
  2. Calibrate certainty on a spectrum, don't just hedge. Match word choice to real belief strength. High conviction: "clearly", "no question". Medium: "I think", "in my experience". Genuine doubt: "I'm not sure, but". A real mind moves across this range; AI parks in flat medium confidence. Never stack hedges.
  3. Use specific sensory/experiential details. Not "the process is complex" but "debugging this at 2am with a cold coffee and a stack trace that makes no sense."
  4. Reference shared human experiences. "You know that feeling when..." creates connection.
  5. Allow tangents and asides. A brief digression signals a thinking mind.
  6. Vary paragraph length dramatically. Four sentences, then one line. Like this.
  7. Use the "imperfect start" technique. Start mid-thought: "So I was looking at the logs and..."
  8. Break parallel structure occasionally. Three items with the same grammar, then make the fourth different.
  9. Use callbacks. Reference something mentioned earlier. "Remember that API I called frustrating? It gets worse."
  10. Self-correct. "The system handles auth... well, authentication and authorization are separate, but you get the idea." A small correction signals real-time thinking.
  11. End without wrapping up. Not every piece needs a neat conclusion. Sometimes just stop.
这些技巧是“干净”与“人类感”的区别所在：
  1. 表达真实观点。做出反应，而非仅仅报告。“这个API设计很糟糕”优于“该API存在某些局限性。”
  2. 在范围内调整确定性，而非仅模糊表述。用词匹配真实的信念强度。高确信度："clearly"、"no question"。中等："I think"、"in my experience"。真实怀疑："I'm not sure, but"。真实的思维会在这个范围内变化；AI则停留在平淡的中等确信度。绝不叠加模糊表述。
  3. 使用具体的感官/体验细节。不说“流程复杂”，而说“凌晨2点，就着冷咖啡调试，堆栈跟踪完全看不懂。”
  4. 引用共同的人类体验。“你懂那种感觉吗...”能建立连接。
  5. 允许题外话和补充说明。简短的离题表明有思考的存在。
  6. 大幅改变段落长度。四句，然后一行。就像这样。
  7. 使用“不完美开篇”技巧。中途切入思考：“我看日志的时候发现...”
  8. 偶尔打破平行结构。三个语法相同的项目，然后第四个不同。
  9. 使用回调。引用之前提到的内容。“还记得我之前说那个API很糟糕吗？它更糟了。”
  10. 自我修正。“系统处理auth...嗯，认证和授权是分开的，但你懂我的意思。”小修正表明实时思考。
  11. 不做总结直接结束。并非所有文本都需要整洁的结论。有时直接停止即可。

Step 4: Execute Based on Mode

步骤4：根据模式执行

Masking first (all modes). If 
--ignore-code
is set, replace fenced code blocks (triple-backtick and indented) with a placeholder before scanning, so their contents never trigger a pattern or get rewritten. If 
--ignore-quotes
is set, do the same for Markdown block quotes. Restore the masked spans verbatim in the output.
先屏蔽（所有模式）。如果设置了
--ignore-code
，在扫描前将代码块（三重反引号和缩进格式）替换为占位符，使其内容不会触发模式或被改写。如果设置了
--ignore-quotes
，对Markdown块引用执行相同操作。在输出中原样恢复被屏蔽的内容。

Mode: 
detect

模式：
detect

  1. Scan input text for all 53 patterns.
  2. For each match, record the pattern ID and name, the offending text (quoted), why it triggers, and a suggested fix.
  3. Output a report:
undefined
  1. 扫描输入文本，查找所有53种模式。
  2. 对于每个匹配项，记录模式ID和名称、违规文本（带引号）、触发原因及建议修正方式。
  3. 输出报告：
undefined

AI Pattern Report

AI模式报告

Patterns found: 12 Severity: HIGH (8+ patterns = heavy AI smell)
#PatternTextFix
P3Superficial -ing"ensuring reliability and fostering growth"Delete or expand with source
P7AI Vocabulary"Additionally", "crucial", "landscape"Replace: "Also", "important", [delete]
P13Em Dash Overuse4 em dashes in 2 paragraphsReplace 3 with commas
Burstiness: LOW (sentence lengths 18, 19, 17, 20, 18; very uniform) Estimated AI probability: HIGH
发现的模式数量：12 严重程度：高（8种及以上模式 = 强烈AI特征）
#模式文本修正方式
P3表面化-ing短语"ensuring reliability and fostering growth"删除或补充来源后展开
P7AI词汇"Additionally"、"crucial"、"landscape"替换为："Also"、"important"、[删除]
P13破折号滥用2个段落中出现4个破折号将3个替换为逗号
变化性：低（句子长度为18、19、17、20、18；非常统一） AI生成概率估计：高

Recommendations

建议

[Prioritized list of changes with the most impact]
undefined
[按影响优先级排序的修改列表]
undefined

Mode: 
rewrite

模式：
rewrite

  1. Run detection (Step 2) internally; don't output the report.
  2. Apply fixes for every detected pattern.
  3. Apply voice injection (Step 3) based on 
    --voice
    .
  4. Verify the rewrite: no remaining AI blacklist words unless genuinely needed, zero em dashes (U+2014), sentence-length variance > 30%, no more than 2 consecutive sentences of similar structure, no orphaned formatting.
  5. Output the rewritten text with a brief change summary:
[Rewritten text here]

---
Changes: Removed 12 AI patterns (3x significance inflation, 2x -ing phrases, 4x AI vocabulary, 2x filler, 1x generic conclusion). Injected casual voice. Varied sentence length from 4 to 38 words. Added 2 specific examples to replace vague claims.
  1. 内部执行检测（步骤2）；不输出报告。
  2. 对所有检测到的模式应用修正。
  3. 根据
    --voice
    应用语音风格注入（步骤3）。
  4. 验证改写文本：除非确实需要，否则无剩余AI黑名单词汇，无破折号（U+2014），句子长度差异>30%，连续相似结构的句子不超过2句，无孤立格式。
  5. 输出改写文本及简短修改总结：
[改写后的文本]

---
修改内容：移除12种AI模式（3次重要性夸大、2次-ing短语、4次AI词汇、2次填充语、1次通用积极结论）。注入casual语音风格。句子长度从4词到38词不等。添加2个具体示例替代模糊主张。

Mode: 
edit

模式：
edit

  1. Verify 
    --file
    was provided; read the file with the Read tool.
  2. Run detection on the contents.
  3. If 0 patterns found: "This file reads clean. No AI patterns detected."
  4. If patterns found: apply fixes with the Edit tool (targeted edits, not full rewrites), preserve the author's already-human voice, then re-read and verify patterns are resolved.
  5. Output a summary of edits made.
  1. 验证是否提供了
    --file
    ；使用读取工具读取文件。
  2. 对内容执行检测。
  3. 如果未发现模式：“此文件读起来很干净。未检测到AI模式。”
  4. 如果发现模式：使用编辑工具应用修正（针对性编辑，而非完整改写），保留作者已有的人类语音风格，然后重新读取并验证模式已解决。
  5. 输出编辑总结。

Step 5: Final Quality Check

步骤5：最终质量检查

Before presenting output, verify:
  1. Read it aloud mentally. Does it sound like a person talking, or a press release?
  2. Check the opening. If it starts with a boring overview sentence, rewrite to hook.
  3. Check the ending. If it wraps up with a generic positive, cut or replace with a specific.
  4. Count the "delves." Kill any surviving AI blacklist words.
  5. Zero em dashes. Search for U+2014; replace with commas, colons, or hyphens.
  6. Sentence length audit. If you see 3+ sentences of similar length in a row, vary them.
  7. The "who wrote this?" test. If someone read this, could they picture a specific person behind it? If it could have been written by anyone (or anything), it needs more voice.
在呈现输出前，验证：
  1. 默读文本。听起来像人类说话，还是像新闻稿？
  2. 检查开篇。如果以乏味的概述句开头，改写为钩子。
  3. 检查结尾。如果以通用积极表述结尾，删除或替换为具体内容。
  4. 统计“delves”数量。删除所有残留的AI黑名单词汇。
  5. 零破折号。搜索U+2014；用逗号、冒号或连字符替代。
  6. 句子长度审核。如果连续3句及以上句子长度相似，调整它们。
  7. “谁写的？”测试。如果有人读了这段文本，能想象出具体的作者吗？如果任何人（或任何机器）都能写出来，那么它需要更多语音风格。

Draft, self-audit, final (cheap quality pass, distinct from 
--iterate
)

草稿、自我审核、最终版本（低成本质量检查，与
--iterate
不同）

After the first rewrite, ask one question of your own draft: "What still makes this read as AI?" Answer honestly in two or three bullets, then do one corrective pass targeting exactly those. This metacognitive step is cheaper than a full 
--iterate
detect loop and catches the tells a checklist misses. It complements 
--iterate
, it does not replace it.
第一次改写后，问自己一个问题：“这段文本还有哪些地方读起来像AI？”用两三个要点诚实回答，然后针对这些要点做一次修正。这个元认知步骤比完整的
--iterate
检测循环更便宜，能捕捉到清单遗漏的特征。它是
--iterate
的补充，而非替代。

Scoring rubric (used when 
--score
is set)

评分标准（设置
--score
时使用）

Compute a 0-100 AI-tell density score. Lower is more human.
RangeVerdictWhat it means
0-20PristineReads like a specific human wrote it. No detector should flag it.
21-40Mostly humanOne or two minor tells, easy to clean.
41-60MixedHalf-AI half-human; partial editing likely.
61-80AI-leaningMultiple structural tells; detectors will probably catch it.
81-100Pure AI smellWholesale chatbot output with no editing.
Compute as: 
score = 4 × patterns_hit + 25 × (1 - burstiness_normalized) + 15 × (vocabulary_blacklist_ratio)
, clamped to 0-100. Show the score on the first line of output before the rewrite.
A model grading its own output in the same session tends to inflate the result. Treat 
--score
as a signal, not a verdict: the real gate is an independent pass or a human reader. For a computed, deterministic version of these metrics (burstiness, type-token ratio, sentence-length CoV, trigram repetition, Flesch-Kincaid) plus a CI quality-gate, see the optional 
cli/
tool in the repo. The skill core here needs none of it.
计算0-100分的AI特征密度评分。分数越低越接近人类风格。
范围结论含义
0-20完美读起来像特定人类撰写。任何检测器都不会标记它。
21-40基本人类风格存在一两个次要特征，易于清理。
41-60混合风格半AI半人类；可能经过部分编辑。
61-80偏向AI风格存在多种结构特征；检测器很可能会标记它。
81-100纯AI特征未经编辑的聊天机器人输出。
计算公式：
score = 4 × 检测到的模式数 + 25 × (1 - 标准化变化性) + 15 × (黑名单词汇占比)
，结果限制在0-100之间。在改写文本前的第一行显示评分。
同一会话中模型对自身输出评分往往会偏高。将
--score
视为信号，而非定论：真正的判断标准是独立审核或人类读者。如需计算确定性指标（变化性、类型标记比、句子长度变异系数、三元组重复率、Flesch-Kincaid可读性）并设置CI质量门限，请查看仓库中的可选
cli/
工具。本工具核心无需依赖这些。

Iterate handling (used when 
--iterate N
is set)

迭代处理（设置
--iterate N
时使用）

After producing the rewrite, re-run Step 2 (Detect) on the output. If patterns_hit > 0 AND iteration_count < N, recurse with the rewritten text as the new input. Stop when patterns_hit == 0 OR iteration_count == N. In the final change summary, note how many iterations ran (e.g., "Converged in 2 iterations").
Worked before/after examples for technical docs, blog posts, and LinkedIn are in 
references/patterns.md
.
生成改写文本后，对输出重新执行步骤2（检测）。如果检测到的模式数>0且迭代次数<N，则以改写文本作为新输入递归执行。当检测到的模式数==0或迭代次数==N时停止。在最终修改总结中注明迭代次数（例如：“2次迭代后收敛”）。
技术文档、博客文章和LinkedIn帖子的前后示例请查看
references/patterns.md
。

Always-On Mode

始终启用模式

To make an agent write clean by default, not only when you invoke 
/humanizer
, bake the core rules into its standing instructions. Ready copy-paste blocks for 
CLAUDE.md
, 
SOUL.md
, a system prompt, and ChatGPT custom instructions live in 
references/always-on-templates.md
. This keeps the skill on-demand while giving power users an always-on option.
Write like a human. Be weird, specific, inconsistent.
要让Agent默认输出干净的文本，而非仅在调用
/humanizer
时生效，请将核心规则融入其常驻指令。适用于
CLAUDE.md
、
SOUL.md
、系统提示和ChatGPT自定义指令的现成复制粘贴块位于
references/always-on-templates.md
。这既保留了工具的按需使用功能，又为高级用户提供了始终启用的选项。
像人类一样写作。保持怪异、具体、不一致。