translate-polisher

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

译者

Translator

四步精翻工作流，产出出版级翻译：分析 → 初译 → 审校 → 终稿。

Four-step professional translation workflow, producing publication-level translations: Analysis → Initial Translation → Review → Final Draft.

用法

Usage

/translate [--from <lang>] [--to <lang>] [--audience <audience>] [--style <style>] [--glossary <file>] <source>

```
<source>
```
：文件路径、URL，或直接在对话中贴入文章正文
```
--from
```
：源语言（省略则自动检测）
```
--to
```
：目标语言（默认
```
zh-CN
```
）
```
--audience
```
：目标读者（默认
```
general
```
）
```
--style
```
：翻译风格（默认
```
auto
```
）
```
--glossary
```
：额外术语表文件，与内置术语表合并
支持语言对：
```
ZH↔EN
```
、
```
ZH↔JA
```
不支持：
```
EN↔JA
```
直译

/translate [--from <lang>] [--to <lang>] [--audience <audience>] [--style <style>] [--glossary <file>] <source>

```
<source>
```
: File path, URL, or directly paste the article text in the conversation
```
--from
```
: Source language (auto-detected if omitted)
```
--to
```
: Target language (default
```
zh-CN
```
)
```
--audience
```
: Target readers (default
```
general
```
)
```
--style
```
: Translation style (default
```
auto
```
)
```
--glossary
```
: Additional glossary file, merged with built-in glossary
Supported language pairs:
```
ZH↔EN
```
,
```
ZH↔JA
```
Not supported: Direct translation between
```
EN↔JA
```

读者预设

Audience Presets

值	说明	效果
`general`	普通读者（默认）	平实语言，术语多加译注
`technical`	开发者/工程师	常见技术术语少加注释
`academic`	研究者/学者	正式语域，术语精确
`business`	商务人士	商务友好，解释技术概念

也接受自定义描述，如

--audience "对 AI 感兴趣的普通读者"

。

Value	Description	Effect
`general`	General readers (default)	Plain language, with additional annotations for technical terms
`technical`	Developers/engineers	Fewer annotations for common technical terms
`academic`	Researchers/scholars	Formal register, precise terminology
`business`	Business professionals	Business-friendly, explains technical concepts

Custom descriptions are also accepted, e.g.,

--audience "general readers interested in AI"

风格预设

Style Presets

值	说明
`auto`	根据分析阶段识别的原文口吻自动匹配翻译风格（默认）
`storytelling`	叙事流畅，过渡自然，措辞生动
`formal`	专业严谨，结构清晰，无口语化
`technical`	精确简洁，术语密集，少修饰
`literal`	贴近原文结构，最少重组
`academic`	学术严谨，复杂从句可接受
`business`	简洁结果导向，行动化
`humorous`	保留并适配幽默
`conversational`	口语化，亲切随和
`elegant`	文学性，韵律考究，字斟句酌

也接受自定义描述，如

--style "诗意而抒情"

。

Value	Description
`auto`	Automatically match the translation style based on the original tone identified in the analysis phase (default)
`storytelling`	Smooth narration, natural transitions, vivid wording
`formal`	Professional and rigorous, clear structure, no colloquialisms
`technical`	Precise and concise, terminology-intensive, minimal embellishments
`literal`	Close to original structure, minimal restructuring
`academic`	Academically rigorous, complex clauses are acceptable
`business`	Concise and result-oriented, action-oriented
`humorous`	Preserve and adapt humor
`conversational`	Colloquial, friendly and approachable
`elegant`	Literary, pays attention to rhythm and word choice

Custom descriptions are also accepted, e.g.,

--style "poetic and lyrical"

工作流

Workflow

第一步：准备

Step 1: Preparation

根据语言对加载内置术语表：
- ZH→EN / EN→ZH：references/glossary-en-zh.md
- ZH→JA / JA→ZH：references/glossary-zh-ja.md
- 当中日文本中混有英文技术术语、产品名或缩写时，可额外参照 references/glossary-en-ja.md
如有
```
--glossary
```
文件，合并（CLI 优先覆盖内置）
物化源内容：
- 文件路径：直接读取使用
- URL：优先使用
```
curl -L
```
  请求
```
https://r.jina.ai/http://<url>
```
  获取正文 Markdown，保存到
```
translate/{slug}.md
```
  - 不要优先使用
```
urllib
```
    之类的默认 HTTP 客户端；这类实现更容易遇到
```
403
```
    或被目标侧拦截
  - 若返回缺少正文、只拿到导航/摘要、被 paywall 截断、或请求失败，立即停止流程
  - 停止时明确告诉用户：当前 URL 无法稳定抓到完整正文，请用户自行下载、复制或粘贴原文后再翻译
  - 不允许在正文缺失时继续分析、概括或翻译
- 对话中贴入的正文：直接使用，无需保存文件
创建临时工作目录用于中间文件（最终会清理）
自动检测源语言（除非指定
```
--from
```
）

Load built-in glossaries based on language pairs:
- ZH→EN / EN→ZH: references/glossary-en-zh.md
- ZH→JA / JA→ZH: references/glossary-zh-ja.md
- When Chinese-Japanese texts contain English technical terms, product names or abbreviations, additionally refer to references/glossary-en-ja.md
If
```
--glossary
```
file is provided, merge it (CLI settings override built-in entries)
Materialize source content:
- File path: Directly read and use
- URL: Prioritize using
```
curl -L
```
  to request
```
https://r.jina.ai/http://<url>
```
  to get the body content in Markdown, save to
```
translate/{slug}.md
```
  - Do not prioritize default HTTP clients like
```
urllib
```
    ; such implementations are more likely to encounter
```
403
```
    errors or be blocked by the target side
  - If the returned content lacks the main body, only gets navigation/summaries, is truncated by paywalls, or the request fails, stop the process immediately
  - When stopping, clearly inform the user: "The full body content cannot be stably fetched from the current URL. Please download, copy or paste the original text and try again."
  - Do not continue analysis, summarization or translation with incomplete content
- Text pasted in conversation: Directly use without saving to file
Create a temporary working directory for intermediate files (will be cleaned up eventually)
Auto-detect source language (unless
```
--from
```
is specified)

第二步：深度分析 →

01-analysis.md

Step 2: In-depth Analysis →

01-analysis.md

翻译前深入分析源材料，聚焦直接影响翻译质量的维度。

Conduct in-depth analysis of the source material before translation, focusing on dimensions that directly affect translation quality.

2.1 概要

2.1 Overview

3-5 句话概括：内容主题、核心论点、最有价值的观点。

Summarize in 3-5 sentences: content theme, core arguments, most valuable insights.

2.2 核心内容

2.2 Core Content

核心论点：一句话总结
关键概念：作者使用了哪些关键概念？如何定义？
结构：论证如何展开？各部分如何衔接？
论据：使用了哪些具体例子、数据或权威引用？

Core Argument: One-sentence summary
Key Concepts: What key concepts does the author use? How are they defined?
Structure: How is the argument developed? How do the parts connect?
Evidence: What specific examples, data or authoritative citations are used?

2.3 背景语境

2.3 Context

作者：谁？背景和立场？
写作语境：回应什么现象、趋势或争论？
目的：试图解决什么问题？影响谁？
隐含假设：论证背后有哪些未明说的前提？

Author: Who is the author? What is their background and stance?
Writing Context: What phenomenon, trend or debate is it responding to?
Purpose: What problem is it trying to solve? Who is the target audience?
Implicit Assumptions: What unstated premises underlie the argument?

2.4 术语提取

2.4 Term Extraction

列出所有技术术语、专有名词、品牌名、缩写
与已加载术语表交叉比对
不在术语表中的术语，查找标准译法
记录到工作术语表

List all technical terms, proper nouns, brand names, abbreviations
Cross-check with loaded glossaries
For terms not in glossaries, find standard translations
Record to working glossary

2.5 语气与风格

2.5 Tone and Style

原文是正式还是口语化？
是否使用幽默、隐喻或文化典故？
考虑目标读者，翻译应采用什么语域？
若目标语言为日语，必须在分析阶段先锁定文体：
```
敬体（です・ます）
```
或
```
常体（だ・である）
```
二选一，并说明选择依据。默认规则：评论、分析、技术文章、博客正文优先常体；面向读者的说明、公告、邮件、FAQ、客服式文本优先敬体。除引语、UI 固有文案、引用原文片段外，正文不得混用

Is the original text formal or colloquial?
Does it use humor, metaphors or cultural allusions?
Considering the target audience, what register should the translation adopt?
If the target language is Japanese, the style must be locked in the analysis phase: choose either Keigo (です・ます) or Plain Form (だ・である) and explain the basis for selection. Default rules: comments, analyses, technical articles, blog bodies prioritize plain form; reader-facing instructions, announcements, emails, FAQs, customer service-style texts prioritize keigo. Except for quotations, inherent UI copy, or cited original fragments, the main text must not mix styles

2.6 读者理解障碍

2.6 Reader Comprehension Barriers

识别目标读者可能遇到困难的地方（根据目标受众校准）：

领域术语：缺乏广为人知的译法，或直译无意义
文化典故：源文化特有的习语、历史事件、流行文化
隐含知识：原作者假设读者具备但目标读者可能缺乏的背景
文字游戏与隐喻：跨语言无法直接传递的修辞
命名概念：有专门名称的理论、效应（如"梳头效应"、"邓宁-克鲁格效应"）
中国特有名词：平台、App、机构、政策、文化概念等在英语或日语读者中未必熟悉的专名；中译英/中译日时通常保留原名词，并在括号内补一条目标语言简注

每个障碍记录：原文术语/段落 → 为何可能困惑 → 简明通俗解释（用作译注）。若是中译英/中译日中的中国特有名词，可直接记录为：原名词 → 为何陌生 → 目标语言简注。

Identify areas that may be difficult for target readers (calibrated based on target audience):

Domain Terminology: Lacks well-known translations, or literal translation is meaningless
Cultural Allusions: Idioms, historical events, pop culture unique to the source culture
Implicit Knowledge: Background that the original author assumes readers have but target readers may lack
Wordplay and Metaphors: Rhetoric that cannot be directly conveyed across languages
Named Concepts: Theories, effects with specific names (e.g., "Hair Combing Effect", "Dunning-Kruger Effect")
China-Specific Nouns: Platforms, apps, institutions, policies, cultural concepts that may be unfamiliar to English or Japanese readers; when translating from Chinese to English/Japanese, retain the original noun and add a brief target-language annotation in parentheses

Record each barrier as: Original term/paragraph → Why it may cause confusion → Concise, plain explanation (for annotations). For China-specific nouns in Chinese-to-English/Japanese translation, record directly as: Original noun → Why it is unfamiliar → Target-language brief annotation.

2.7 修辞与隐喻映射

2.7 Rhetoric and Metaphor Mapping

识别源文中所有隐喻、明喻、习语和修辞表达。逐一分析：

原文表达：原句
意图含义：作者真正想传达什么
直译风险：逐字翻译是否会不自然、丢失内涵或困惑读者？
目标语言策略：
- 意译：放弃源语言意象，用目标语言自然表达传达原意
- 替换：用目标语言中传达相同含义和情感效果的习语/意象替代
- 保留：原意象在目标语言中同样有效，保留

同时标记：

词语选择中的情感内涵：如"alarming"不仅是客观描述，还传达主观感受——记录需保留的情感效果
言外之意：表面简单但含义更丰富的句子——记录作者真正想说什么

Identify all metaphors, similes, idioms and rhetorical expressions in the source text. Analyze each one:

Original Expression: Original sentence
Intended Meaning: What does the author really want to convey?
Literal Translation Risks: Will word-for-word translation be unnatural, lose connotation or confuse readers?
Target Language Strategy:
- Free Translation: Abandon the source language imagery, use natural target-language expressions to convey the original meaning
- Replacement: Replace with idioms/imagery in the target language that convey the same meaning and emotional effect
- Retention: Keep the original imagery if it is equally effective in the target language

Also mark:

Emotional Connotations in Word Choice: For example, "alarming" is not just an objective description but conveys subjective feelings — record the emotional effect that needs to be preserved
Implied Meaning: Sentences that seem simple but have richer meanings — record what the author really means

2.8 结构与创意挑战

2.8 Structural and Creative Challenges

需要重组的复杂句式（长从句、嵌套修饰语、分词短语）
不可译的结构挑战（双关、歧义）
需要创意适配的作者声音或幽默

保存
01-analysis.md
，结构：

undefined

Complex sentence structures that need restructuring (long clauses, nested modifiers, participial phrases)
Untranslatable structural challenges (pun, ambiguity)
Author's voice or humor that requires creative adaptation

Save as
01-analysis.md
, with the following structure:

undefined

概要

Overview

[3-5 句]

[3-5 sentences]

核心内容

Core Content

核心论点：[一句话] 关键概念：[列表] 结构：[大纲]

Core Argument: [One sentence] Key Concepts: [List] Structure: [Outline]

背景语境

Context

作者：[谁，背景，立场] 写作语境：[回应什么] 目的：[目标和受众] 隐含假设：[未明说的前提]

Author: [Who, background, stance] Writing Context: [What it responds to] Purpose: [Goals and audience] Implicit Assumptions: [Unstated premises]

术语表

Glossary

[术语 → 译法, ...]

[Term → Translation, ...]

语气与风格

Tone and Style

[评估] [若目标语言为日语，额外记录：文体（敬体/常体）、选择依据、允许例外]

[Evaluation] [If target language is Japanese, additionally record: Style (Keigo/Plain Form), basis for selection, allowed exceptions]

读者理解障碍

Reader Comprehension Barriers

[术语/段落] → [为何困惑] → [建议译注]
...

[Term/Paragraph] → [Why it may cause confusion] → [Suggested annotation]
...

修辞与隐喻映射

Rhetoric and Metaphor Mapping

[原文表达] → [意图含义] → [策略：意译/替换/保留] → [建议译法]
...

[Original Expression] → [Intended Meaning] → [Strategy: Free Translation/Replacement/Retention] → [Suggested Translation]
...

结构与创意挑战

Structural and Creative Challenges

[句式重组需求、双关、创意适配需求]

undefined

[Sentence restructuring needs, puns, creative adaptation needs]

undefined

第三步：组装翻译提示 →

02-prompt.md

Step 3: Assemble Translation Prompt →

02-prompt.md

读取

01-analysis.md

，组装完整翻译提示。将风格指令、内容背景、合并术语表、理解障碍内联到提示中。

Read

01-analysis.md

and assemble a complete translation prompt. Integrate style instructions, content background, merged glossaries, and comprehension barriers into the prompt.

风格推导

Style Derivation

auto
模式（默认）：从
```
01-analysis.md
```
的"语气与风格"章节提取原文口吻特征，将其直接转化为翻译风格指令。不套用预设模板，而是用具体描述告诉译者"原文是什么口吻，翻译应保持什么口吻"。例如：原文分析为"轻松口语化，频繁使用第一人称和反问句"，则风格指令为"保持轻松口语化的语气，延续第一人称叙述和反问句式，译文应读起来像一篇自然的中文随笔而非正式文章"
显式指定风格（如
```
--style formal
```
）：使用对应预设描述，覆盖原文口吻
若目标语言为日语，无论
```
auto
```
还是显式风格，都必须把分析阶段选定的文体（敬体或常体）写入提示，并要求正文全文保持一致

提示模板见 references/subagent-prompt-template.md。

此提示用于 subagent（分块时）或主 agent 自身（不分块时）。

auto
mode (default): Extract the original tone features from the "Tone and Style" section of
```
01-analysis.md
```
, and directly convert them into translation style instructions. Do not apply preset templates, but use specific descriptions to tell the translator "what the original tone is, and what tone the translation should maintain". For example, if the original analysis is "Relaxed and colloquial, frequent use of first person and rhetorical questions", the style instruction would be "Maintain a relaxed and colloquial tone, continue using first-person narration and rhetorical questions, and the translation should read like a natural Chinese essay rather than a formal article"
Explicitly specified style (e.g.,
```
--style formal
```
): Use the corresponding preset description, overriding the original tone
If the target language is Japanese, regardless of
```
auto
```
or explicit style, the style (Keigo or Plain Form) selected in the analysis phase must be written into the prompt, and the translator must be required to maintain consistency throughout the main text

Prompt template can be found in references/subagent-prompt-template.md.

This prompt is used for subagents (when splitting into chunks) or the main agent itself (when not splitting into chunks).

第四步：初译 →

03-draft.md

Step 4: Initial Translation →

03-draft.md

长文本分块处理

Long Text Chunk Processing

估算字数。若超过 4000 词：

按 markdown 块边界拆分，每块不超过 5000 词
保持标题、段落、列表、代码块等结构完整
为每个 chunk 创建一个 subagent，全部并行执行
每个 subagent 读取
```
02-prompt.md
```
获取共享上下文，翻译其 chunk
所有 subagent 完成后，按顺序合并为
```
03-draft.md
```
接缝检查（合并后立即执行）：
- 可选译注去重：每个 subagent 独立工作，可能会对同一术语重复加注。合并后扫描全文，只保留真正有助于理解的首次注释；后续重复且不再提供新信息的注释删除
- 语气连贯性：检查 chunk 边界处前后段落的语气、文体是否一致，修正突兀的风格跳变
- 日文文体统一：若目标语言为日语，逐段检查 chunk 边界与全文句尾，确保敬体/常体一致；除引语、UI 固有文案、引用原文片段外不得切换
- 上下文指代：检查跨 chunk 边界的指代词（"前面提到的"、"上述"、"如前所述"等），确认指代目标在译文中确实存在且可被读者定位

Estimate word count. If over 4000 words:

Split along Markdown block boundaries, each chunk no more than 5000 words
Keep structures such as titles, paragraphs, lists, code blocks intact
Create a subagent for each chunk, execute all in parallel
Each subagent reads
```
02-prompt.md
```
to get shared context and translates its chunk
After all subagents complete, merge the chunks in order into
```
03-draft.md
```
Seam Check (execute immediately after merging):
- Optional Annotation Deduplication: Since each subagent works independently, it may add duplicate annotations for the same term. After merging, scan the full text and only retain the first annotation that truly helps with understanding; delete subsequent duplicate annotations that do not provide new information
- Tone Consistency: Check if the tone and style of paragraphs before and after chunk boundaries are consistent, and revise abrupt style changes
- Japanese Style Consistency: If the target language is Japanese, scan the chunk boundaries and sentence endings throughout the text to ensure consistency between Keigo/Plain Form and the style locked in the analysis phase; no switching is allowed except for quotations, inherent UI copy, or cited original fragments
- Contextual Reference: Check referential terms across chunk boundaries (e.g., "mentioned earlier", "above", "as mentioned before") to confirm that the referenced target exists in the translation and can be located by readers

短文本

Short Text

直接翻译，保存为

03-draft.md

。

Translate directly and save as

03-draft.md

翻译原则（所有翻译必须遵守）

Translation Principles (Must Be Followed for All Translations)

原文至上：翻译必须基于原文逐段逐句进行，严禁总结、概括、压缩或改写原文内容。不得省略句子或用自己的话重述。段落可以根据标点与分段优化规则重新拆分（允许比原文多分段），但不得合并原文的多个段落为一段。原文的所有标题必须保留，不得省略或合并。如果通过 URL 抓取的内容与原文有出入（被摘要化、截断、付费墙拦截或改写），必须直接停止并要求用户提供完整正文，不得基于残缺内容继续翻译
准确第一：事实、数据、逻辑必须与原文完全一致
达意而非逐字：翻译作者的意思，而非字面。当直译不自然或无法传达效果时，自由重组以地道目标语言表达相同含义
修辞处理：按意图含义理解隐喻、习语和修辞表达。当源语言意象在目标语言中不具备相同内涵时，用传达相同含义和情感效果的自然表达替代。参照第二步的修辞映射
情感保真：保留用词的情感内涵，而非仅字典释义。带有主观感受的词（如"alarming"、"haunting"）应在目标语言中唤起相同反应
自然流畅：使用地道的目标语言语序和句式；当源语言结构在目标语言中不自然时，自由断句或重组
日文文体锁定：若目标语言为日语，动笔前先确定敬体或常体。正文叙述、解释、过渡和结论统一使用同一套终止形；仅引语、UI 固有文案、引用原文片段等明确例外可保留不同语体
术语一致：优先使用标准译法；仅在首次出现且确实有助于目标读者理解、消歧或检索时，才在译文后括注原文
中国特有名词注释：当翻译成英文或日文时，若遇到目标读者可能不熟悉的中国特有平台、产品、机构或文化名词，优先保留原名词，并在首次出现时紧跟目标语言简注。格式：
```
原名词（target-language explanation）
```
。例如：
```
小红书（a Chinese lifestyle and social commerce platform）
```
、
```
小红书（中国のライフスタイル共有・ECプラットフォーム）
```
保留格式：保持所有 markdown 格式（标题、粗体、斜体、图片、链接、代码块）
Frontmatter 转换：若源文有 YAML frontmatter，保留并做以下变更：(1) 描述源文章的元数据字段加
```
source
```
前缀（camelCase）：
```
url
```
→
```
sourceUrl
```
、
```
title
```
→
```
sourceTitle
```
等；(2) 翻译文本字段值，作为新顶层字段添加；(3) 其他字段保持原样，适当翻译值
尊重原文：保持原意和意图；不添加、删除或评论——但句式和意象可自由适配以服务于含义
译注：仅在缺少注释会妨碍理解时，为术语、概念或文化典故添加简明解释。常规格式：
```
译文（原文，通俗解释）
```
；若保留原名词本身，使用
```
原名词（target-language explanation）
```
。根据目标受众校准注释深度：普通读者可适度加注，技术/学术/商务读者默认少注或不注。短文本（< 5 句）进一步减少注释，不过度教学化

Original Text First: Translation must be done paragraph by paragraph, sentence by sentence based on the original text. Strictly prohibited to summarize, generalize, compress or rewrite the original content. Do not omit sentences or rephrase in your own words. Paragraphs can be re-split according to punctuation and paragraph optimization rules (allowing more paragraphs than the original), but do not merge multiple original paragraphs into one. All titles in the original text must be retained, not omitted or merged. If the content fetched from the URL differs from the original (summarized, truncated, blocked by paywalls or rewritten), stop immediately and ask the user to provide the full text. Do not continue translation based on incomplete content
Accuracy First: Facts, data, and logic must be completely consistent with the original text
Convey Meaning, Not Literal Words: Translate the author's meaning, not the literal words. When literal translation is unnatural or cannot convey the intended effect, freely restructure to express the same meaning in authentic target language
Rhetorical Processing: Understand metaphors, idioms and rhetorical expressions based on their intended meaning. When the source language imagery does not have the same connotation in the target language, replace it with natural expressions that convey the same meaning and emotional effect. Refer to the rhetoric mapping in Step 2
Emotional Fidelity: Retain the emotional connotation of words, not just dictionary definitions. Words with subjective feelings (e.g., "alarming", "haunting") should evoke the same response in the target language
Natural Fluency: Use authentic target language word order and sentence structure; when the source language structure is unnatural in the target language, freely split sentences or restructure
Japanese Style Lock: If the target language is Japanese, determine Keigo or Plain Form before starting translation. The main text's narration, explanation, transitions and conclusions must use the same set of terminal forms; only explicit exceptions such as quotations, inherent UI copy, or cited original fragments can retain different styles
Terminology Consistency: Prioritize using standard translations; only add the original term in parentheses after the translation when it first appears and truly helps the target reader understand, disambiguate or retrieve
China-Specific Noun Annotations: When translating to English or Japanese, if encountering China-specific platforms, products, institutions or cultural concepts that may be unfamiliar to target readers, retain the original noun and add a brief target-language annotation immediately after its first occurrence. Format:
```
Original Noun (target-language explanation)
```
. For example:
```
小红书（a Chinese lifestyle and social commerce platform）
```
,
```
小红书（中国のライフスタイル共有・ECプラットフォーム）
```
Retain Formatting: Keep all Markdown formatting (titles, bold, italic, images, links, code blocks)
Frontmatter Conversion: If the source text has YAML frontmatter, retain it and make the following changes: (1) Add
```
source
```
prefix (camelCase) to metadata fields describing the source article:
```
url
```
→
```
sourceUrl
```
,
```
title
```
→
```
sourceTitle
```
, etc.; (2) Translate the values of text fields and add them as new top-level fields; (3) Keep other fields unchanged, translate values appropriately
Respect Original Text: Retain the original meaning and intent; do not add, delete or comment — but sentence structures and imagery can be freely adapted to serve the meaning
Annotations: Only add concise explanations for terms, concepts or cultural allusions when lack of annotations will hinder understanding. Regular format:
```
Translation (Original Term, plain explanation)
```
; if retaining the original noun itself, use
```
Original Noun (target-language explanation)
```
. Calibrate the depth of annotations based on the target audience: ordinary readers can have moderate annotations, while technical/academic/business readers have fewer or no annotations by default. Further reduce annotations for short texts (<5 sentences) to avoid excessive instructional tone

第五步：批判审校 →

04-critique.md

Step 5: Critical Review →

04-critique.md

主 agent 对照原文批判审校初译。此步只产出诊断，不改写。

The main agent conducts a critical review of the initial draft against the original text. This step only produces diagnoses, not rewrites.

5.1 准确性与完整性

5.1 Accuracy and Completeness

结构对比：统计原文与译文的标题数、段落数。若译文标题数少于原文，或段落数显著少于原文（差异超过 20%），标记为遗漏风险，逐一排查
逐段对照原文，逐句比较
核实所有事实、数字、日期、专有名词
标记任何意外添加、删除或篡改的内容
检查术语是否全文一致使用术语表
确认无遗漏段落或章节

Structure Comparison: Count the number of titles and paragraphs in the original text and translation. If the number of titles in the translation is less than the original, or the number of paragraphs is significantly less than the original (difference exceeds 20%), mark it as a missing risk and check each item one by one
Compare paragraph by paragraph, sentence by sentence against the original text
Verify all facts, figures, dates, proper nouns
Mark any unexpected additions, deletions or alterations
Check if terminology is used consistently throughout the text according to the glossary
Confirm no paragraphs or sections are missing

5.2 翻译腔诊断（按目标语言）

5.2 Translationese Diagnosis (By Target Language)

英文目标语言：

直译句法：保留中文/日文题述结构、话题先行结构或过度前置状语，导致英文不自然
冠词与限定词缺失：
```
a/an/the
```
、this/that、some/these 等使用不当，暴露源语言迁移痕迹
介词与搭配生硬：逐字映射中文/日文搭配，形成非地道英文表达
名词化过重：本可用动词推进的句子被写成僵硬抽象名词链
指代不清：中文省略主语或日文省略照应直接搬到英文，导致 referent 模糊
语域不稳：在自然英文里应简洁处写得过分书面，或在正式文本中过度口语化
信息流不顺：句子旧信息/新信息顺序不符合英文读者预期，读起来别扭

中文目标语言：

多余连接词：过度使用因此/然而/此外/另外，上下文已暗示关系时不需要
被动语态滥用：过多被/由/受到，主动语态更自然时
名词堆砌：长修饰链应拆为短句
分裂句：从英语"It is...that"硬搬的不自然"是...的"结构
过度名词化：动词或形容词更自然时使用抽象名词（如"进行了讨论"→"讨论了"）
代词冗余：可省略时过度使用他/她/它/我们/你

日文目标语言：

敬体/常体混用：正文叙述在
```
です・ます
```
与
```
だ・である
```
之间来回切换，破坏专业感与连贯性
终止形漂移：小标题后的首句、补充说明、总结句或 chunk 接缝处忽然换成另一套句尾
片假名过多：可用和语或汉语表达的概念不必要地使用片假名外来语
主语过剩：日文常省略主语，过度保留英文的主语结构会不自然
被动形滥用：英文被动语态直译为「〜される」，主动表达更自然时
名词修饰语连锁：英文的长前置修饰直译不自然，应拆分为后置修饰或独立句
不自然的连接词：「しかしながら」「それゆえ」等硬连接词多用，上下文已明确关系时应省略
翻译调语序：保持英文语序的不自然日语，应重组为自然语序

English Target Language:

Literal Syntax: Retain Chinese/Japanese topic-comment structure, topic-first structure or overly preposed adverbials, resulting in unnatural English
Missing Articles and Determiners: Improper use of
```
a/an/the
```
, this/that, some/these, etc., revealing traces of source language transfer
Stiff Prepositions and Collocations: Word-for-word mapping of Chinese/Japanese collocations, forming non-authentic English expressions
Excessive Nominalization: Sentences that could be advanced with verbs are written as rigid chains of abstract nouns
Unclear Reference: Chinese omitted subjects or Japanese omitted references directly transferred to English, resulting in ambiguous referents
Unstable Register: Overly formal writing where natural English should be concise, or overly colloquial in formal texts
Poor Information Flow: The order of old/new information in sentences does not meet the expectations of English readers, making it awkward to read

Chinese Target Language:

Redundant Conjunctions: Excessive use of 因此/然而/此外/另外 when the contextual relationship is already implied
Overuse of Passive Voice: Excessive use of 被/由/受到 when active voice is more natural
Noun Stacking: Long modifier chains should be split into short sentences
Cleft Sentences: Unnatural "是...的" structures directly copied from English "It is...that"
Excessive Nominalization: Use of abstract nouns when verbs or adjectives are more natural (e.g., "进行了讨论"→"讨论了")
Redundant Pronouns: Overuse of 他/她/它/我们/你 when they can be omitted

Japanese Target Language:

Mixed Keigo and Plain Form: Switching between
```
です・ます
```
and
```
だ・である
```
in main text narration, damaging professionalism and coherence
Terminal Form Drift: Sudden switch to another set of sentence endings in the first sentence after a subheading, supplementary explanation, summary sentence or chunk boundary
Excessive Katakana: Unnecessary use of Katakana loanwords for concepts that can be expressed in Wago or Kango
Excessive Subjects: Japanese often omits subjects; overly retaining English subject structures will be unnatural
Overuse of Passive Voice: Direct translation of English passive voice to 「〜される」 when active expression is more natural
Long Noun Modifier Chains: Literal translation of English long premodifiers is unnatural; should be split into postmodifiers or independent sentences
Unnatural Conjunctions: Overuse of hard conjunctions like 「しかしながら」「それゆえ」 when the contextual relationship is already clear and should be omitted
Translationese Word Order: Unnatural Japanese that retains English word order; should be restructured into natural word order

5.3 修辞与情感保真

5.3 Rhetoric and Emotional Fidelity

对照
```
01-analysis.md
```
的隐喻映射：所有标记的隐喻/习语是否按建议策略（意译/替换/保留）处理？
标记直译后不自然或丢失原意的隐喻或修辞表达
检查情感内涵：源文中带主观感受的词是否在译文中唤起相同反应，还是被扁平化为中性/客观描述？
标记丢失的言外之意：因译者过于贴近表面含义而未传达作者深层意图的句子

Compare with the metaphor mapping in
```
01-analysis.md
```
: Are all marked metaphors/idioms handled according to the suggested strategy (free translation/replacement/retention)?
Mark metaphors or rhetorical expressions that are unnatural or lose original meaning after literal translation
Check emotional connotation: Do words with subjective feelings in the source text evoke the same response in the translation, or are they flattened into neutral/objective descriptions?
Mark lost implied meaning: Sentences where the translator was too close to the surface meaning and failed to convey the author's deep intent

5.4 策略执行

5.4 Strategy Execution

```
02-prompt.md
```
中的翻译策略是否被实际遵循？
分析中识别的语气和语域是否被应用？
```
01-analysis.md
```
中的理解障碍是否以适当译注处理？
术语是否一致使用？

Was the translation strategy in
```
02-prompt.md
```
actually followed?
Was the tone and register identified in the analysis applied?
Were the comprehension barriers identified in
```
01-analysis.md
```
handled with appropriate annotations?
Is terminology used consistently?

5.5 表达与逻辑

5.5 Expression and Logic

标记读起来像"翻译腔"的句子——不自然的语序、硬搬、僵硬措辞
检查句间和段间逻辑流
识别重组可提升可读性的地方
指出遗漏目标语言习惯表达的地方

Mark sentences that read like "translationese" — unnatural word order, forced copying, stiff wording
Check logical flow between sentences and paragraphs
Identify areas where restructuring can improve readability
Point out places where target language customary expressions are missing

5.6 译注质量

5.6 Annotation Quality

注释是否准确、简洁、确实有帮助？
识别遗漏的需要注释的理解障碍
检查中译英/中译日中的中国特有名词是否在首次出现时得到足够简明的目标语言解释
标记对目标受众显而易见的术语的过度注释
检查文化典故是否在需要时得到解释

Are annotations accurate, concise and truly helpful?
Identify missing annotations for comprehension barriers
Check if China-specific nouns in Chinese-to-English/Japanese translation have sufficient brief target-language explanations when first mentioned
Mark over-annotations for terms that are obvious to the target audience
Check if cultural allusions are explained when needed

5.7 文化适配

5.7 Cultural Adaptation

隐喻和习语在目标语言中是否有效？
是否有在目标文化中可能困惑或冒犯的引用？
是否有因文化语境差异可能被误解的段落？

保存
04-critique.md
，结构：

undefined

Are metaphors and idioms effective in the target language?
Are there references that may confuse or offend the target culture?
Are there paragraphs that may be misunderstood due to cultural context differences?

Save as
04-critique.md
, with the following structure:

undefined

准确性与完整性

Accuracy and Completeness

[问题]：[位置] — [描述]

[Issue]: [Location] — [Description]

翻译腔问题

Translationese Issues

[问题类型]：[初译示例] → [建议修正]

[Issue Type]: [Initial Draft Example] → [Suggested Revision]

修辞与情感保真

Rhetoric and Emotional Fidelity

[直译隐喻]：[原文] → [初译] → [建议意译]
[扁平化情感]：[原文词/短语] → [初译] → [如何恢复情感效果]

[Literal Metaphor]: [Original Text] → [Initial Draft] → [Suggested Free Translation]
[Flattened Emotion]: [Original Word/Phrase] → [Initial Draft] → [How to Restore Emotional Effect]

策略执行

Strategy Execution

[策略]：[遵循/遗漏] — [详情]

[Strategy]: [Followed/Missed] — [Details]

表达与逻辑

Expression and Logic

[位置]：[问题] → [建议]

[Location]: [Issue] → [Suggestion]

译注

Annotations

[添加/删除/修改]：[术语] — [原因]

[Add/Delete/Modify]: [Term] — [Reason]

文化适配

Cultural Adaptation

[问题]：[描述] — [建议]

[Issue]: [Description] — [Suggestion]

总结

Summary

[整体评估：X 个关键问题，Y 个改进，Z 个小建议]

undefined

[Overall Evaluation: X key issues, Y improvements, Z minor suggestions]

undefined

第六步：终稿

Step 6: Final Draft

根据

04-critique.md

的所有发现，对

03-draft.md

进行修订与润色，一步到位产出出版级译文。读取初译和审校发现，也可回溯源文和

01-analysis.md

：

修正所有审校发现的准确性问题
将翻译腔表达改写为自然的目标语言模式
按隐喻映射重新意译直译的隐喻和修辞表达
恢复被扁平化的情感内涵
应用遗漏的翻译策略
重组僵硬或别扭的句子以提升流畅度
按审校建议添加、删除或修改译注
改善段落间过渡
按建议适配文化典故
若目标语言为日文，逐段扫描句尾与判断句，确认敬体/常体与分析阶段锁定的文体一致，修正局部漂移
通读全文——作为独立作品是否流畅自然？
确保叙事声音全文一致，一致应用所选翻译风格
最终扫描：任何仍读起来像"翻译"而非"写作"的表达都应重铸为自然的目标语言表达
全文术语一致性最终检查
验证格式正确保留（标题、粗体、链接、代码块）
段落与节奏：
- 以目标语言的自然行文为准，不机械照搬源文句法，也不为了形式要求强行拆成单句段
- 仅在可读性显著提升时拆段或断句；每段应服务一个清晰层次或动作
- 中文目标语言使用规范全角中文标点，日文目标语言使用标准日文标点，英文目标语言使用标准英文标点

Revise and polish

03-draft.md

based on all findings in

04-critique.md

, producing a publication-level translation in one step. Read the initial draft and review findings, and can also refer back to the original text and

01-analysis.md

Correct all accuracy issues identified in the review
Rewrite translationese expressions into natural target-language patterns
Re-translate literal metaphors and rhetorical expressions according to metaphor mapping
Restore flattened emotional connotations
Apply missed translation strategies
Restructure stiff or awkward sentences to improve fluency
Add, delete or modify annotations according to review suggestions
Improve transitions between paragraphs
Adapt cultural allusions according to suggestions
If the target language is Japanese, scan sentence endings and judgment sentences paragraph by paragraph to confirm that Keigo/Plain Form is consistent with the style locked in the analysis phase, and correct local drift
Read through the full text — does it flow naturally as an independent work?
Ensure consistent narrative voice throughout the text, and consistently apply the selected translation style
Final scan: Any expression that still reads like a "translation" rather than "writing" should be recast into natural target-language expression
Final check for terminology consistency throughout the text
Verify that formatting is correctly retained (titles, bold, links, code blocks)
Paragraphs and Rhythm:
- Follow natural target-language writing standards, do not mechanically copy source text syntax, nor forcibly split into single-sentence paragraphs for formal requirements
- Split paragraphs or sentences only when readability is significantly improved; each paragraph should serve a clear level or action
- Use standard full-width Chinese punctuation for Chinese target language, standard Japanese punctuation for Japanese target language, and standard English punctuation for English target language

第七步：输出

Step 7: Output

默认输出

Default Output

默认直接输出最终译文正文，不额外添加"原文："头部、导读或编者说明
仅当用户明确要求发布版包装（如"加原文链接"、"写一段导读"、"做成译介稿"）时，才在正文前附加这些信息

By default, directly output the final translation body without additional headers like "Original Text:", introductions or editor's notes
Only add such information before the main text when the user explicitly requests publication packaging (e.g., "Add original link", "Write an introduction", "Make a translation feature article")

输出格式

Output Format

根据译文内容自动选择输出格式：

含 markdown 格式的内容（标题、粗体、链接、代码块、列表等）：用 markdown 代码块包裹输出，让用户可以直接复制源码：

```markdown
{译文内容}
```

纯文本内容（无 markdown 格式标记）：直接输出，无需代码块包裹

判断依据：译文中是否包含

、

**

、

[]()

、

等 markdown 语法。只要存在任何 markdown 格式标记，就用代码块包裹。

Automatically select output format based on translation content:

Content with Markdown formatting (titles, bold, links, code blocks, lists, etc.): Wrap the output in a Markdown code block so users can directly copy the source code:

```markdown
{Translation Content}
```

Plain text content (no Markdown formatting markers): Directly output without wrapping in code blocks

Judgment basis: Whether the translation contains Markdown syntax such as

**

[]()

, etc. As long as any Markdown formatting marker exists, wrap it in a code block.

默认行为：直接输出

Default Behavior: Direct Output

默认将最终译文输出到对话中（不写文件）。翻译过程中产生的所有中间文件（

01-analysis.md

、

02-prompt.md

、

03-draft.md

、

04-critique.md

、

chunks/

目录）在输出前全部删除，临时工作目录也一并删除。

By default, output the final translation directly in the conversation (do not write to file). All intermediate files generated during translation (

01-analysis.md

02-prompt.md

03-draft.md

04-critique.md

chunks/

directory) are deleted before output, and the temporary working directory is also deleted.

写入文件模式

File Writing Mode

仅当用户明确要求写入文件时（如"保存到文件"、"写成文件"、"输出到文件"），才将最终译文保存为

translation.md

。此时同样删除所有中间文件，输出目录中只保留

translation.md

。

Only save the final translation as

translation.md

when the user explicitly requests writing to file (e.g., "Save to file", "Write to file", "Output to file"). In this case, all intermediate files are also deleted, and only

translation.md

is retained in the output directory.

图片语言检查

Image Language Check

翻译完成后，做轻量级图片语言检查：

收集译文中的图片引用
识别可能含大量文字的图片（封面、截图、图表、框架图、信息图）
若图片主要文字语言与译文语言不匹配，主动提醒用户
提醒仅列出清单，不自动本地化图片

提醒格式：

可能需要图片本地化：
- ![示例封面](attachments/example-cover.png)：可能仍含源语言文字
- ![示例图表](attachments/example-diagram.png)：文字密集的框架图，检查标签是否需要翻译

After translation is completed, perform a lightweight image language check:

Collect image references in the translation
Identify images that may contain a lot of text (covers, screenshots, charts, framework diagrams, infographics)
If the main text language of the image does not match the translation language, actively remind the user
The reminder only lists the inventory, does not automatically localize images

Reminder format:

Image localization may be needed:
- ![Example Cover](attachments/example-cover.png): May still contain source language text
- ![Example Diagram](attachments/example-diagram.png): Text-dense framework diagram, check if labels need translation

摘要

Summary

**翻译完成**（精翻模式）

源文件：{source-path}
语言：{from} → {to}
应用术语：{count} 条

若写入文件模式，摘要中额外显示：

输出文件：{output-dir}/translation.md

若发现图片语言不匹配，在摘要后附上提醒清单。

**Translation Completed** (Professional Mode)

Source File: {source-path}
Language: {from} → {to}
Terms Applied: {count}

If in file writing mode, additionally display in the summary:

Output File: {output-dir}/translation.md

If image language mismatch is found, attach the reminder list after the summary.