article-rewriter

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Article Rewriter

面向“翻译 + 改写 + 发布”的文章处理技能。

它不是纯翻译器，也不是只做措辞润色的改稿器。遇到外文原文、混合语言原文、需要重组结构、需要适配公众号/知乎/Newsletter、或用户要求成稿级输出时，都要把“理解原文”和“生产成稿”拆开做。

三模式：

```
quick
```
：轻量翻译或轻量改写，不主动外查，尽量少留中间文件
```
standard
```
：分析原文，必要时补充少量资料，走一轮显式草稿和批评修订
```
publish
```
：分析、核验、重组、批评、修订、抛光、表述优化、质检，产出可直接发布版本

默认模式：

standard

Article processing skill oriented towards "translation + rewriting + publication".

It is not a pure translator, nor is it a revision tool that only polishes wording. When dealing with foreign language originals, mixed-language originals, content that requires structural reorganization, content that needs adaptation to WeChat Official Account/Zhihu/Newsletter, or when users request draft-level output, "understanding the original text" and "producing the final draft" must be handled separately.

Three Modes:

```
quick
```
: Lightweight translation or lightweight rewriting, no active external research, minimize intermediate files
```
standard
```
: Analyze the original text, supplement small amounts of information if necessary, go through explicit draft and critical revision rounds
```
publish
```
: Analysis, verification, reorganization, critique, revision, polishing, expression optimization, quality inspection, produce a directly publishable version

Default Mode:

standard

输入类型

Input Types

```
inline text
```
：用户直接粘贴正文
```
file path
```
：读取本地文件
```
URL
```
：抓取正文后再处理(优先 WebReader，失败尝试用
```
https://r.jina.ai/{url}
```
读取，仍失败则尝试使用
```
https://defuddle.md/{url}
```
，如果都失败则明确说明无法获取内容)

```
inline text
```
: Users directly paste the main text
```
file path
```
: Read local files
```
URL
```
: Capture the main text before processing (prioritize WebReader; if failed, try reading with
```
https://r.jina.ai/{url}
```
; if still failed, try using
```
https://defuddle.md/{url}
```
; if all fail, clearly state that content cannot be obtained)

默认值

Default Values

设置	默认值	说明
输出语言	`zh-CN`	最终统一输出中文
模式	`standard`	用户未指定时使用
平台	`generic`	若用户提到公众号/知乎/Newsletter，则改用对应预设
标题数	`5`	用户未指定时提供 5 个标题候选
调研范围	`2-4` 个主题	`publish` 模式可扩展到 `4-6` 个
输出目录	`rewrites/{slug}-{timestamp}/`	所有中间文件和最终稿统一放在这里
长文阈值	`> 2500` 英文词或 `> 4000` 中文字	超过后优先采用分块写作与统一审校

Setting	Default Value	Description
Output Language	`zh-CN`	Unified final output in Chinese
Mode	`standard`	Used when not specified by users
Platform	`generic`	Switch to corresponding preset if users mention WeChat Official Account/Zhihu/Newsletter
Number of Titles	`5`	Provide 5 title candidates when not specified by users
Research Scope	`2-4` topics	Extend to `4-6` in `publish` mode
Output Directory	`rewrites/{slug}-{timestamp}/`	All intermediate files and final drafts are stored here uniformly
Long Article Threshold	`> 2500` English words or `> 4000` Chinese characters	Prioritize chunked writing and unified review when exceeded

模式

Modes

模式	触发词	核心步骤	适用场景
Quick	“简单润色”“快改一下”“不要查资料”	约束确认 → 术语统一（如需要）→ 直接成稿	短文、轻量润色、用户明确不要外查
Standard	默认	分析 → 选择性调研 → Prompt 快照 → 草稿 → 批评 → 修订 → 表述优化 → 成稿	大多数文章改写需求，尤其是“翻译并改写”
Publish	“深度改写”“公众号成稿”“适合发布”“补充资料并重写”	分析 → 核验 → 重组 → Prompt 快照 → 草稿 → 批评 → 修订 → 抛光 → 表述优化 → 质检 → 成稿	要发布、事实敏感、风格要求高或长文任务

自动判断规则：

用户明确说“不用查资料”“只改表达”时，用
```
quick
```
用户明确要“发公众号/知乎/Newsletter”“深度改写”“补资料”“像一篇正式文章”时，用
```
publish
```
只要任务包含明显翻译难点、长文、或用户要“成稿级中文”，至少用
```
standard
```
其他情况默认
```
standard
```

Mode	Trigger Words	Core Steps	Applicable Scenarios
Quick	"Simple polish", "Quick revision", "No external research"	Constraint confirmation → Terminology unification (if needed) → Direct draft generation	Short articles, light polishing, users explicitly request no external research
Standard	Default	Analysis → Selective research → Prompt snapshot → Draft → Critique → Revision → Expression optimization → Final draft	Most article rewriting needs, especially "translate and rewrite"
Publish	"In-depth rewrite", "WeChat Official Account draft", "Suitable for publication", "Supplement information and rewrite"	Analysis → Verification → Reorganization → Prompt snapshot → Draft → Critique → Revision → Polishing → Expression optimization → Quality inspection → Final draft	Needs publication, fact-sensitive content, high style requirements or long article tasks

Automatic Judgment Rules:

Use
```
quick
```
when users explicitly say "no external research" or "only revise expression"
Use
```
publish
```
when users explicitly request "publish on WeChat Official Account/Zhihu/Newsletter", "in-depth rewrite", "supplement information" or "like a formal article"
Use at least
```
standard
```
when the task involves obvious translation difficulties, long articles, or users request "draft-level Chinese"
Use
```
standard
```
by default in other cases

输出目录

Output Directory

为每次任务创建输出目录：

rewrites/{slug}-{timestamp}/

文件	何时需要	说明
`source.md`	全部	落盘后的原始内容
`00-brief.md`	全部	任务约束、平台、受众、篇幅、检索边界、是否允许重组
`01-analysis.md`	`standard` / `publish`	主题、结构、受众、保留项、难点、理解障碍、改写策略
`01-terms.md`	有翻译或术语密集时	本次任务的术语表、推荐译法、禁用译法、待确认译法
`02-research.md`	做了外查时	外部资料、事实核验、冲突信息、可引用要点
`03-outline.md`	`publish` 或重组明显时	新结构、大纲、标题角度、叙述顺序
`04-prompt.md`	非平凡任务默认产出	写作指令快照，供草稿、分块和后续修订复用
`05-draft.md`	`standard` / `publish`	初稿；必要时带 translator notes
`06-critique.md`	`standard` / `publish`	诊断问题，不直接改文
`07-revision.md`	`standard` / `publish`	吃掉批评意见后的修订稿
`08-polish.md`	`publish`	中文化、节奏、标题、平台适配后的抛光稿
`09-expression.md`	`standard` / `publish` 的非平凡任务	去除翻译腔、AI 腔、套话和机械连接感后的表述优化稿
`qa.md`	`publish` 或事实敏感任务	最终质检清单与残余风险
`rewrite.md`	全部	最终交付稿

如果用户只要求聊天里直接返回结果，仍然默认保存

rewrite.md

，并在回复里给出路径。

Create an output directory for each task:

rewrites/{slug}-{timestamp}/

File	When Needed	Description
`source.md`	All tasks	Original content after saving to disk
`00-brief.md`	All tasks	Task constraints, platform, audience, length, retrieval boundaries, whether reorganization is allowed
`01-analysis.md`	`standard` / `publish`	Topic, structure, audience, reserved items, difficulties, comprehension barriers, rewriting strategies
`01-terms.md`	When there is translation or dense terminology	Terminology table for this task, recommended translations, forbidden translations, pending translations
`02-research.md`	When external research is conducted	External materials, fact verification, conflicting information, quotable key points
`03-outline.md`	`publish` or obvious reorganization needed	New structure, outline, title angles, narrative order
`04-prompt.md`	Default for non-trivial tasks	Writing instruction snapshot, reused for drafting, chunking and subsequent revisions
`05-draft.md`	`standard` / `publish`	First draft; include translator notes if necessary
`06-critique.md`	`standard` / `publish`	Diagnose problems without directly revising the text
`07-revision.md`	`standard` / `publish`	Revised draft after addressing critique comments
`08-polish.md`	`publish`	Polished draft after localization, rhythm adjustment, title optimization and platform adaptation
`09-expression.md`	Non-trivial tasks in `standard` / `publish`	Expression optimization draft after removing translationese, AI tone, clichés and mechanical connections
`qa.md`	`publish` or fact-sensitive tasks	Final quality inspection checklist and residual risks
`rewrite.md`	All tasks	Final deliverable

If users only request direct results in the chat, still save

rewrite.md

by default and provide the path in the response.

平台与文风

Platforms and Writing Styles

平台、段落节奏、标题风格不要写死在主文件里。按需阅读：

references/style-presets.md：公众号、知乎、Newsletter、通用版的文风和标题约束
references/glossary-en-zh.md：英文原文或中英混排材料中的特殊译法
references/quality-rubric.md：批评稿与最终质检时的固定检查维度
references/expression-optimizer.md：最终表述优化时要清理的翻译腔、AI 腔和套话模式

选择逻辑：

用户明确指定平台时，优先使用对应预设
用户未指定平台时，先根据原文类型和目标受众判断；仍不明确则使用
```
generic
```
技术、商业、社会议题都可以有叙事感，但不要强行“鸡血化”或“爆款化”

Do not hardcode platforms, paragraph rhythms or title styles in the main file. Read on demand:

[references/style-presets.md]: Writing style and title constraints for WeChat Official Account, Zhihu, Newsletter and generic versions
[references/glossary-en-zh.md]: Special translations for English originals or mixed Chinese-English materials
[references/quality-rubric.md]: Fixed inspection dimensions for critique drafts and final quality inspection
[references/expression-optimizer.md]: Patterns of translationese, AI tone and clichés to be cleaned up during final expression optimization

Selection Logic:

Prioritize using corresponding presets when users explicitly specify a platform
When no platform is specified, first judge based on original text type and target audience; use
```
generic
```
if still unclear
Technical, commercial and social topics can have narrative elements, but avoid forced "鸡血化 (overly motivational)" or "爆款化 (viral-style)" styles

术语处理

Terminology Handling

当任务包含以下任一情况时，必须先做术语统一，再进入改写：

原文是英文
原文是中英混排
原文涉及 AI、产品、商业、技术等高频专有名词
用户明确要求“翻译并改写”

处理顺序：

优先读取内置术语表 references/glossary-en-zh.md
从原文额外提取专有名词、缩写、产品名、方法论名词
形成本次任务的会话术语表，写入
```
01-terms.md
```
标出推荐译法、禁用译法、保留原词、待定项
后续草稿、批评、修订都以
```
01-terms.md
```
为准，不得随意漂移

执行原则：

常见且约定俗成的词，直接用标准中文
容易误译、风格差异大的词，优先遵循内置术语表

首次出现时，可使用

中文（English）

或

中文（英文原词，简短解释）

如果术语存在多种合理译法，选择最适合目标平台和受众的一种，并全文保持一致
改写允许重组句子，但不要把关键术语“改丢”或改成过度口语化的说法

Terminology unification must be completed before rewriting if any of the following situations apply:

The original text is in English
The original text is mixed Chinese-English
The original text involves high-frequency proper nouns in AI, product, business or technical fields
Users explicitly request "translate and rewrite"

Processing Sequence:

Prioritize reading the built-in terminology table [references/glossary-en-zh.md]
Extract additional proper nouns, abbreviations, product names and methodology terms from the original text
Form a session-specific terminology table for this task and write it to
```
01-terms.md
```
Mark recommended translations, forbidden translations, retained original terms and pending items
Subsequent drafts, critiques and revisions must follow
```
01-terms.md
```
and must not deviate arbitrarily

Implementation Principles:

Use standard Chinese directly for common and established terms
Prioritize following the built-in terminology table for terms that are easily mistranslated or have large style differences

When first appearing, use

Chinese (English)

Chinese (English original, brief explanation)

If multiple reasonable translations exist for a term, choose the one most suitable for the target platform and audience, and maintain consistency throughout the text
Rewriting allows sentence reorganization, but do not "lose" key terms or convert them to overly colloquial expressions

工作流

Workflow

Step 1: 获取约束并写入

00-brief.md

Step 1: Obtain Constraints and Write to

00-brief.md

优先从当前对话提取，不要为了默认值反复追问。至少确认：

输入来源：文本、文件、URL
用户目标：润色、扩写、翻译并改写、发文成稿、平台适配
是否允许外部检索
是否要保留原文结构
是否要控制篇幅
目标平台与受众
事实敏感度：普通 / 中 / 高

如果用户没有说明：

默认允许“必要时补充少量资料”，但不要为静态内容滥搜
默认允许重组结构，但要保留原文核心论点
默认提供 5 个标题候选

00-brief.md

建议记录：

Source type
Goal
Platform
Audience
Length target
Research boundary
Preserve structure: yes / no / partial
Fact sensitivity
Output promise

Prioritize extracting from the current conversation; do not repeatedly ask for default values. At least confirm:

Input source: text, file, URL
User goal: polish, expand, translate and rewrite, publishable draft, platform adaptation
Whether external retrieval is allowed
Whether to retain the original structure
Whether to control length
Target platform and audience
Fact sensitivity: normal / medium / high

If users do not specify:

Default to "supplement small amounts of information if necessary", but avoid excessive searching for static content
Default to allowing structural reorganization, but retain the core arguments of the original text
Default to providing 5 title candidates

00-brief.md

should record:

Source type
Goal
Platform
Audience
Length target
Research boundary
Preserve structure: yes / no / partial
Fact sensitivity
Output promise

Step 2: 落盘原文并创建输出目录

Step 2: Save Original Text to Disk and Create Output Directory

将输入内容统一落盘到
```
source.md
```
创建输出目录
```
rewrites/{slug}-{timestamp}/
```
检测原文语言
如果原文不是中文，先加载内置术语表并提取会话术语，写入
```
01-terms.md
```
如果原文很长，先确定分块边界和合并策略，再进入草稿

要求：

URL 输入优先抽取正文，避免导航、推荐位、评论区污染
文件或 URL 抽取失败时，明确说明阻塞点；如果仍可继续，退回到用户已提供的正文
分块时按语义边界切，不要机械按字数均分

Save all input content to
```
source.md
```
uniformly
Create output directory
```
rewrites/{slug}-{timestamp}/
```
Detect the language of the original text
If the original text is not in Chinese, first load the built-in terminology table and extract session-specific terms, then write to
```
01-terms.md
```
If the original text is long, first determine chunk boundaries and merging strategies before entering the drafting stage

Requirements:

For URL input, prioritize extracting the main text to avoid contamination from navigation, recommendation sections and comment areas
If file or URL extraction fails, clearly state the blocking point; if still proceedable, fall back to the main text provided by the user
Chunk by semantic boundaries, not mechanically by word count

Step 3: 分析原文并写入

01-analysis.md

Step 3: Analyze Original Text and Write to

01-analysis.md

standard

和

publish

模式必须产出

01-analysis.md

。至少包含：

原文主题、核心论点、作者意图
目标受众推断
原文结构拆解
必须保留的事实、数字、案例、引用
关键术语及建议译法
读者理解障碍：背景知识缺口、代词指代、长句、跳步论证
比喻与文化映射：哪些隐喻应保留、转译、解释或降级
语气目标：克制 / 叙事 / 评论 / 方法论 / 发布稿
可优化点：过长、重复、跳跃、空泛、事实老旧、结论太早或太晚
改写策略：保守重写 / 结构重组 / 翻译后二次创作

如果原文已经写得很好，分析里要明确“以增量和可读性优化为主”，不要强行颠覆。

01-analysis.md

must be produced in

standard

and

publish

modes. At least include:

Original text topic, core arguments, author's intent
Inferred target audience
Original text structure breakdown
Facts, figures, cases and citations that must be retained
Key terms and recommended translations
Reader comprehension barriers: background knowledge gaps, pronoun references, long sentences, jumpy arguments
Metaphors and cultural mappings: which metaphors should be retained, translated, explained or downplayed
Tone target: restrained / narrative / comment / methodological / press release
Optimizable points: overly long, repetitive, jumpy, vague, outdated facts, conclusions too early or too late
Rewriting strategy: conservative rewriting / structural reorganization / secondary creation after translation

If the original text is already well-written, clearly state in the analysis that "optimization should focus on incremental improvements and readability" instead of forced subversion.

Step 4: 调研与事实核验

Step 4: Research and Fact Verification

何时需要做

02-research.md

：

用户明确要求补充资料、最新信息、案例、数据
原文含有明显时效性内容
原文中的数字、研究、公司信息可能已经过期
主题本身需要最基本的事实校验

按需阅读：

references/research-rules.md：检索范围、来源优先级、冲突处理、引用格式

执行原则：

```
quick
```
模式默认不主动外查
用户明确禁止检索时，跳过外查，并在最终结果里注明“本次基于原文改写，未补充外部资料”
```
standard
```
模式只查能显著提高文章可信度的关键信息
```
publish
```
模式要把“外部事实”“作者观点”“你的推断”分开
调研文件只记录“事实与来源”，不要把语言问题混写进去

When to create

02-research.md

Users explicitly request supplementary materials, latest information, cases or data
The original text contains obviously time-sensitive content
Figures, studies or company information in the original text may be outdated
The topic itself requires basic fact verification

Read on demand:

[references/research-rules.md]: Retrieval scope, source priority, conflict handling, citation format

Implementation Principles:

```
quick
```
mode defaults to no active external research
Skip external research if users explicitly prohibit it, and note in the final result "This rewrite is based on the user-provided original text with no additional external materials"
```
standard
```
mode only searches for key information that can significantly improve article credibility
```
publish
```
mode must separate "external facts", "author's opinions" and "your inferences"
Research files only record "facts and sources", do not mix in language issues

Step 5: 规划新结构

Step 5: Plan New Structure

以下情况必须写

03-outline.md

：

```
publish
```
模式
原文结构明显混乱
用户要求平台适配且风格变化较大
需要从翻译转为“适合中文发布的成稿”

至少包括：

标题方向：洞察型 / 利益型 / 反常识型 / 场景型 / 提问型
开头策略：故事、场景、反差、问题、结论先行
主体段落顺序
每段要完成的作用
结尾动作：总结、提醒、方法、下一步建议

standard

模式可以不单独落盘大纲，但心里必须先确定结构再写。

03-outline.md

must be written in the following cases:

```
publish
```
mode
The original text structure is obviously chaotic
Users request platform adaptation with significant style changes
Need to convert from translation to "final draft suitable for Chinese publication"

At least include:

Title directions: insight-based / benefit-based / counterintuitive / scenario-based / question-based
Opening strategies: story, scenario, contrast, question, conclusion first
Main body paragraph order
Purpose of each paragraph
Closing actions: summary, reminder, method, next-step suggestions

standard

mode does not need to save the outline to disk separately, but must determine the structure before writing.

Step 6: 组装 Prompt 快照并写入

04-prompt.md

Step 6: Assemble Prompt Snapshot and Write to

04-prompt.md

非平凡任务默认产出

04-prompt.md

。它不是给用户看的终稿，而是给后续草稿、分块、修订复用的统一指令。

至少写清楚：

任务目标
输出语言与平台
目标受众
必保留信息
允许重组的程度
术语约束
禁止事项
调研结论或时间边界
文风要求
标题要求

如果任务需要分块：

每个 chunk 都复用同一份
```
04-prompt.md
```
每个 chunk 只处理自己的正文，不自行发明全局结论
合并后必须再做一次整体批评和统一修订

04-prompt.md

is produced by default for non-trivial tasks. It is not a final draft for users, but a unified instruction reused for subsequent drafting, chunking and revisions.

At least clearly state:

Task goal
Output language and platform
Target audience
Information that must be retained
Allowable degree of reorganization
Terminology constraints
Prohibited items
Research conclusions or time boundaries
Writing style requirements
Title requirements

If the task requires chunking:

Each chunk reuses the same
```
04-prompt.md
```
Each chunk only processes its own main text and does not invent global conclusions on its own
Must conduct a unified critique and revision of the merged full draft

Step 7: 生成

05-draft.md

Step 7: Generate

05-draft.md

05-draft.md

是初稿，不追求一次成稿，但必须把原文的骨架和信息先落稳。

执行原则：

原文是外文时，初稿先保证信息完整、关系准确、术语稳定
允许带简短 translator notes，专门记录难点、歧义、取舍
不在初稿阶段强行抖机灵或堆修辞
如果是分块任务，先得到所有 chunk 的初稿，再合并成统一
```
05-draft.md
```

05-draft.md

is the first draft, which does not need to be perfect, but must stabilize the framework and information of the original text first.

Implementation Principles:

If the original text is in a foreign language, first ensure complete information, accurate relationships and stable terminology in the first draft
Allow short translator notes to record difficulties, ambiguities and trade-offs
Avoid forced wit or rhetorical stacking in the drafting stage
If it is a chunked task, first obtain the first draft of all chunks, then merge into a unified
```
05-draft.md
```

Step 8: 生成

06-critique.md

Step 8: Generate

06-critique.md

06-critique.md

只做诊断，不直接改文。优先读取 references/quality-rubric.md。

至少从以下维度逐项检查：

Accuracy：是否误译、漏译、过译、关系错位
Completeness：关键限定条件、事实、例子是否丢失
Natural Chinese：是否有欧化句法、硬译、别扭搭配
Terminology：术语是否一致，首次解释是否合理
Structure：段落顺序和因果链是否清楚
Register / Platform Fit：是否符合公众号、知乎、Newsletter 或正式文风
Fact Boundary：推断和事实是否混淆，时效信息是否处理得当
Title / Framing：标题、开头、正文承诺是否一致

写法要求：

逐项点出问题和影响
优先说“哪里不对，为什么不对，应该往哪改”
不要在批评稿里顺手重写全文

06-critique.md

only diagnoses problems and does not directly revise the text. Prioritize reading [references/quality-rubric.md].

At least check item by item from the following dimensions:

Accuracy: Whether there is mistranslation, omission, over-translation or relationship misalignment
Completeness: Whether key constraints, facts and examples are missing
Natural Chinese: Whether there are Europeanized syntax, literal translations or awkward collocations
Terminology: Whether terminology is consistent and first-time explanations are reasonable
Structure: Whether paragraph order and causal chains are clear
Register / Platform Fit: Whether it conforms to WeChat Official Account, Zhihu, Newsletter or formal writing styles
Fact Boundary: Whether inferences and facts are confused, and time-sensitive information is properly handled
Title / Framing: Whether titles, openings and text promises are consistent

Writing Requirements:

Point out problems and their impacts item by item
Prioritize stating "what is wrong, why it is wrong, and how to revise it"
Do not rewrite the full text directly in the critique draft

Step 9: 生成

07-revision.md

Step 9: Generate

07-revision.md

把

06-critique.md

的问题逐条吃掉，得到结构和表达都更稳定的修订稿。

执行原则：

先修准确性、完整性、逻辑，再修语言和节奏
如果批评中有互相冲突的建议，优先保事实与清晰度
修订时不得引入新事实，除非这些事实已在
```
02-research.md
```
中核实

Address all problems in

06-critique.md

to get a revised draft with more stable structure and expression.

Implementation Principles:

First revise accuracy, completeness and logic, then revise language and rhythm
If there are conflicting suggestions in the critique, prioritize facts and clarity
Do not introduce new facts during revision unless they have been verified in
```
02-research.md
```

Step 10: 生成

08-polish.md

Step 10: Generate

08-polish.md

publish

模式必须做抛光，

standard

模式在用户要求“像成稿”“可直接发”时也应做。

抛光只解决以下问题：

把翻译腔改成自然中文
拉顺段落节奏和过渡
调整标题、开头、结尾的抓力与克制感
清理重复表达、空泛修辞、解释过度
统一小标题、标点、格式和篇章密度

抛光阶段不要：

重新发明观点
引入未核实事实
因为追求“顺”而删掉关键限定词

Polishing must be done in

publish

mode, and should also be done in

standard

mode when users request "like a final draft" or "directly publishable".

Polishing only addresses the following problems:

Convert translationese to natural Chinese
Smooth out paragraph rhythm and transitions
Adjust the attractiveness and restraint of titles, openings and closings
Clean up repetitive expressions, vague rhetoric and over-explanations
Unify subheadings, punctuation, formatting and chapter density

Do not do the following during polishing:

Reinvent viewpoints
Introduce unverified facts
Delete key qualifiers for the sake of "smoothness"

Step 11: 生成

09-expression.md

Step 11: Generate

09-expression.md

在结构、事实、节奏都已经稳定后，再单独做一次“优化表述”。优先读取 references/expression-optimizer.md。

这一步只处理表达层，不改论点和事实。核心目标：

去掉翻译腔、AI 腔、宣传腔
删掉空泛拔高、机械连接词、公式化三段式
把模糊归因改成具体归因，或改成保守表达
收紧过度修辞、口号句、像摘抄金句的句子
让段落更像真实作者写出来的中文，而不是“很会写的模型”

重点排查：

过度强调意义、历史地位、宏大趋势
“此外 / 值得注意的是 / 不仅如此 / 归根结底”这类连接拐杖
“不仅……而且……”“这不仅是……更是……”等否定式排比
生硬的三项并列、假完整感
模糊归因，如“专家认为”“业内普遍认为”“有观点指出”
过度使用破折号、粗体、抽象名词和口号化结尾
为显得不重复而频繁换同义词，导致术语或主语漂移

执行原则：

保留原意，不借“人味”之名擅自加观点
平台越正式，越要克制；不要把正式文风硬改成社交媒体口吻
允许自然的节奏变化，但不要故意“演”出松弛感
如果原文本来就是高度克制的技术或正式文体，优化目标是去机械感，不是加情绪

After the structure, facts and rhythm are stable, conduct a separate "expression optimization" round. Prioritize reading [references/expression-optimizer.md].

This step only handles the expression layer and does not revise arguments or facts. Core goals:

Remove translationese, AI tone and propaganda tone
Delete vague exaggerations, mechanical conjunctions and formulaic three-part structures
Convert vague attributions to specific attributions or conservative expressions
Tighten excessive rhetoric, slogan sentences and sentences that look like quoted golden lines
Make paragraphs look like Chinese written by a real author instead of a "very capable model"

Key Inspection Points:

Overemphasis on significance, historical status and grand trends
Conjunction crutches such as "此外 / 值得注意的是 / 不仅如此 / 归根结底" (furthermore / it is worth noting that / not only that / in the final analysis)
Negative parallel structures such as "不仅……而且……" (not only...but also...) and "这不仅是……更是……" (this is not only...but also...)
Stilted three-item parallelism, false sense of completeness
Vague attributions such as "experts believe", "it is widely believed in the industry", "some opinions point out"
Excessive use of dashes, bold text, abstract nouns and slogan-style closings
Frequent synonym replacement to avoid repetition, leading to terminology or subject drift

Implementation Principles:

Retain the original meaning; do not add opinions in the name of "human touch"
The more formal the platform, the more restrained the style; do not forcefully convert formal writing styles to social media tones
Allow natural rhythm changes, but do not deliberately "perform" a relaxed tone
If the original text is already a highly restrained technical or formal style, the optimization goal is to remove mechanicalness, not add emotion

Step 12: 质检并保存最终稿

Step 12: Quality Inspection and Save Final Draft

publish

模式和事实敏感任务必须写

qa.md

，至少包含：

是否有漏译、误译、过译
是否有未经核实的新事实
术语是否前后一致
标题是否准确反映正文
是否保留了原文最重要的事实、数字、案例、引用
是否仍有需要保守表述的地方
```
09-expression.md
```
是否只改表述，没有偷改事实和判断

最终稿始终保存到

rewrite.md

。

如果做了

09-expression.md

：

```
rewrite.md
```
基于
```
09-expression.md
```

如果没有单独做表述优化，但做了

08-polish.md

：

```
rewrite.md
```
基于
```
08-polish.md
```

如果没有单独抛光：

```
rewrite.md
```
基于
```
07-revision.md
```

qa.md

must be written for

publish

mode and fact-sensitive tasks, at least including:

Whether there is mistranslation, omission or over-translation
Whether there are unverified new facts
Whether terminology is consistent throughout
Whether titles accurately reflect the main text
Whether the most important facts, figures, cases and citations of the original text are retained
Whether there are still areas requiring conservative expressions
Whether
```
09-expression.md
```
only revised expressions without altering facts and judgments

The final draft is always saved to

rewrite.md

09-expression.md

was created:

```
rewrite.md
```
is based on
```
09-expression.md
```

If no separate expression optimization was done but

08-polish.md

was created:

```
rewrite.md
```
is based on
```
08-polish.md
```

If no separate polishing was done:

```
rewrite.md
```
is based on
```
07-revision.md
```

Step 13: 生成标题候选

Step 13: Generate Title Candidates

默认提供 5 个标题，优先覆盖不同角度，而不是只换几个词：

洞察型
利益型
反常识型
场景型
提问型

标题规则：

必须准确反映正文核心价值
不伪造数字，不夸大结论，不制造不存在的紧迫感
平台越正式，标题越克制
同一组标题之间角度要明显不同

如果用户明确说“不需要标题”，则跳过。

Provide 5 title candidates by default, prioritizing different angles instead of just changing a few words:

Insight-based
Benefit-based
Counterintuitive
Scenario-based
Question-based

Title Rules:

Must accurately reflect the core value of the main text
Do not fabricate figures, exaggerate conclusions or create non-existent urgency
The more formal the platform, the more restrained the title
The angles of titles in the same group must be significantly different

Skip this step if users explicitly say "no titles needed".

长文与分块策略

Long Article and Chunking Strategy

长文不要直接硬翻到底。采用“统一约束、局部分块、整体复核”的方式。

执行顺序：

先完成

00-brief.md

、

01-analysis.md

、

01-terms.md

、

04-prompt.md

按语义边界分块，不按字符数机械切块
每个 chunk 共享同一术语表和 Prompt 快照
合并所有 chunk 初稿到
```
05-draft.md
```
对合并后的完整稿统一做
```
06-critique.md
```
统一修订后再决定是否进入
```
08-polish.md
```
无论是否单独抛光，最终都要对合并稿做一次表述优化，再输出
```
rewrite.md
```

特别注意：

分块只解决上下文窗口问题，不意味着每块都能独立定调
所有全局性的判断、标题、开头、结尾，都应该在合并后统一处理
合并稿必须检查术语漂移、重复解释、前后口径不一致
表述优化必须在合并后的全文上做，不能每个 chunk 各自“人性化”

Do not directly translate long articles in one go. Adopt the approach of "unified constraints, local chunking, overall review".

Implementation Sequence:

First complete

00-brief.md

01-analysis.md

01-terms.md

and

04-prompt.md

Chunk by semantic boundaries, not mechanically by character count
Each chunk shares the same terminology table and prompt snapshot
Merge all chunk first drafts into
```
05-draft.md
```
Conduct unified
```
06-critique.md
```
for the merged full draft
After unified revision, decide whether to enter
```
08-polish.md
```
Regardless of whether separate polishing is done, always conduct expression optimization on the merged full draft before outputting
```
rewrite.md
```

Special Notes:

Chunking only solves the context window problem and does not mean each chunk can set its own tone independently
All global judgments, titles, openings and closings should be handled uniformly after merging
Must check for terminology drift, repeated explanations and inconsistent statements in the merged draft
Expression optimization must be done on the merged full text, not separately "humanized" for each chunk

最终输出格式

Final Output Format

rewrite.md

默认使用以下结构：

markdown

undefined

rewrite.md

uses the following structure by default:

markdown

undefined

标题候选

Title Candidates

标题一
标题二
标题三
标题四
标题五

Title 1
Title 2
Title 3
Title 4
Title 5

正文

Main Text

[改写后的完整文章]

[Rewritten full article]

参考来源

Reference Sources

来源标题
来源标题

Source Title
Source Title

改写说明

Rewrite Notes

模式：standard
平台：wechat
是否补充外部资料：是
是否使用术语表：是
是否经过批评修订：是
是否经过表述优化：是


如果未外查，把“参考来源”改为：

```markdown

Mode: standard
Platform: wechat
Supplementary external materials: Yes
Terminology table used: Yes
Critique and revision conducted: Yes
Expression optimization conducted: Yes


If no external research was done, change "Reference Sources" to:

```markdown

参考来源

Reference Sources

本次基于用户提供原文改写，未补充外部资料

undefined

This rewrite is based on the user-provided original text with no additional external materials

undefined

质量门槛

Quality Threshold

信息密度应高于原文，而不是仅仅换说法
如果扩写，新增内容要么解释关键概念，要么补足案例/数据/边界
事实敏感内容应优先核验，不确定时宁可保守表达
风格要有人味，但不要表演式抒情
如果原文质量本来很高，应以“精修”代替“重写”
有翻译成分的任务，至少要做一次显式批评与修订
长文或分块任务，必须对合并后的全文再做统一审校
非平凡任务在最终交付前，必须再做一次独立的表述优化

Information density should be higher than the original text, not just rephrasing
If expanding, new content should either explain key concepts or supplement cases/data/boundaries
Fact-sensitive content should be verified first; use conservative expressions if uncertain
The style should have a human touch, but avoid performative sentimentality
If the original text is already high-quality, use "precision polishing" instead of "rewriting"
Tasks involving translation must include at least one explicit critique and revision round
Long articles or chunked tasks must include a unified review of the merged full draft
Non-trivial tasks must undergo a separate expression optimization round before final delivery

禁止事项

Prohibited Items

不要把未经核实的信息包装成“最新研究表明”
不要为了制造传播感而扭曲结论
不要照搬大段原句，只做词语替换
不要堆砌空泛修辞、套话或过量感叹句
不要默认所有文章都适合“爆款标题”
不要在
```
06-critique.md
```
里顺手重写全文，混淆诊断和修订
不要让
```
08-polish.md
```
变成又一轮无边界改写
不要让
```
09-expression.md
```
借“优化表述”之名偷改事实、立场或信息密度

Do not package unverified information as "latest research shows"
Do not distort conclusions to create viral potential
Do not copy large sections of the original text with only word substitutions
Do not stack vague rhetoric, clichés or excessive exclamation marks
Do not assume all articles are suitable for "viral titles"
Do not rewrite the full text directly in
```
06-critique.md
```
and confuse diagnosis with revision
Do not allow
```
08-polish.md
```
to become another round of unbounded rewriting
Do not allow
```
09-expression.md
```
to alter facts, positions or information density in the name of "expression optimization"