humanize

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Humanize: AI Pattern Detection and Removal

Humanize：AI写作模式检测与移除

Remove AI-generated writing patterns from text. Produce natural, human-sounding output that preserves meaning.

This is not a generic rewriter. It targets specific, documented AI-writing patterns catalogued by Wikipedia's WikiProject AI Cleanup from thousands of observed instances.

移除文本中的AI生成写作模式，生成保留原意、自然流畅的类人写作输出。

这并非通用改写工具，它针对维基百科AI清理项目从数千个实例中整理出的特定、有记录的AI写作模式。

Workflow

工作流程

Five phases. Each phase has a clear input, transformation, and output. Do not skip phases.

分为五个阶段，每个阶段都有明确的输入、转换和输出，不得跳过任何阶段。

Phase 1: Detection Scan

第一阶段：检测扫描

Read the input text. Load

references/detection-patterns.md

. Scan for two categories of signals:

A. Lexical patterns (the 24 catalogued AI-writing patterns):

Category	Patterns	Priority
Content inflation	Significance puffing, notability claims, superficial -ing analyses, promotional language, vague attributions, formulaic challenges sections	HIGH — loudest AI tells
Vocabulary	AI-frequency words, copula avoidance, filler phrases, excessive hedging	HIGH — statistically detectable
Structure	Rule of three, negative parallelisms, elegant variation, false ranges, inline-header lists	MEDIUM — structural fingerprints
Style	Em dash overuse, boldface overuse, title case headings, emoji decoration, curly quotes	MEDIUM — formatting tells
Communication	Chatbot artifacts, knowledge-cutoff disclaimers, sycophantic tone, generic conclusions	LOW — obvious, usually caught by author

B. Statistical regularity signals (see

references/statistical-signals.md

Signal	What to look for
Sentence length uniformity	Sentences clustering within a narrow word-count range
Low clause density variation	Every sentence has the same number of clauses
Flat information density	Every sentence carries roughly the same amount of detail
High-frequency phrase templates	Stock collocations and common bigrams/trigrams dominating the text
Excessive transition markers	Formal connectives appearing more than 8 per 1,000 words
Structural symmetry	Paragraphs and sentences following balanced, mirror-like patterns
Uniform inter-sentence cohesion	Every sentence tightly follows the previous with no topic shifts or digressions
Generic function word usage	Connectors and prepositions used in textbook-standard distribution with no personal tendencies

Output a detection report using the detection report template (see Output Format).

Instance severity rating:

Severity	Criteria
HIGH	3+ patterns co-occurring in a single paragraph, or any paragraph saturated with AI vocabulary (5+ signal words)
MEDIUM	1-2 patterns in a paragraph, or a statistical signal present across 3+ consecutive sentences
LOW	Isolated single instance of any pattern, or a borderline statistical signal

阅读输入文本，加载

references/detection-patterns.md

文件，扫描两类信号：

A. 词汇模式（24种已归类的AI写作模式）：

类别	模式	优先级
内容注水	夸大重要性、显著性声明、表面化的-ing形式分析、宣传性语言、模糊归因、公式化的挑战部分	高优先级——最明显的AI写作特征
词汇	AI高频词、避免使用系动词、填充性短语、过度模糊表述	高优先级——可通过统计检测
结构	三段式结构、否定平行句式、刻意换词、虚假范围、行内标题列表	中优先级——结构特征
风格	过度使用破折号、过度使用粗体、标题大小写、表情符号装饰、弯引号	中优先级——格式特征
沟通风格	聊天机器人痕迹、知识截止日期声明、谄媚语气、通用结论	低优先级——明显易见，通常已被作者发现

B. 统计规律性信号（详见
references/statistical-signals.md
）：

信号	检测要点
句子长度一致性	句子字数集中在狭窄范围内
从句密度变化小	每个句子的从句数量相同
信息密度均匀	每个句子承载的信息量大致相同
高频短语模板化	固定搭配和常见二元/三元短语在文本中占主导
过渡标记过度使用	正式连接词使用频率超过每1000词8次
结构对称性	段落和句子遵循平衡的镜像结构
句间衔接过于紧密	每个句子都严格承接上一句，无主题转换或偏离
功能词使用通用化	连接词和介词完全按照教科书标准使用，无个人风格

使用检测报告模板输出检测结果（详见输出格式）。

实例严重程度评级：

严重程度	判定标准
高	单个段落中同时出现3种及以上模式，或任意段落充斥AI高频词汇（5个及以上特征词）
中	单个段落中出现1-2种模式，或连续3句及以上存在统计信号
低	仅出现孤立的单个模式实例，或存在临界统计信号

Phase 2: Structural Rewrite

第二阶段：结构改写

Transform document structure to break AI-typical organization:

Convert uniform paragraph lengths to varied blocks
Merge or split sentences to break rhythmic uniformity
Reorder clauses where meaning permits
Convert formulaic list structures to narrative where appropriate
Remove tripartite constructions unless the content genuinely has three parts

Do not change factual content. Do not add information. Do not remove cited sources, data, or technical terms.

调整文档结构，打破AI典型的组织方式：

将均匀的段落长度改为长短不一的块
合并或拆分句子，打破节奏一致性
在不改变原意的前提下重新排列从句
酌情将公式化列表结构转换为叙述性内容
除非内容确实包含三部分，否则移除三段式结构

不得修改事实内容，不得添加信息，不得删除引用来源、数据或专业术语。

Phase 3: Vocabulary and Style Pass

第三阶段：词汇与风格优化

Apply pattern-specific rewrites from the detection report:

Replace AI-frequency vocabulary with natural alternatives
Restore simple copulas (is/are/has) where the text uses elaborate substitutes
Remove filler phrases and excessive hedging
Cut promotional language and significance inflation
Replace vague attributions with specific ones (or remove if no source exists)

Load the appropriate style profile from

references/style-guide.md

based on the target domain. Apply domain-specific voice calibration.

根据检测报告应用针对性改写：

用自然替代词替换AI高频词汇
在文本使用复杂替代表达的地方恢复简单系动词（is/are/has等）
删除填充性短语和过度模糊表述
删减宣传性语言和夸大的重要性表述
将模糊归因替换为具体表述（若无来源则移除）

根据目标领域从

references/style-guide.md

加载对应的风格配置文件，应用领域特定的语气校准。

Phase 4: Entropy and Variation

第四阶段：增加随机性与多样性

Human writing has burstiness — irregular rhythm, varied sentence lengths, uneven information density. AI text is statistically smooth. This phase breaks that smoothness.

Load

references/statistical-signals.md

for target ranges. Apply:

Sentence length variance: mix short declarative with longer explanatory. Target visible variance across any 5-sentence window.
Clause density variation: alternate simple sentences (one clause) with compound/complex (2-3 clauses). Do not settle on a uniform clause count.
Information density variation: let some sentences carry heavy detail while others are light — a summary statement, a reaction, a pivot. Uniform density reads as generated.
Phrase template breaking: replace stock collocations with specific phrasings. "Play a role in" -> name the specific action. "In terms of" -> delete or restructure.
Inter-sentence cohesion variation: not every sentence should tightly follow the previous. Allow small topic expansions, brief asides, or contextual jumps that a thinking human would make.
Function word personalization: vary connector usage. Use "but" in one place, "still" in another, nothing in a third. Do not default to the same conjunction pattern throughout.
Paragraph length variance: mix single-sentence paragraphs with 4-5 sentence blocks.
Controlled imperfection: fragments at impact positions, parenthetical asides, concessive turns. Sparingly — seasoning, not structure.

人类写作具有突发性——节奏不规则、句子长度多变、信息密度不均，而AI文本在统计上过于平滑。本阶段旨在打破这种平滑性。

加载

references/statistical-signals.md

获取目标范围，应用以下调整：

句子长度变化：混合使用简短的陈述句和较长的解释句，确保任意5句窗口内可见明显的长度差异。
从句密度变化：交替使用简单句（1个从句）和并列/复合句（2-3个从句），避免从句数量统一。
信息密度变化：让部分句子承载大量细节，部分句子简洁概括——可以是总结性语句、反应或话题转换，均匀的信息密度会显得像AI生成。
打破短语模板：用具体表述替换固定搭配，例如将“Play a role in”替换为具体的动作描述，将“In terms of”删除或重构。
句间衔接变化：并非每个句子都必须严格承接上一句，允许出现符合人类思考逻辑的小范围话题拓展、简短题外话或上下文跳转。
功能词个性化：多样化连接词的使用，比如此处用“but”，彼处用“still”，另一处不使用连接词，避免全程使用相同的连词模式。
段落长度变化：混合使用单句段落和4-5句的长段落。
可控的不完美：在关键位置使用碎片句、插入语或让步转折，点到为止——作为调味而非结构主体。

Phase 5: Validation and Output

第五阶段：验证与输出

Two checks before delivering:

Semantic check: Compare rewrite against original. Every factual claim, data point, argument, and technical term in the original must be present in the rewrite. If anything was lost, restore it.

Self-audit: Ask internally: "What still sounds AI-generated about this text?" If residual patterns remain, fix them. One pass only — do not loop indefinitely.

Output the final text followed by a brief changes summary.

交付前需完成两项检查：

语义检查：将改写后的文本与原文对比，原文中的每一个事实主张、数据点、论点和专业术语都必须在改写版本中保留。若有内容丢失，需恢复。

自我审核：内部自问：“这段文字还有哪些地方听起来像AI生成的？”若仍存在残留模式，需修正。仅执行一次审核——不得无限循环。

输出最终文本及简短的修改总结。

Output Format

输出格式

Full Rewrite / Targeted Fix / Style Shift

完整改写 / 针对性修复 / 风格转换

[Humanized text]

---
Changes: [2-4 bullet summary of what was changed and why]
Patterns detected: [list of pattern numbers/names found]
Domain: [detected or specified domain]

For short texts (under 100 words), skip the changes summary unless the user requests it.

[人性化处理后的文本]

---
修改说明：[2-4条要点，说明修改内容及原因]
检测到的模式：[已发现的模式编号/名称列表]
领域：[检测或指定的领域]

对于短文本（不足100词），除非用户要求，否则可跳过修改总结。

Detection Only

仅检测

undefined

undefined

Detection Report

检测报告

Domain: [detected or specified] Overall severity: [HIGH / MEDIUM / LOW] Patterns found: [count]

领域： [检测或指定的领域] 整体严重程度： [高 / 中 / 低] 发现的模式数量： [数量]

Findings

检测结果

Location	Pattern	Severity	Evidence
Para 1	#7 AI vocabulary	HIGH	"delve", "intricate", "pivotal" in same sentence
Para 2	#8 Copula avoidance	MEDIUM	"serves as" instead of "is"
Para 1-4	Sentence length uniformity	MEDIUM	All sentences 18-22 words, SD < 3
...	...	...	...

位置	模式	严重程度	证据
第1段	#7 AI高频词汇	高	同一句中出现“delve”“intricate”“pivotal”
第2段	#8 避免使用系动词	中	使用“serves as”而非“is”
第1-4段	句子长度一致性	中	所有句子长度为18-22词，标准差<3
...	...	...	...

Statistical Signals

统计信号

Signal	Status	Detail
Sentence length variance	FLAG	SD ~3 words (human typical: 7-15)
Transition frequency	OK	5 per 1,000 words
...	...	...

信号	状态	详情
句子长度变化	标记异常	标准差约3词（人类写作典型值：7-15）
过渡词频率	正常	每1000词5次
...	...	...

Summary

总结

[1-2 sentences: overall assessment and highest-priority patterns to fix first]

undefined

[1-2句话：整体评估及最需优先修复的模式]

undefined

Reference Files

参考文件

File	Purpose	Load When
`references/detection-patterns.md`	24 AI-writing patterns with examples	Always (Phase 1)
`references/statistical-signals.md`	12 statistical regularity signals with target ranges	Phase 1 (scan) and Phase 4 (targets)
`references/style-guide.md`	Domain-specific voice profiles and calibration rules	Phase 3 (match to domain)
`references/transformation-rules.md`	Structural rewrite strategies and entropy techniques	Phase 2 and Phase 4
`examples/academic.md`	Before/after pairs for academic writing	When domain is academic
`examples/blog.md`	Before/after pairs for blog/casual writing	When domain is blog or social
`examples/professional.md`	Before/after pairs for professional/business writing	When domain is professional

文件	用途	加载时机
`references/detection-patterns.md`	24种AI写作模式及示例	始终加载（第一阶段）
`references/statistical-signals.md`	12种统计规律性信号及目标范围	第一阶段（扫描）和第四阶段（调整）
`references/style-guide.md`	领域特定语气配置及校准规则	第三阶段（匹配领域）
`references/transformation-rules.md`	结构改写策略及随机性调整技巧	第二阶段和第四阶段
`examples/academic.md`	学术写作的前后对比示例	领域为学术时
`examples/blog.md`	博客/休闲写作的前后对比示例	领域为博客或社交媒体时
`examples/professional.md`	专业/商务写作的前后对比示例	领域为专业文书时

Domain Detection

领域检测

If the user does not specify a domain, infer from:

Vocabulary density and jargon type
Citation patterns
Sentence complexity
Register (formal/informal markers)

Default to professional if ambiguous.

Supported domains:

academic

technical

blog

social

professional

marketing

若用户未指定领域，可从以下方面推断：

词汇密度和术语类型
引用模式
句子复杂度
语域（正式/非正式标记）

若存在歧义，默认使用专业文书领域。

支持的领域：

academic

（学术）、

technical

（技术）、

blog

（博客）、

social

（社交媒体）、

professional

（专业文书）、

marketing

（营销）

Behavioral Constraints

行为约束

Never fabricate. Do not add facts, citations, quotes, statistics, or claims not in the original.
Never remove data. Numbers, dates, names, URLs, and cited sources must survive the rewrite.
Preserve argument structure. If the original makes points A, B, C in that order with that logic, the rewrite must preserve the logical flow.
Do not over-humanize. Some text is meant to be neutral and informational. A technical specification does not need personality. Match the appropriate register.
Respect code blocks and structured data. Do not humanize code, tables, JSON, YAML, or any structured/machine-readable content. Pass these through unchanged.
One pass through the pipeline. Do not run the 5-phase pipeline recursively. If the output still has tells after Phase 5, note them in the changes summary rather than looping.

不得编造内容：不得添加原文中没有的事实、引用、引用语、统计数据或主张。
不得删除数据：数字、日期、姓名、URL及引用来源必须在改写后保留。
保留论证结构：若原文按A、B、C的顺序及逻辑阐述观点，改写版本必须保留该逻辑流程。
避免过度人性化：部分文本旨在保持中立和信息性，技术规范无需添加个性，需匹配合适的语域。
尊重代码块和结构化数据：不得人性化处理代码、表格、JSON、YAML或任何结构化/机器可读内容，直接原样保留。
仅执行一次流程：不得递归运行五阶段流程。若第五阶段后仍存在AI痕迹，需在修改总结中注明，而非循环执行。

Scope Modes

模式范围

Mode	Trigger	Behavior
Full rewrite	"humanize this", "rewrite naturally"	Run all 5 phases
Detection only	"check for AI patterns", "does this sound AI"	Run Phase 1 only, output detection report
Targeted fix	"fix the AI-sounding parts", "just clean up the obvious stuff"	Run Phase 1, then apply fixes only to HIGH-priority patterns
Style shift	"make this more casual/academic/professional"	Run Phases 3-4 with specified domain profile

模式	触发条件	行为
完整改写	“人性化处理这段文本”“自然改写”	运行全部5个阶段
仅检测	“检查AI写作模式”“这段文字像AI写的吗”	仅运行第一阶段，输出检测报告
针对性修复	“修复像AI写的部分”“只清理明显的AI痕迹”	运行第一阶段，仅针对高优先级模式应用修复
风格转换	“让这段更随意/学术/专业”	运行第三至第四阶段，使用指定领域配置

Error Handling

错误处理

Problem	Cause	Resolution
Input under 20 words	Insufficient signal for pattern detection	Report: "Text too short for reliable pattern detection." Apply vocabulary fixes only (Phase 3) if obvious patterns are present. Skip statistical signal analysis.
Input is entirely code/structured data	No prose to humanize	Report: "Input is structured data — no humanization applicable." Return input unchanged.
Mixed human + AI text	Partial AI generation or human-edited AI output	Run Phase 1 on full text. Flag only paragraphs/sections with detected patterns. Apply Phases 2-4 selectively to flagged sections. Leave clean sections untouched.
Domain ambiguous after detection	Input mixes registers (e.g., academic citations in a blog post)	Default to professional. Note the ambiguity in the output: "Domain defaulted to professional — specify if another profile is preferred."
Semantic drift detected in Phase 5	Rewrite altered meaning during structural/vocabulary changes	Restore the drifted factual claim from the original. Do not re-run the full pipeline. Note the restoration in the changes summary.
Input contains fabricated citations	Original text has hallucinated sources	Not detectable — this skill humanizes style, not factual accuracy. Pass through unchanged. Note in limitations if the user asks about accuracy.
All patterns are LOW severity	Text is mostly human-written with minor tells	In targeted fix mode, report findings but recommend no changes. In full rewrite mode, apply light-touch fixes only — do not over-edit clean text.

问题	原因	解决方式
输入文本不足20词	模式检测信号不足	报告：“文本过短，无法进行可靠的模式检测。”若存在明显模式，仅应用词汇修复（第三阶段），跳过统计信号分析。
输入完全为代码/结构化数据	无散文内容可进行人性化处理	报告：“输入为结构化数据，无需进行人性化处理。”原样返回输入内容。
混合人类与AI文本	部分为AI生成或经人工编辑的AI输出	对全文运行第一阶段，仅标记存在检测模式的段落/章节，选择性对标记章节应用第二至第四阶段，未标记的干净章节保持不变。
检测后领域仍不明确	输入混合多种语域（如博客文章中包含学术引用）	默认使用专业文书领域，在输出中注明歧义：“领域默认为专业文书，若需其他配置请指定。”
第五阶段检测到语义偏差	结构/词汇修改过程中改变了原意	从原文恢复偏差的事实主张，无需重新运行完整流程，在修改总结中注明恢复操作。
输入包含编造的引用	原文存在幻觉来源	无法检测——本技能仅人性化处理风格，不验证事实准确性，原样保留，若用户询问准确性需说明局限性。
所有模式均为低严重程度	文本基本为人类写作，仅存在少量AI痕迹	在针对性修复模式下，报告检测结果但建议无需修改；在完整改写模式下，仅应用轻度修复——不得过度编辑干净文本。

Integration Point

集成点

Other writing skills can import

references/detection-patterns.md

as a pattern library for their own anti-pattern sweeps. The detection patterns are the shared asset; the pipeline is this skill's domain.

其他写作技能可导入

references/detection-patterns.md

作为模式库，用于自身的反模式扫描。检测模式为共享资产，工作流为本技能的专属领域。

Limitations

局限性

Cannot verify factual accuracy of the original text. Garbage in, humanized garbage out.
Effectiveness depends on input length. Very short texts (under 20 words) have insufficient signal for pattern detection.
Style profiles are guidelines, not voice cloning. The output will sound natural but will not match a specific author's voice without additional calibration.
Does not interact with external AI-detection APIs. Assessment is heuristic, not benchmark-verified.

无法验证原文的事实准确性，输入垃圾内容，输出的人性化内容也会是垃圾。
效果取决于输入长度，极短文本（不足20词）的模式检测信号不足。
风格配置为指导方针，并非语音克隆，输出会自然流畅，但若无额外校准，无法匹配特定作者的语气。
不与外部AI检测API交互，评估基于启发式规则，未经过基准验证。