article-rewriter

Article Rewriter

An article processing skill oriented toward "translation + rewriting + publishing".

It is not a pure translator, nor an editor that only polishes wording. When encountering foreign-language source text, mixed-language source text, the need to restructure, the need to adapt to WeChat Official Account/Zhihu/Newsletter, or when the user requests finished-manuscript-level output, it must separate "understanding the source text" and "producing the finished manuscript".

Three modes:

• quick: Lightweight translation or lightweight rewriting, does not proactively search externally, and keeps intermediate files to a minimum
• standard: Analyzes the source text, supplements a small amount of material when necessary, and goes through one round of explicit drafting and critical revision
• publish: Analysis, verification, restructuring, criticism, revision, polishing, expression optimization, quality check, producing a directly publishable version

Default mode:standard

Input type

• inline text: The user directly pastes the body text
• file path: Reads a local file
• URL: Fetches the body text before processing (prefer WebReader; if it fails, try using https://r.jina.ai/{url} to read; if it still fails, try using https://defuddle.md/{url} ; if all fail, clearly state that the content cannot be obtained)

Default values

Setting	Default value	Description
Output language	zh-CN	Final output is uniformly in Chinese
Mode	standard	Used when the user does not specify
Platform	generic	If the user mentions WeChat Official Account/Zhihu/Newsletter, switch to the corresponding preset
Number of titles	5	Provide 5 title candidates when the user does not specify
Research scope	2-4 topics	publish mode can be expanded to 4-6 items
Output directory	rewrites/{slug}-{timestamp}/	All intermediate files and the final manuscript are placed here uniformly
Long-text threshold	> 2500 English words or > 4000 Chinese characters	After exceeding this, prioritize chunked writing and unified review

Mode

Mode	Trigger words	Core steps	Applicable scenarios
Quick	"Simple polishing", "quick edit", "do not look up materials"	Constraint confirmation → terminology unification (if needed) → direct finished manuscript	Short texts, lightweight polishing, user explicitly does not want external lookup
Standard	Default	Analysis → selective research → Prompt snapshot → draft → criticism → revision → expression optimization → finished manuscript	Most article rewriting needs, especially "translate and rewrite"
Publish	"Deep rewrite", "WeChat Official Account finished manuscript", "suitable for publication", "supplement materials and rewrite"	Analysis → verification → restructuring → Prompt snapshot → draft → criticism → revision → polishing → expression optimization → quality check → finished manuscript	Tasks to be published, fact-sensitive, with high style requirements, or long-form tasks

Automatic judgment rules:

• When the user explicitly says "no need to look up materials" or "only change the expression", use quick
• When the user explicitly asks to "post to WeChat Official Account/Zhihu/Newsletter," "deep rewrite," "supplement materials," or "make it like a formal article," use publish
• As long as the task contains obvious translation difficulties, long-form content, or the user wants "publication-grade Chinese," use at least standard
• In all other cases, default to standard

Output directory

Create an output directory for each task:rewrites/{slug}-{timestamp}/

File	When needed	Description
source.md	All	Raw content after being written to disk
00-brief.md	All	Task constraints, platform, audience, length, retrieval boundaries, whether restructuring is allowed
01-analysis.md	standard / publish	Theme, structure, audience, items to retain, difficulties, comprehension obstacles, rewriting strategy
01-terms.md	When there is translation or dense terminology	Terminology list for this task, recommended translations, prohibited translations, translations pending confirmation
02-research.md	When external research was done	External materials, fact verification, conflicting information, citable key points
03-outline.md	publish or when restructuring is obvious	New structure, outline, title angles, narrative order
04-prompt.md	Default output for non-trivial tasks	Writing instruction snapshot, for reuse in drafts, chunking, and subsequent revisions
05-draft.md	standard / publish	First draft; with translator notes if necessary
06-critique.md	standard / publish	Diagnose problems, do not directly revise the text
07-revision.md	standard / publish	Revised draft after incorporating criticism
08-polish.md	publish	Polished draft after sinicization, rhythm, title, and platform adaptation
09-expression.md	standard / publish non-trivial tasks	Expression optimization draft after removing translationese, AI tone, clichés, and mechanical transitions
qa.md	publish or fact-sensitive tasks	Final quality check checklist and residual risks
rewrite.md	All	Final deliverable draft

If the user only asks for the result to be returned directly in chat, still save by default to rewrite.md, and provide the path in the reply.

Platform and writing style

Do not hard-code the platform, paragraph rhythm, or title style in the main file. Read as needed:

• references/style-presets.md: Writing style and title constraints for WeChat Official Account, Zhihu, Newsletter, and general versions
• references/glossary-en-zh.md: Special translation methods in English source text or mixed Chinese-English materials
• references/quality-rubric.md: Fixed inspection dimensions for criticism drafts and final quality checks
• references/expression-optimizer.md: Translationese, AI tone, and cliché patterns to clean up during final expression optimization

Selection logic:

• When the user explicitly specifies a platform, prioritize the corresponding preset
• When the user does not specify a platform, first judge based on the source text type and target audience; if still unclear, use generic
• Technology, business, and social issues can all have a narrative feel, but do not forcibly make them "hyped up" or "viral"

Terminology handling

When a task contains any of the following, terminology must be unified first before entering rewriting:

• The source text is in English
• The source text is mixed Chinese-English
• The source text involves high-frequency proper nouns in AI, products, business, technology, etc.
• The user explicitly requests "translate and rewrite"

Processing order:

1. Prioritize reading the built-in terminology list references/glossary-en-zh.md
2. Additionally extract proper nouns, abbreviations, product names, and methodology terms from the source text
3. Form the session terminology list for this task and write it to 01-terms.md
4. Mark recommended translations, prohibited translations, terms to keep in the original language, and pending items
5. Subsequent drafts, critiques, and revisions shall all be based on 01-terms.md and must not drift arbitrarily

Execution principles:

• For common and conventional terms, use standard Chinese directly
• For terms that are easily mistranslated or vary greatly in style, prioritize the built-in glossary
• On first occurrence, you may use 中文（English） or 中文（英文原词，简短解释）
• If a term has multiple reasonable translations, choose the one best suited to the target platform and audience, and keep it consistent throughout
• Rewriting may reorganize sentences, but do not "lose" key terms or turn them into overly colloquial expressions

Workflow

Step 1: Obtain constraints and write them to 00-brief.md

Prioritize extracting from the current conversation; do not repeatedly ask for defaults. At minimum confirm:

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

If the user has not specified:

• By default, allow "supplementing a small amount of material when necessary," but do not over-search for static content
• By default, allow restructuring, but preserve the original core arguments
• By default, provide 5 title candidates

00-brief.md Suggested record:

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

Step 2: Save the original text to disk and create the output directory

1. Save the input content uniformly to source.md
2. Create the output directory rewrites/{slug}-{timestamp}/
3. Detect the original language
4. If the original text is not Chinese, first load the built-in glossary and extract session terms, and write them to 01-terms.md
5. If the original text is very long, first determine chunk boundaries and the merge strategy, then proceed to drafting

Requirements:

• For URL input, prioritize extracting the main text and avoid contamination from navigation, recommendations, and comment sections
• If file or URL extraction fails, clearly state the blocking point; if you can still continue, fall back to the body text already provided by the user
• When chunking, split by semantic boundaries; do not mechanically divide by word count

Step 3: Analyze the original text and write it to 01-analysis.md

standard and publish mode must produce 01-analysis.md. At minimum include:

• Original topic, core arguments, author intent
• Inferred target audience
• Breakdown of the original structure
• Facts, numbers, cases, and quotations that must be preserved
• Key terms and suggested translations
• Reader comprehension obstacles: background knowledge gaps, pronoun references, long sentences, skipped reasoning
• Metaphors and cultural mapping: which metaphors should be preserved, transcreated, explained, or downgraded
• Tone target: restrained / narrative / commentary / methodology / publication draft
• Optimization points: too long, repetitive, jumpy, vague, outdated facts, conclusion too early or too late
• Rewriting strategy: conservative rewrite / structural reorganization / secondary creation after translation

If the original text is already well written, the analysis must clearly state "focus on incremental and readability optimization"; do not forcibly overturn it.

Step 4: Research and fact-checking

When it is necessary to do 02-research.md：

• The user explicitly requests supplementary material, the latest information, cases, or data
• The original text contains clearly time-sensitive content
• The numbers, research, or company information in the original text may already be outdated
• The topic itself requires the most basic fact-checking

Read as needed:

• references/research-rules.md: retrieval scope, source priority, conflict handling, citation format

Execution principles:

• quick mode does not proactively search externally by default
• When the user explicitly prohibits retrieval, skip external search, and note in the final result: "This revision is based on the original text, with no external materials added"
• standard mode only looks up key information that can significantly improve the article's credibility
• publish mode must separate "external facts," "the author's views," and "your inferences"
• Research files only record "facts and sources"; do not mix language issues into them

Step 5: Plan the new structure

In the following cases, you must write 03-outline.md：

• publish mode
• The original structure is clearly chaotic
• The user requests platform adaptation and the style changes significantly
• It needs to shift from translation to a "finished draft suitable for Chinese publication"

At minimum include:

• Title direction: insight-driven / benefit-driven / counterintuitive / scenario-based / question-based
• Opening strategy: story, scene, contrast, question, conclusion first
• Order of body paragraphs
• The purpose each paragraph must accomplish
• Closing action: summary, reminder, method, next-step suggestion

standard mode may skip saving a separate outline, but the structure must be determined in your mind before writing.

Step 6: Assemble the Prompt snapshot and write it to 04-prompt.md

Non-trivial tasks produce by default 04-prompt.md. It is not the final draft for the user to read, but a unified instruction for subsequent drafting, chunking, and revision reuse.

At minimum state clearly:

• Task objective
• Output language and platform
• Target audience
• Information that must be retained
• Degree of restructuring allowed
• Terminology constraints
• Prohibited items
• Research conclusions or time boundaries
• Style requirements
• Title requirements

If the task needs chunking:

• Every chunk reuses the same 04-prompt.md
• Each chunk only handles its own body text and does not invent global conclusions on its own
• After merging, a full overall critique and unified revision must be done again

Step 7: Generate 05-draft.md

05-draft.md is a first draft; it does not aim to be finished in one pass, but it must first firmly establish the original text's skeleton and information.

Execution principles:

• When the original text is in a foreign language, the first draft should first ensure complete information, accurate relationships, and stable terminology
• Brief translator notes may be included, specifically recording difficulties, ambiguities, and tradeoffs
• Do not force cleverness or pile on rhetoric at the first-draft stage
• If it is a chunked task, first obtain first drafts of all chunks, then merge them into a unified 05-draft.md

Step 8: Generate 06-critique.md

06-critique.md only performs diagnosis and does not directly revise the text. Prioritize reading references/quality-rubric.md。

Check item by item at least from the following dimensions:

• Accuracy: whether there is mistranslation, omission, over-translation, or misplacement of relationships
• Completeness: whether key qualifiers, facts, and examples are lost
• Natural Chinese: whether there is Europeanized syntax, hard translation, or awkward collocations
• Terminology: whether terminology is consistent and whether the first explanation is reasonable
• Structure: whether paragraph order and the causal chain are clear
• Register / Platform Fit: whether it fits WeChat official accounts, Zhihu, Newsletter, or a formal style
• Fact Boundary: Whether inference and facts are confused, and whether time-sensitive information is handled appropriately
• Title / Framing: Whether the title, opening, and body promises are consistent

Writing requirements:

• Point out the problems and impact item by item
• Prioritize saying "what is wrong, why it is wrong, and where it should be changed"
• Do not casually rewrite the entire text in a critique draft

Step 9: Generate 07-revision.md

Address 06-critique.md 's issues one by one to obtain a revised draft with more stable structure and expression.

Execution principles:

• First fix accuracy, completeness, and logic, then fix language and rhythm
• If there are conflicting suggestions in the critique, prioritize preserving facts and clarity
• When revising, do not introduce new facts unless these facts have already been verified in 02-research.md Verify in

Step 10: Generate 08-polish.md

publish mode must be polished,standard mode should also be done when the user requests "like a finished draft" or "ready to publish."

Polishing only addresses the following issues:

• Change translationese into natural Chinese
• Smooth out paragraph rhythm and transitions
• Adjust the grip and restraint of the title, opening, and ending
• Clean up repetitive expressions, vague rhetoric, and over-explanation
• Unify subheadings, punctuation, formatting, and text density

During the polishing stage, do not:

• Reinvent viewpoints
• Introduce unverified facts
• Delete key qualifiers in pursuit of "smoothness"

Step 11: Generate 09-expression.md

After the structure, facts, and rhythm are all stable, do a separate round of "expression optimization." Prioritize reading references/expression-optimizer.md。

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

• Remove translationese, AI tone, and promotional tone
• Delete vague elevation, mechanical connectives, and formulaic three-part structures
• Change vague attribution into specific attribution, or into conservative expression
• Tighten excessive rhetoric, slogan-like sentences, and sentences that read like copied quotes
• Make paragraphs read more like Chinese written by a real author, rather than "a model that writes very well"

Focus on checking for:

• Excessive emphasis on significance, historical status, and grand trends
• Connective crutches such as "in addition / it is worth noting that / moreover / in the final analysis"
• Negative-style parallel structures such as "not only... but also..." and "this is not only... but even more..."
• Stiff three-item parallelism and a false sense of completeness
• Vague attribution, such as "experts believe," "the industry generally believes," and "some views point out"
• Overuse of dashes, bold text, abstract nouns, and sloganized endings
• Frequent synonym swapping to appear non-repetitive, causing terminology or subject drift

Execution principles:

• Preserve the original meaning; do not add viewpoints without authorization in the name of "human touch"
• The more formal the platform, the more restraint is needed; do not forcibly turn a formal style into a social media tone
• Allow natural rhythm changes, but do not deliberately "perform" a sense of looseness
• If the original text is already a highly restrained technical or formal style, the optimization goal is to remove mechanicalness, not to add emotion

Step 12: Quality check and save the final draft

publish mode and fact-sensitive tasks must write qa.md, including at least:

• Whether there are omissions, mistranslations, or over-translations
• Whether there are unverified new facts
• Whether terminology is consistent throughout
• Whether the title accurately reflects the body text
• Whether the most important facts, figures, cases, and quotations from the original text have been retained
• Whether there are still places that need to be expressed conservatively
• 09-expression.md Whether only the wording was changed, without sneakily altering facts and judgments

The final draft is always saved to rewrite.md。

If 09-expression.md：

• rewrite.md based on 09-expression.md

If expression optimization was not done separately, but 08-polish.md：

• rewrite.md based on 08-polish.md

If polishing was not done separately:

• rewrite.md based on 07-revision.md

Step 13: Generate title candidates

By default, provide 5 titles, prioritizing coverage of different angles rather than just swapping a few words:

• Insight type
• Benefit type
• Counterintuitive type
• Scenario type
• Question type

Title rules:

• Must accurately reflect the core value of the body text
• Do not fabricate numbers, do not exaggerate conclusions, do not create nonexistent urgency
• The more formal the platform, the more restrained the title
• The angles among titles in the same set must be clearly different

If the user explicitly says "no title needed," then skip.

Long-text and chunking strategy

Do not directly force-translate long texts all the way through. Use the approach of "unified constraints, local chunking, overall review."

Execution order:

1. First complete 00-brief.md、01-analysis.md、01-terms.md、04-prompt.md
2. Chunk by semantic boundaries, not mechanically by character count
3. Each chunk shares the same terminology list and Prompt snapshot
4. Merge all chunk drafts into 05-draft.md
5. Perform a unified 06-critique.md
6. After unified revision, then decide whether to proceed to 08-polish.md
7. Regardless of whether polishing is done separately, the merged draft must ultimately undergo one round of expression optimization before outputting rewrite.md

Special note:

• Chunking only solves the context window problem; it does not mean each chunk can independently set the tone
• All global judgments, titles, openings, and endings should be handled uniformly after merging
• The merged draft must be checked for terminology drift, repeated explanations, and inconsistent wording before and after
• Expression optimization must be done on the full merged text; each chunk cannot be separately "humanized"

Final output format

rewrite.md By default, use the following structure:

## 标题候选

1. 标题一
2. 标题二
3. 标题三
4. 标题四
5. 标题五

---

## 正文

[改写后的完整文章]

---

## 参考来源

- [来源标题](链接)
- [来源标题](链接)

---

## 改写说明

- 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

If no external search was performed, change "Reference Sources" to:

## 参考来源

- 本次基于用户提供原文改写，未补充外部资料

Quality threshold

• Information density should be higher than the original text, not merely a rewording
• If expanding, the added content must either explain key concepts or supplement cases/data/boundaries
• Fact-sensitive content should be verified first; when uncertain, it is better to express it conservatively
• The style should feel human, but do not engage in performative sentimentality
• If the original text is already of high quality, "fine polishing" should replace "rewriting"
• For tasks involving translation, at least one explicit critique and revision must be done
• For long-text or chunked tasks, a unified review must be performed again on the full merged text
• Before final delivery of nontrivial tasks, one more independent expression optimization must be done

Prohibited items

• Do not package unverified information as "the latest research shows"
• Do not distort conclusions just to create a sense of virality
• Do not copy large passages verbatim and only replace words
• Do not pile up vague rhetoric, clichés, or excessive exclamatory sentences
• Do not assume all articles are suited to "viral headlines"
• Do not in 06-critique.md casually rewrite the entire text, confusing diagnosis with revision
• Do not let 08-polish.md become yet another round of unbounded rewriting
• Do not let 09-expression.md secretly alter facts, stance, or information density under the guise of "optimizing expression"