de-ai-polish

De AI Polish

Quick Start

Detect AI-style expressions in articles:

# 全文扫描检测
/de-ai-polish detect @article.md

# 直接修改
/de-ai-polish fix @article.md

Core Detection Rules

General Principles: Detection Principles and Common Patterns

Core Principle: The following checklist and rules are areference framework, not an exhaustive list. AI writing has infinite variations, and regex matching will inevitably miss some. Detection must primarily rely on the Agent reading through the full text, with the checklist serving only as auxiliary reference.

AI-style expressions often appear asvariants, and will not exactly match the literal form of the checklist. Follow these principles:

1. Trace upward to the source: For expressions not listed, determine whether they belong to a variant of a known template family (contrastive sentence pattern → progressive structure → formulaic transition → neat parallelism → stacked degree adverbs → negative declaration → detached contrast → synonymous repetition)
2. Frequency red line: The following words/structures appearing more than2 timesin the full text are considered AI formulaic:越来越/越...越...、其实、往往/通常/很多时候、经常/常常、更重要的是/更麻烦的是、看起来/看来、真正、不只/不只是、更像/更像一个/更像一种、很容易、更容易
3. Consecutive repetition: The same intensifying word appearing consecutively in adjacent sentences is an AI trace (才...才...、不是...而是...、就是X，就是Y，就是Z）
4. Structural mechanicalness: Three or more parallel items guided by the same function word (容易X、容易Y、容易Z); sentences beginning with formulaic lead-in words (更麻烦的是、说到这里); fixed collocations of degree adverb + abstract noun (有很强的体感）
5. Synonymous repetition: Different words repeatedly expressing the same semantic field, such as窄/小/平/浅/局限、变化/改变/转变/变革, exceeding 3 times within a short distance is a sign of AI vocabulary poverty

Low AI Contamination Typology (v1.4.0)

The following 7 categories are used to summarize this skill's fragmented prohibited styles into stable sources of problems. The original 17-category checklist remains, responsible for specific hits; this typology is responsible for first determining "why this sentence seems AI-like," then deciding how to revise it.
Conclusion: This skill'snegative style cleanup itemscan be fully categorized into these 7 categories;personal-style-guide.md The positive personal style in is the rewriting goal and does not belong to contamination items.

Contamination Type	Source of Problem	Covers Rules in This Skill	Rewriting Priority
Posture contamination	The text is evaluating itself, flattering the reader, or posing as the author	Concluding posture sentences, sycophantic tone, self-statements, evaluative statements, excessive emphasis on fame, objective concession posture	Delete the posture layer, directly retain facts, judgments, or experience
Template contamination	Using fixed sentence patterns to organize viewpoints, replacing genuine judgment	Contrastive sentence patterns, formulaic connectives, AI transitions, lead-in verbal tics, comparative and assertive expressions	Rewrite the entire sentence, changing it to direct statements, causation, scenes, or parallelism
Semantic dilution	Vague words, degree adverbs, and big words dilute information	Absolutization / philosophizing / dramatization, stacked degree adverbs, vague frequency words, synonymous repetition, empty meaning elevation	Complete the object, condition, scope, and result; delete uninformative modifiers
Rhythm contamination	Sentence patterns, parallelism, short sentences, and line breaks create a machine rhythm	Neat parallelism, question-word parallelism, consecutive short paragraphs, long cascades 越...越..., sentence rhythm problems	Break up neat structures, merge fragmented sentences, restore natural paragraph rhythm
Concept contamination	Use jargon, metaphor, slogans, or technical terms to create a sense of sophistication	Slogan-like metaphors, internet jargon, misuse of quotation marks,格局/生态/闭环/模板/叙事 and other abstract words	Replace with ordinary words whenever possible; when a term must be used, explain its usage and boundaries in this text
Process leakage	Expose AI, research, collaboration, or writing processes to readers	Vague attribution, knowledge cutoff disclaimers, traces of collaborative communication,以上是/请告诉我, material-status expressions	Delete process traces; keep only confirmable facts and the author's judgment in the main text
Format contamination	Use formatting packaging to replace main text expression	Abuse of unordered lists, inline heading vertical lists, emojis, excessive bold, blockquote declarations, misuse of quotation marks	Restore to natural paragraphs, valid lists, or genuine quotations

How to use：

1. First assign each suspected problem a contamination type, then check the 17 detailed categories below to confirm the basis for the hit.
2. When the same sentence hits multiple categories, keep only one primary type and write the rest as co-occurring causes. For example, “**这是 X** appears in a blockquote” has template contamination as the primary type and format contamination as co-occurring.
3. For rewriting suggestions, give a complete sentence pattern according to the primary type; do not perform synonym substitution.

I. Absolutely prohibited categories (must be removed)

The following patterns should be removed under any circumstances; they are the most obvious AI traces:

1. Contrastive sentence patterns

Patterns that must be removed：

• 不是.*而是
• 并非.*而是
• 不在于.*在于
• 不在.*而在(Abbreviated form, such as “the problem is not the tool, but...”)
• 不再是.*而是
• 与.*是两回事
• 与.*不等同于
• 看似.*实则(Such as: seemingly discussing metaphor, but actually discussing the mode of production)
• 看起来.*其实(Such as: looks like X, but is actually Y)
• 可以.*但不能.*(Capability-limiting contrast, such as "cold water can wake a person up, but cannot see the future for them")
• 不在于.*也不在于(Double negation variant)
• 远不只是.*(Intensified variant of "not only")
• 这不是X。这是Y。(Separated contrast: splitting "not... but" into two sentences, essentially the same)
• 这不是X(Negative declaration: used as the beginning of a paragraph or a standalone assertion, such as "this is not imagination" or "this is not chasing trends")
• **这是X**(Positive declaration: used as a standalone bold line or blockquote, essentially the same as a negative declaration, only omitting the negative prefix, such as “This is moving one's observation position forward in advance」「This is establishing a position for oneself within a new trust mechanism」）
• 低水平X...高水平Y / 表层X...深层Y(Neat symmetrical contrast, often appearing in blockquotes)

Rewriting strategy: Choose “delete / merge / rewrite” according to context. Split into independent declarative sentences or progressive relations; avoid forced contrast for emphasis. When rewriting, do not apply the same structure ("essentially X" "the crux lies in X"); instead, flexibly choose alternatives such as statement, question and answer, cause and effect, or coordination according to the context.

Rewriting negative declarations and positive declarations: Simply delete the "this is not" or "this is" prefix and keep the core content. The bolded line format itself is enough to convey emphasis; there is no need to start with "this is".

Synonym cycling：

Detection pattern: using multiple synonyms within a short distance to refer to the same thing

Example：

• 主人公…主要角色…中心人物…英雄(using 4 words in the same paragraph to refer to the same person)
• 系统…平台…工具…解决方案(frequently switching words but referring to the same thing)

Revision: consistently use the most accurate word, and repeat the same word when the context is clear.

2. Formulaic connectives

Already prohibited in the writing rules：

• 首先、其次、此外、但是、因此、然而
• 一方面、另一方面
• 综上所述、总而言之

AI vocabulary bank：

• 深入探讨、深入分析(overuse)
• 强调、凸显、彰显(excessive repetition)
• 持久的、持久的证明
• 增强、增强其
• 培养、培养/促进、促进
• 获得、获得.*的认可
• 相互作用、复杂的相互作用
• 复杂/复杂性、错综复杂的
• 格局、整体格局、格局的演变(abuse of abstract nouns)
• 织锦、社会织锦(abuse of abstract nouns)
• 宝贵的、宝贵的经验
• 充满活力的、充满活力的社区

Superficial analysis expressions：

• 突出/强调/彰显…，确保…
• 反映/象征…，为…做出贡献
• 培养/促进…，涵盖…
• 展示…，体现…

3. Closing stance sentences

Expressions that must be detected：

• 方向已经明确
• 未来可期
• 拭目以待
• 这只是开始
• 让我们期待

Formulaic outlook：

Detection keywords：

• 尽管其…面临若干挑战…
• 尽管存在这些挑战、尽管面临挑战
• 挑战与遗产、机遇与挑战并存
• 未来展望、下一步计划
• 我们将继续关注、值得关注

Revision: delete empty outlooks, end with specific facts or a specific plan, or simply stop.

4. Colloquial and casual expressions

Vocabulary that must be detected：

• 稳、挺稳、很稳、靠谱
• 飘、摇摆、不稳
• 搞定、没问题
• 迟早得、肯定、必须得
• Overused affirmative words:肯定、绝对、一定
• 跑偏、很顺、挺顺and other overly colloquial state descriptions

Casual expression patterns：

• 写长了、做多了、搞定了and other overly casual expressions
• 迟早得拆、肯定会、必须得and other colloquial expressions
• 接下来要盯的东西、接下来要关注的and other colloquial expressions
• Use short colloquial wording instead of professional explanations

Revision direction: use more formal and more specific expressions ("will have to be split sooner or later" → "needs to be considered for splitting", "off track" → "deviating from expectations", "very smooth" → "smooth").

Copula avoidance：

• 作为/代表/标志着/充当 [一个]、拥有/设有/提供 [一个]、位于.*的中心、坐落于
• Revision: "Y as X" → "X is Y", "is provided with Z W" → "has Z W", "is situated at" → "is located at"

Flattering tone：

• 好问题！、您说得完全正确！、这就是一个很好的观点
• Revision: state the core content directly and delete flattering expressions.

5. Absolute, philosophical, and dramatic expressions

Expressions that must be detected：

• 并不是一个绝对命题
• 本质上、从根本上说、从本质上讲
• 必然、无疑、绝对、必须
• 真正的、核心的、关键的(overuse)
• 推到底、刺痛感
• 这词听着玄、道理其实很朴素
• 天然、天生、自然而然(absolute natural expression)
• 立刻、马上、立即(overly absolute time expression)
• 其实(overused, appears pretentious)
• 其实已经能看清、其实很简单and other expressions with "actually"
• 很现实(AI-style absolute judgment)

Overemphasis on significance：

• 作为.*的体现/证明/提醒、充当.*的证明
• 标志着.*关键时刻、见证了.*历史
• 是.*不可或缺的组成部分、是.*的重要体现
• 凸显/强调/彰显了其重要性/意义
• 反映了更广泛的、象征着其持续的/永恒的/持久的
• 为.*做出贡献、为.*奠定基础
• 关键转折点、不断演变的格局

False scope：

• 从.*到.*(check whether it is on a meaningful scale)
• For example, meaningless large scopes like "from the Big Bang to dark matter"
• Problem: forcibly using the "from X to Y" structure when X and Y are not on the same scale

Stacked degree adverb structure：

• 很+形容词/动词: very direct, very exploitative, very dependent, very troublesome
• 非常+形容词/动词: extremely dependent, extremely troublesome, extremely clear
• 有+很+的+名词: have a very direct feeling, have a very strong bodily sense, have a very deep feeling
• Revision: replace degree adverbs with more precise verbs, or directly state the specific state

Dramatic feeling words：

• shaken, shocking, touched, eye-catching, fascinating, inspiring
• Revision: switch to more plain expressions ("interesting", "different", "worth noting")

Revision: Replace absolutes with flexible expressions ("inevitably" → "often"/"mostly"/"possibly", "essentially" → "usually"/"to some extent"). Preserve a sense of openness; do not seal off conclusions excessively.

6. Neat parallelism and symmetrical structures

Structures that must be detected：

• Three or more "both...and...and..." parallel structures
• Forced antithetical parallel sentence patterns
• Neat symmetrical expression structures

Specific examples：

• "both explain 'who I am,' and explain 'what I am doing,' and explain 'how I do it'"
• "make the taste clear, make the bottom line clear, make the habits clear" (three "make...clear")
• Overly neat "on the one hand...on the other hand..."

Parallelism of question words：

Parallel items guided by three or more question words are a typical structure when AI lists scenarios:

• 能不能...能不能...能不能...(e.g., "can it be downloaded, can it be run directly, can a few fewer forms be filled out")
• 怎样...怎样...怎样...(e.g., "how to collect information, how to organize judgment, how to accumulate experience")
• 如何...如何...如何...(e.g., "how to identify problems, how to turn them into tools, how to translate work")
• 越...越...越...(e.g., "the more you understand tools, the easier it is to improve efficiency; the more you improve efficiency, the easier it is to output")

Rewrite: merge some items into noun phrases or verb-object phrases to break the uniform sentence pattern. For example, "how to collect information, how to organize judgment, how to accumulate experience" → "how to collect information and organize judgment, how to accumulate experience" (merge + reduce).

Rewrite: break the neat structure, change it to declarative sentences or mixed structures, and avoid three or more parallel items with the same structure.

7. AI transitions

Transitions that must be detected：

• 先把.*摆出来、摆上台面
• 不妨把.*拆成、不妨
• 这两个问题一问、答案就清晰了
• 这话对了一半、这话没错
• 我后来发现、我也想过
• 一个直接的原因、一个现实的原因、一个根本的原因
• 一个很现实的原因是、一个很直接的原因是
• 原因很简单、原因很直接
• 这也是为什么(conclusion-inversion transition; simply delete the prefix)
• 从一开始就(absolute time framework, such as "it was not aligned with the problem from the very beginning")

Traces of collaborative communication：

• 希望这对您有帮助、当然！、您说得完全正确！、请告诉我、以上是…
• Rewrite: delete all traces of collaborative communication and directly state the core content.

8. Self-statement

Expressions that must be detected：

• 我更愿意、我更倾向于
• 我给自己定、我设定了
• 我想强调、我想指出
• 我期待、我希望

Progressive self-statement：

• 我越来越觉得、我越来越认为、我越来越倾向于、我现在越来越
• "more and more" appears more than 2 times in the entire article
• Rewrite: directly state the viewpoint and delete the "I increasingly" prefix. Keep at most one instance of "I gradually realized."

9. Lead-in verbal tics

Expressions that must be detected：

• 表面上看
• 更关键的是、更值得写的是、更重要的是
• 更大的变化在于、更实质的变化是、更X的是/更X的在于(progressive declaration guided by the word "more")
• 从经验看、在实践中
• 这当然成立、毫无疑问
• 接下来会、接下来我们

Repeated frequency expressions：

• 很多时候(appears more than 2 times in the entire article),往往、通常(excessive repetition)
• Rewrite: keep the first occurrence; replace subsequent ones with "in most cases"/"often"/"generally speaking," or delete them directly.

Vague attribution：

• 行业报告显示、观察者指出、专家认为、多个来源(no specific source)
• Rewrite: delete the vague attribution and state it directly, or provide a specific source such as "according to the XX report of XX year."

Knowledge cutoff disclaimer：

• 截至 [日期]、根据我最后的训练更新、基于可用信息…、在现有资料中…
• Rewrite: directly state known facts and delete the disclaimer.

10. Slogan-like metaphors

Expressions that must be detected：

• signal, narrative, gear shifting, battlefield
• combination punch, closed loop, template, ecosystem
• move...into, drag...into...

11. Overly emotional

Vocabulary that must be detected：

• heartbreaking, hard, valuable, biggest
• repeatedly flipping

12. Excessive emphasis on fame

Phrases that must be detected：

• 被.*多家媒体报道、受到.*广泛关注
• 独立报道、地方/区域/国家媒体
• 由知名专家撰写、专家团队
• 活跃的社交媒体账号
• 拥有超过.*粉丝、关注者

Rewrite: Remove empty emphasis on popularity, or provide specific and meaningful content.

II. Flexible detection category (judge based on context)

The following patterns need to be judged based on the article's style, context, and target readers to determine whether revision is needed:

13. Unordered lists

Detection criteria: Whether the list is "piling up opinions" rather than "explaining steps"; if unordered lists appear more than 2-3 times in the entire article, they should be changed to paragraphs.

Acceptable scenarios: Operation steps, checklists, enumeration of core frameworks.

Inline heading vertical list：

• - **用户体验：** 用户体验通过新界面得到显著改善(mechanical list of bold headings + colons)
• Rewrite: Change to paragraph-style expression or normal list format.

Overuse of emojis and bold：

• 🚀 **启动阶段：**、💡 **关键洞察：**, overuse of **OKRs** and other bold formatting
• Rewrite: Remove emojis and excessive bold.

14. Quotation mark usage

Detect overuse：

• Quotation mark density is too high (quotation marks are used heavily throughout the entire article)
• Adding quotation marks to ordinary words (not terminology, not quotations)
• Using quotation marks to express "irony" or "emphasis"
• Half-width English quotation marks "..."
• Half-width single quotation marks '...'
• Overuse of 「」

Quotation mark misuse patterns specific to AI writing (must be detected)：

The following types of quotation marks are usually unnecessary and should be removed: descriptive phrases ("可调用的能力"), figurative descriptions ("Skill 的外壳"), colloquial expressions ("能不能一键跑起来"), action/state descriptions ("换一个 Skill"), attribute enumeration ("可组织、可迭代"), problem expressions ("Skill 很火"), value judgments ("工作标准"）。

Scenarios for correct use of quotation marks: Quoting original text, defining core terminology (first occurrence), avoiding ambiguity, distinguishing special concepts.

Detection criteria：

• If the meaning is still clear after removing the quotation marks, quotation marks are usually unnecessary
• If it is the descriptive use of ordinary words (nouns, verbs, adjectives), quotation marks are unnecessary
• If it is a metaphor, colloquial expression, or action description, quotation marks are unnecessary
• The density of quotation marks should be minimized throughout the entire article
• Prioritize retaining: definitions of core terminology, quotations from original text

15. Evaluative statements and overgeneralization

Patterns that must be detected：

• 把.*讲得很清楚、把.*说得很明白
• 更尖、更锐利、更深刻(evaluative adjectives modifying abstract concepts)
• 更硬的定义、更.*的.*(overly modified adjective structures)
• 被.*卡住、被.*阻挡(passive sentence pattern + overgeneralization)
• 一旦把视角从.*就更容易看清(formulaic perspective shift)
• 更具体一点、更准确地说(AI-style transitional phrases)

Rewrite: State the content directly, and remove evaluative modifiers and transitional phrases. "Stuck by two things" → "Two things are blocking the way," "a little more specific" → directly enter the specific content.

16. Comparative and assertive expressions

Patterns that must be detected：

• 它更像一种、更像一个.*
• 它其实是、这里其实、其实是第二类
• 天然会、天生就、自然而然就
• 会更清楚、会更.*(overused "will be more" pattern)
• 就.*了、就.*能(overly absolute assertions)

Negative assertive expressions：

• 事实并不是这样、事情并不是这样、并不是.*而是、其实并不是
• Rewrite: Remove the negative prefix and state the core content directly. "But this is not the case" or directly enter the facts.

Rewrite direction：

• "It is more like a kind of" → directly say what it is
• "It is actually" → directly state the fact
• "naturally" → use flexible expressions such as "usually", "tend to", etc.
• "will be clearer" → state the effect directly
• Overall strategy: remove the prefix and state the core content directly

17. Vague frequency words and objective concessive stance

Vague frequency words (false universality)：

• 经常、常常、常常会 — AI uses these words to create the illusion of a "common phenomenon"
• 很容易 + abstract judgment (e.g., "discussion easily becomes narrow") — a combination of vague frequency + vague consequence
• Rewriting direction: replace with specific frequency ("the last few times we talked about X" instead of "often talk about X"), or delete the frequency modifier directly; supplement vague consequences with specific references ("becomes narrow" → "limited to the product level")

Evaluative pronouncements (objective concessive stance)：

• 太绝对、太武断、话说得太死、说过头了、容易被反驳
• 这个判断太X、这种说法太Y — AI's formulaic concession of setting up a target first and then knocking it down
• Rewriting direction: replace evaluative pronouncements with a personal stance ("I don't see it that way" instead of "this judgment is too absolute"), or skip the concession directly and state your own judgment

Rewriting comparison：

Original text	Rewritten	Explanation
I often hear a judgment	The last few times I've run into a judgment	Specific frequency replaces vague words
Discussion easily becomes narrow	Discussion easily becomes limited to the product level	Add "narrow in what way"
This judgment is too absolute and easily refuted	Speaking of X, I don't see it that way	Personal stance replaces objective evaluation

Workflow

Startup gate: scenario determination

The intensity, tolerance, and protection scope of de-AI-flavoring vary by scenario. Before starting to read through, first determine which type of scenario this belongs to, and adjust the intensity and focus of the subsequent Steps 2-5 accordingly. If the user does not specify, handle it as "general" by default.

Scenario	Default intensity	Pollution categories to focus scanning on	Protected Spans width
Legal documents(judgments/opinions/contracts/pleadings/legal opinions)	Restrained	Stance pollution / process leakage / format pollution	Widest
Public account · public commentary	Aggressive	Template pollution / rhythm pollution / semantic dilution	Standard
Spoken · instant reply(chat/status/standup)	Minimal	Sycophancy / collaboration traces	Standard
General(unspecified)	Use public account intensity	All categories	Standard

Scenario determination constrains the following three things: ① rewriting intensity (restrained/aggressive/minimal); ② Protected Spans width (widest for legal documents, see Step 1); ③ which categories of pollution to focus scanning on (the remaining pollution categories are still handled by default, not skipped, just lower priority). In the legal document scenario, formal register, procedural terminology, and fixed formulaic expressions are retained by default and not treated as AI flavor.

Step 1: Reading comprehension

Must read the entire text completelyComprehend the core viewpoints, article style, and contextual background. Do not detect merely through regular expression/grep matching—regular expressions can only cover the exact forms of known patterns and will miss a large number of semantically equivalent variants with different forms.

While reading through, simultaneously delineate Protected Spans (items that must not be modified): mark the scope that must be preserved as-is and must not be touched in subsequent Steps 2-4. Default coverage:

• Numbers and citations of statutes, judicial interpretations, and regulations (e.g., Article 143 of the Civil Code, (2023) Zui Gao Fa Min Shen No. 12)
• Full names of parties, institutions, law firms, and companies
• Procedural terms (jurisdiction, objection, appeal period, burden of proof, court summons, etc.)
• Direct quotations, citations, contract clause numbers
• Numerical values, dates, proportions, amounts, time limits
• URLs, document numbers, case numbers

Determination principle: prefer retaining more rather than mistakenly modifying. If a Protected Span superficially appears AI-flavored (e.g., a company's full name containing "Co., Ltd." is misjudged as promotional), still retain it as-is and only note in remarks that the user should review it. When the scenario is "legal documents," the scope of Protected Spans shall be the broadest.

Step 2: Identify AI-style expressions

Identify AI-style expressions by having the Agent read the full text. The core detection rules in the preceding text (General Provisions + Sections I-II) are areference framework, not an exhaustive list. The Agent should:

Before entering this step, first declare the scan scope: which sections were covered this time (the specific subsections of the General Provisions, Section I, and Section II), and also explicitly list the sections not covered this time. Let the user see at a glance where the boundaries of detection lie.

The identification process must not touch the Protected Spans delineated in Step 1: even if suspected AI patterns appear inside a Protected Span (contrastive sentence structures, promotional language, etc.), do not include them in findings or rewrite them; at most, note in the delivery remarks that the user should review them.

1. Type classification: first assign a primary type to suspected issues according to the "low-AI-contamination typology": stance contamination / template contamination / semantic dilution / rhythm contamination / concept contamination / process leakage / format contamination. Do not start with word-list substitution.
2. Vocabulary-level scan: read through the full text and identify all suspected AI-style words and phrases—whether or not they appear in the list.Focus the scan on the following hotspot areas：
   • Section headings: patterns such as "not X" and "only Y" in headings are often overlooked
   • Bold lines and block quotes: AI tends to place the most polished contrastive sentences and declarative assertions in **粗体** or > 引用 Middle
   • First and last sentences of paragraphs: AI transition phrases and summary sentences most often appear in these positions
3. Structure-level scan: after the vocabulary scando not declare completion. Independently conduct a structure scan specifically to identify AI-prone sentence skeletons.Mandatorily scanthe following structural patterns:
   • Parallel structures guided by 3+ consecutive identical function words：是 X。是 Y。是 Z。 / 生成 X，生成 Y，生成 Z / 先把 X。先把 Y。先把 Z。 / 我们要的是 X，是 Y，是 Z
   • Parallel structures guided by 3+ consecutive interrogative words：能不能 X？能不能 Y？能不能 Z？ / 怎样 X，怎样 Y，怎样 Z / 如何 X，如何 Y
   • 3+ consecutive passive parallel structures: by X, by Y, by Z / can be reviewed, adopted, transmitted
   • 3+ consecutive short paragraphs (strings of single-sentence paragraphs): each paragraph stands alone, the content structure is similar, and the rhythm is uniformly neat
   • "X is X is X" nesting: the same verb appears multiple times to introduce parallel structures in the same form
4. then check against the list: supplement terms whose frequency exceeds the limit that may have been missed during a full read-through (e.g., 越来越 真正 不只 and other words appearing more than 2 times)
5. trace upward to the source: each rule in the list represents a type of pattern. When you identify wording not listed, determine whether it belongs to a variant of a known pattern family (contrastive sentence pattern → progressive structure → formulaic transition → neat parallelism → stacked degree adverbs → negative declaration → separated contrast → synonymous loop)

Prohibited practices: only using grep/regular expressions to match keywords in the list, then claiming the check is complete. Regex matching can only serve as an auxiliary tool for frequency counting, and cannot replace reading comprehension.You cannot declare completion after scanning the lexical layer—you must go through the structural layer scan again.

Why the structural layer must be gone through independently: lexical-layer flags (such as "not X but Y", "actually", "truly") are easy to anchor on, and an agent can easily put all its attention on word replacement. Structural-layer flags (consecutive question marks, consecutive short paragraphs, parallel structures guided by consecutive function words) will not appear in word-level grep results—you must scan them separately by position, by rhythm, and by the number of parallel items. No matter how thoroughly you scan the words, if you do not scan the structure, it is as if you did not scan at all.

Step 3: Determine whether to revise

Not all matches need to be revised. Criteria: whether it affects naturalness of expression, whether it fits the article's style, and whether it serves a rhetorical purpose.

Step 4: Rephrase the sentence

Core principle: rephrase the entire sentence; do not merely delete vocabulary or replace synonyms.

Incorrect (×):"它更像一种可维护的模块" → "这是一种可维护的模块"(only words deleted, structure unchanged) Correct (✓):"它更像一种可维护的模块" → "这种设计便于后续维护"(restructure the sentence)

Synonym replacement is ineffective rewriting: replacing "narrow" with "limited", replacing "small" with "finite", and replacing "this is not" with "that is not" leaves the sentence pattern and tone unchanged, and readers can still sense the AI flavor. On the premise of understanding the core meaning of the original sentence, you must switch to an entirely new expression structure.

Original text	Ineffective rewrite (×)	Effective rewrite (✓)	Explanation
the vision was small from the start	the vision was narrow from the start	the discussion just didn't grasp the key point	changed the semantic field, changed the sentence pattern
this is not the AI I understand	that is not what I focus on	the angle I focus on is a bit different	remove the negative declaration and replace it with a personal stance
the vision becomes narrow	the vision becomes limited	but seeing only these is still not enough	completely switch to another expression
the more you X, the more you Y; the more you Y, the more you Z	the more you X, the more you Y; the more you Y, the more you Z	Understand the tool, and efficiency will rise first. With surplus capacity, output will naturally increase	split long cascades into short sentences

Step 5: Inject personal style (optional)

After removing AI-like expressions, personal style features may optionally be injected.Target ratio: technical accuracy 60% / personalization 40%This step is optional; purely technical/legal documents may skip it if no personalization is needed.

Two paths (choose based on whether an author sample is available)：

• Author sample available → Voice Calibration: If the user provides writing samples of the target author (≥500 characters recommended), follow the "Voice Calibration process" in reference/personal-style-guide.md to extract a voice profile (seven dimensions: sentence length distribution / word choice level / paragraph openings / punctuation habits / catchphrases and transitions / opinion density / tone tendency), and match it during rewriting.Samples take priority over the default features below.
• No sample → default voice: Inject the positive features below; for details see the reference/personal-style-guide.md"Default voice" section.

Sample boundaries: Only use author samples provided by the user or confirmed as usable for this task. Voice Calibration only learns expressive features; it does not impersonate the author's identity, copy original sentences or highly distinctive phrases from the samples, or write sample facts into the target text. If the samples contain sensitive information, abstract it only into a voice profile and do not repeat it in the delivered text.

Default positive features(used when no sample is available):

1. Authenticity of the opening: Start with personal feelings/concrete scenes/plain judgments, not "This article will explore…"
2. Flexible expression: Keep "possibly" "often" "to some extent" etc.; do not close off conclusions.
3. Measured colloquialism: "to be honest" "maybe" are OK; "get it done" "go off track" are not.
4. Self-deprecating humor: Unexpected twists and counterintuitive points are OK; boastful self-praise is not.
5. Natural contrast: "Of course… but…" is OK; "not… but rather…" is not.
6. Sentence rhythm: Mix long and short sentences; do not force neatness.

Reference files(read for detailed guidance):

• reference/personal-style-guide.md: Voice Calibration process + default voice features
• reference/expression-transformations.md: Expression transformation comparison table
• reference/sentence-rhythm-guide.md: Sentence rhythm guide
• reference/quality-scoring.md: Quality scoring system

Step 6: Punctuation correction (enforced by code)

Must be performed as the final step. After AI modifies text, English punctuation is often left behind; relying purely on LLM reminders to correct it is unreliable, so use a script to enforce conversion.

python3 scripts/fix_punctuation.py <文件路径>

The script automatically handles:

• English quotation marks " → Chinese quotation marks ""(alternating state machine)
• In Chinese contexts, , → ，、; → ；、? → ？、! → ！
• In Chinese contexts, : → ：、() → （）
• After Chinese characters, . → 。

The script automatically skips the following areas and does not convert them:

• YAML front matter（--- metadata between)
• Markdown image and link syntax (![]()、[]()）
• Code blocks and inline code (```、`）
• URL addresses

Step 7: Pre-delivery scoring gate

After Steps 4-6 are complete and before delivery to the user, score the final text according to the 5 dimensions in reference/quality-scoring.md (naturalness / rhythm / professionalism / personality / conciseness, total score 10):

• Total score < 7.0 → Return to Step 4 to rephrase the problematic sentences, then score again.
• Single-dimension hard failure: Naturalness < 1.5 or personality = 0 → send back for rework even if the total score meets the threshold.
• Scenario adjustments: In legal document scenarios, if professionalism < 1.5, rework is required; in colloquial scenarios, the individuality / conciseness thresholds may be relaxed.
• Voice profile gate when author samples are available: If the final text clearly deviates from the main stable dimensions of the extracted profile, or copies original sentences / highly distinctive phrases from the samples, introduces facts from the samples, or reproduces counterexamples marked by the profile, it must return to Step 4/5 for revision.

When scoring, the two auxiliary perspectives of "directness" and "trusting the reader" may be used to locate the reason for point deductions (see reference/quality-scoring.mdthe "Use as a delivery gate" section); they are not scored separately. The scoring process and results are by default not written into the delivered main text, and are provided only in delivery notes or when requested by the user.