writing-dna-skill

写作蒸馏器.skill

中文 | English

英文名：writing-dna-skill

我们的宗旨是：快乐写作。

所以做了这个可以蒸馏任意账号/作者写作风格的 agent skill，让写作不再痛苦，不再从 0 开始难产。

它并不是简单的“模仿某某风格”提示词，整个 skill 包含了一套完整的蒸馏流程，让 agent 把一批完整文章拆解成可复用的 Writing DNA，包括语言、结构、选题、素材、认知框架与视觉风格。蒸馏完成后，agent 可以在写作前重新读取这些风格文档，从而更稳定地复刻目标写作风格。

该 skill 在 moxt.ai 制作。

在 MoxtHub 打开写作蒸馏器.skill →

它解决什么问题

很多人让 AI 模仿一个作者，通常只会给几篇例文，然后说“照这个风格写”。这种方式很容易失真，因为 agent 只是在凭感觉抓表面语气。

writing-dna-skill 的思路是：不要让 agent 直接模仿，而是先让它系统分析一批完整文章，把风格拆成可复用的规则，再用这些规则指导后续写作。

它会把风格拆成：

• 这个作者常用什么词、句式、标点和节奏
• 文章通常如何开头、展开、转折和收束
• 什么话题会写，什么话题不写
• 常引用什么素材，如何使用案例、数据和截图
• 背后的世界观、价值判断和核心假设
• 如果是图文内容，还会分析配图、排版、字体层级和色彩使用

适合什么

• 分析某个账号/作者的稳定写作风格
• 沉淀自己的写作风格资产
• 对比不同作者在同一议题上的表达差异
• 给 agent/AI 提供更稳定的风格复刻上下文

中英文产物

这个 skill 会将“对话语言”和“产物语言”分开处理。例如，用户可以用中文下达任务，但为英文 Newsletter 生成英文 Writing DNA。

• 中文产物使用 templates/author-corpus/zh/ 和中文文件名。
• 英文产物使用 templates/author-corpus/en/ 和英文文件名。
• 用户明确指定时以用户选择为准，否则跟随语料的主要语言。

它能分析什么

这个 skill 会尽量把“风格”从一个模糊感受，拆成可以被记录、复用和检查的规则。它不只看文字表面的语气，也会看文章为什么这样写、如何组织材料、如何形成判断。

它适合分析：

• 长文、公众号文章、博客、Newsletter、专栏文章
• 小红书、微博、社交平台上的图文内容
• 自己过去写过的一批文章，用来沉淀个人写作资产
• 同一主题下多个作者的文章，用来对比表达差异

如果原文里有截图、配图、数据图或排版样式，也建议一起保留。很多图文账号的风格不只在文字里，图片的出现位置、截图承担的证据功能、标题层级和留白节奏，都会影响最终“像不像”。

蒸馏的六个层次

层次	分析对象	输出结果
L1 表层语言	词频、句长、标点、修辞、常用表达	语言DNA.md
L2 文章结构	开头 hook、正文架构、转折方式、结尾收束	文章结构模板.md
L3 选题逻辑	发布时机、切入角度、话题优先级	写作视角与认知框架.md
L4 素材策略	引用来源、权威对象、案例和数据使用方式	写作视角与认知框架.md
L5 认知框架	世界观、价值判断、核心假设、反复出现的命题	写作视角与认知框架.md
L6 视觉风格	配图策略、排版格式、字体层级、色彩使用	视觉风格指南.md

简单说：

• L1-L2 解决“怎么写”
• L3-L5 解决“怎么想”
• L6 解决“怎么呈现”

完整的风格复刻需要三者结合。只复刻语言，通常会“读着像但想法不像”；只复刻结构，通常会“框架像但语气不像”；忽略视觉风格，图文内容就会“文字像但看着不像”。

语料要求

真实蒸馏建议准备：

• 至少 20 篇完整文章
• 文件格式建议为 .md 或 .txt
• 尽量覆盖不同时期、不同主题、不同类型的文章

examples/format-only/ 只用于演示目录格式，不代表真实蒸馏效果。不要用少量短文评估这个 skill 的上限。

蒸馏流程

一次完整蒸馏通常分成 6 步：

1. 收集原始文章：手动抓取或让 agent 抓取至少 20 篇完整文章，统一放进 raw/ 或 raw-corpus/。
2. 建立元数据：为每篇文章标注标题、日期、类型、主题、hook、结构、素材来源等信息，形成 _meta/。
3. 分析语言 DNA：统计词频、句长、标点、段落节奏、常用表达和中英文混用习惯，输出 语言DNA.md。
4. 提炼结构模板：标注开头、转折、正文组织、结尾方式，把不同文章类型整理成可复用结构，输出 文章结构模板.md。
5. 归纳认知框架：通过深度阅读提取选题逻辑、素材偏好、价值判断、核心假设和反复出现的命题，输出 写作视角与认知框架.md。
6. 整理视觉风格：如果原文包含图片或特殊排版，分析配图策略、截图功能、字体层级、色彩和图文协作方式，输出 视觉风格指南.md。

最后，agent 会把这些结果整合成 Writing-DNA.md。这份文档是后续复刻写作风格时最常用的总入口。

蒸馏产物

一次完整蒸馏通常会输出这些文件：

作者或账号名称/
├── raw/                    # 原始文章语料
├── _meta/                  # 元数据标注
├── 语言DNA.md              # L1 分析产物
├── 文章结构模板.md          # L2 分析产物
├── 写作视角与认知框架.md    # L3-L5 分析产物
├── 视觉风格指南.md          # L6 分析产物
└── Writing-DNA.md          # 最终整合文档

其中 Writing-DNA.md 是最终整合文档，适合后续直接喂给 agent，作为写作前的风格上下文。

快速上手

推荐直接在 MoxtHub 打开 writing-dna-skill：Moxt 拥有强大的上下文空间和 AI 原生的文件格式，方便 agent 处理长文内容。

1. 让 agent 抓取或手动准备至少 20 篇完整文章，放入命名为 raw/ 或 raw-corpus/ 的文件夹。
2. 让 agent 使用 writing-dna-skill，或完整阅读 SKILL.md，并按流程蒸馏所有文章。
3. 等待 agent 输出：
   • 语言DNA.md
   • 文章结构模板.md
   • 写作视角与认知框架.md
   • 视觉风格指南.md
   • Writing-DNA.md
4. 后续写作前，让 agent 先重新阅读所有蒸馏产物，再按该风格写作。

可以这样开始一次蒸馏：

请使用 writing-dna-skill 蒸馏分析这个目录里的所有文章。

使用蒸馏好的风格

想让复刻效果更稳定，不要只在第一次写作时让 agent 读取风格文档。SKILL.md 第六节把每次写作前的读取要求写成了硬性步骤：

1. 先布置写作任务，给 agent 相关选题的足够上下文、认知等。
2. 让 agent 重新阅读全部蒸馏产物——四份分层产物加 Writing-DNA.md，一份都不跳过。整合文档是压缩后的结论，具体语感只存在于分层产物里。
3. 再让 agent 读 5 篇体裁和题材最接近的 raw/ 原文。优先按 _meta/ 的 article_type 和 topic_tags 筛，匹配超过 5 篇时取最近的。读原文不是为了找素材，而是校准分层产物描述不出来的东西：句子的呼吸感、段落怎么接、什么时候突然用一个短句。
4. 要求 agent 按蒸馏产物中的语言、结构、认知和视觉规则写作。
5. 如果输出不像，要求 agent 对照 语言DNA.md、文章结构模板.md、写作视角与认知框架.md 逐项自检后重写。

规则冲突时的优先级：用户本次的明确指令 > 匹配当前体裁的结构模板 > 语言特征与视觉风格 > 认知框架。原文里的具体观点和事实不能直接搬进新文章——复刻的是写法，不是内容。

可以这样说：

请先阅读这个作者目录下的所有蒸馏产物，尤其是 Writing-DNA.md、语言DNA.md、文章结构模板.md、写作视角与认知框架.md 和视觉风格指南.md。

然后从 raw/ 里挑 5 篇和我这次要写的题材最接近的原文通读，说一下这 5 篇的共同语感。

最后根据这些风格规则，帮我写一篇面向读者的文章。主题是：……

写完之后：去掉 AI 味

蒸馏得再准，模型写出来还是常有 AI 痕迹。这是两个不同的问题。蒸馏解决的是「像不像这个作者」，AI 味是「像不像机器写的」。风格对了，句子仍可能动不动就「不是 A 而是 B」、小标题一律「一、二、三」编号、段落开头突然抛一句「听起来像……」。

skills/lieflat-less-ai-tone/ 是流程的最后一道工序，按明确列出的规则清理这类痕迹。它随本仓库一起安装，不需要额外操作。

装好之后：

请用 lieflat-less-ai-tone 的规则，检查并改写刚才那篇文章。
先读这个作者的语言DNA.md，冲突时以蒸馏产物为准。

它采用白名单式改写，只处理清单里列出的规则，没命中的句子逐字保留，也不动文章结构。每条规则都有可定位的触发标记，并经过 300 篇 AI 输出与 329 篇真实文章的对比实测，不靠"读起来像 AI"这种语感判断。完整的研究方法、数据与被推翻的预设见 lieflat-less-ai-tone。

要点：蒸馏产物与去 AI 味规则冲突时，以蒸馏产物为准。 那是目标作者的真实写法，不是 AI 痕迹。比如某作者本来就爱用破折号，就不该按通用规则删掉。

为什么有效

这个 skill 的核心是让 agent 把风格拆成多个可检查的层次，将写作风格用脚本分析的方式沉淀总结，并在写作前将注意力聚焦在蒸馏产物上，因此能复刻写作风格。

这样得到的是一组可以被 agent 反复读取、执行和自检的风格规则。

Make 写作 Happy Again ！

重要边界

这个项目用于学习、分析、风格研究和个人写作资产沉淀。请不要用于冒充作者本人、误导读者，或侵犯他人版权。

开源仓库不应包含未经授权的原文语料。你可以放目录模板、字段模板、说明文档和自己拥有版权的示例文本。

文件结构

writing-dna-skill/
├── SKILL.md
├── README.md
├── README.en.md
├── LICENSE
├── agents/
│   └── openai.yaml
├── references/
│   └── workflow.en.md
├── docs/
│   ├── release-checklist.md
│   └── usage-boundaries.md
├── templates/
│   └── author-corpus/
│       ├── zh/                  # 中文模板与中文文件名
│       └── en/                  # 英文模板与英文文件名
├── skills/
│   └── lieflat-less-ai-tone/    # 去 AI 味规则集（随本仓库安装）
└── examples/
    └── format-only/

License

在 MoxtHub 使用 writing-dna-skill →

MIT

Writing DNA Distiller

中文 | English

Skill name: writing-dna-skill

An agent skill for distilling the writing style of any author or publication into reusable rules.

This is not a simple "write like this person" prompt. It gives an agent a complete distillation workflow for turning a corpus of full articles into reusable Writing DNA: language, structure, topic selection, source strategy, cognitive frames, and visual style. Before writing, the agent can reread these documents to reproduce the target style with greater consistency.

This skill was created at moxt.ai.

Open Writing DNA Distiller on MoxtHub →

The Problem It Solves

When people ask AI to imitate an author, they often provide a few examples and say, "Write in this style." The result is unreliable because the agent is guessing from surface-level tone.

writing-dna-skill takes a different approach. Instead of asking the agent to imitate immediately, it first analyzes a substantial corpus of complete articles, turns recurring style choices into explicit rules, and then uses those rules to guide future writing.

The skill examines:

• recurring words, sentence structures, punctuation, and rhythm
• how articles open, develop, transition, and close
• which topics the author selects or avoids
• which sources, examples, data, and screenshots appear repeatedly
• the worldview, value judgments, and assumptions behind the writing
• image strategy, layout, typography hierarchy, and color when the source is visual

Best For

• analyzing the stable writing style of an author, account, or publication
• turning your own writing history into a reusable style asset
• comparing how different authors approach the same topic
• giving an AI agent more reliable context for style-consistent writing

Chinese And English Artifacts

The skill keeps the conversation language separate from the artifact language. A user may speak Chinese while analyzing an English newsletter, or speak English while building a Chinese writing system.

• Chinese artifacts use templates/author-corpus/zh/ and Chinese filenames.
• English artifacts use templates/author-corpus/en/ and English filenames.
• An explicit user choice takes priority. Otherwise, the source corpus determines the artifact language.

What It Can Analyze

The skill turns "style" from a vague impression into rules that can be recorded, reused, and checked. It looks beyond tone to understand why an article is written in a particular way, how evidence is organized, and how conclusions are formed.

Suitable source material includes:

• long-form articles, newsletters, columns, blogs, and public-account posts
• image-led posts on Xiaohongshu, Weibo, and other social platforms
• your own archive of past writing
• articles by multiple authors writing about the same subject

If the source includes screenshots, illustrations, charts, or distinctive formatting, preserve them. For visual publications, image placement, evidentiary screenshots, heading hierarchy, and whitespace can be as important as the prose itself.

Six Layers of Distillation

Layer	What It Analyzes	Output
L1 Surface Language	Word frequency, sentence length, punctuation, rhetoric, recurring expressions	language-dna.md
L2 Article Structure	Opening hook, body architecture, transitions, closing pattern	structure-patterns.md
L3 Topic Logic	Timing, angle selection, topic priorities	cognitive-framework.md
L4 Source Strategy	References, authorities, examples, and use of data	cognitive-framework.md
L5 Cognitive Frame	Worldview, values, assumptions, and recurring propositions	cognitive-framework.md
L6 Visual Style	Image strategy, formatting, typography hierarchy, and color	visual-style-guide.md

In short:

• L1-L2 explain how the author writes.
• L3-L5 explain how the author thinks.
• L6 explains how the work is presented.

Reliable style reproduction needs all three. Language alone may sound similar while expressing different ideas. Structure alone may resemble the original outline but miss the voice. Ignoring visual style can make image-led content feel wrong even when the prose is close.

Corpus Requirements

For a real distillation, prepare:

• at least 20 complete articles
• .md or .txt source files
• material from different periods, topics, and article types when possible

examples/format-only/ demonstrates the expected directory structure. It is not a reliable style corpus, and a few short samples should not be used to judge the skill's full capability.

Distillation Workflow

A complete distillation normally follows six steps:

1. Collect the source corpus: Gather at least 20 complete articles and place them in raw/ or raw-corpus/.
2. Build metadata: Record the title, date, article type, topic, hook, structure, and source strategy for each article in _meta/.
3. Analyze language DNA: Measure word frequency, sentence length, punctuation, paragraph rhythm, recurring expressions, and mixed-language habits. Write language-dna.md.
4. Extract structure patterns: Mark openings, turns, body organization, and endings, then turn repeated article shapes into structure-patterns.md.
5. Identify cognitive frames: Distill topic logic, source preferences, values, assumptions, and recurring propositions into cognitive-framework.md.
6. Document visual style: When the corpus includes images or distinctive formatting, analyze image roles, screenshot evidence, typography, color, and text-image collaboration. Write visual-style-guide.md.

The agent then combines the results into Writing-DNA.md, the main document to reread before writing in the distilled style.

Output

A complete project usually produces:

author-or-publication/
├── raw/                    # Complete source articles
├── _meta/                  # Article-level metadata
├── language-dna.md          # L1 analysis
├── structure-patterns.md    # L2 analysis
├── cognitive-framework.md   # L3-L5 analysis
├── visual-style-guide.md    # L6 analysis
└── Writing-DNA.md          # Integrated style document

Writing-DNA.md is the primary reusable context for future writing sessions.

Quick Start

We recommend opening writing-dna-skill directly on MoxtHub, where Moxt's large context window and AI-native file workflow make it easier for an agent to process complete long-form corpora.

1. Collect or ask an agent to collect at least 20 complete articles and place them in raw/ or raw-corpus/.
2. Ask the agent to use writing-dna-skill, or to read SKILL.md and references/workflow.en.md, before processing the corpus.
3. Wait for the agent to produce:
   • language-dna.md
   • structure-patterns.md
   • cognitive-framework.md
   • visual-style-guide.md
   • Writing-DNA.md
4. Before future writing tasks, ask the agent to reread all distilled style documents.

Start a distillation with:

Use writing-dna-skill to distill the writing style from every article in this directory.

Writing With a Distilled Style

For more consistent results, do not load the style documents only once. Section six of SKILL.md makes the pre-writing reading a required step:

1. Give the agent the topic, audience, purpose, and relevant context.
2. Ask it to reread every distilled artifact — the four layered files plus Writing-DNA.md, skipping none. The integrated document holds compressed conclusions; the actual cadence lives in the layered files.
3. Then have it read 5 articles from raw/ that are closest in content type and subject. Filter on article_type and topic_tags in _meta/ first; when more than five match, take the most recent. Reading source articles is not about harvesting material — it calibrates what the artifacts cannot describe: how sentences breathe, how paragraphs connect, when a short sentence lands.
4. Require the draft to follow the language, structure, cognitive, and visual rules.
5. If the result does not feel accurate, ask the agent to audit the draft against language-dna.md, structure-patterns.md, and cognitive-framework.md, then rewrite it.

Priority when rules conflict: the user's explicit instructions for this piece > the structure pattern matching the content type > language characteristics and visual style > cognitive frames. Specific claims and facts from the source articles must not be carried into the new piece — you are reproducing how the author writes, not what they wrote.

Example:

Read every distilled style document in this author's directory, especially
Writing-DNA.md, language-dna.md, structure-patterns.md,
cognitive-framework.md, and visual-style-guide.md.

Then pick the 5 articles in raw/ closest to my subject, read them, and tell me
what they share in voice.

Finally, write an article for the target audience using those rules.
Topic: ...

After Writing: Removing AI Tone

However accurate the distillation, model output often still reads as machine-written. These are two separate problems. Distillation addresses does this sound like the author, while AI tone is does this sound like a machine. The style can be right while "it's not A, but B" keeps recurring, every subheading is numbered "one, two, three", and paragraphs open with a bare comment like "Sounds like...".

skills/lieflat-less-ai-tone/ is the final step in the workflow, clearing those tells against an explicit rule list. It installs together with this repository; no extra step is needed.

Then:

Apply the lieflat-less-ai-tone rules to the article you just wrote.
Read this author's language-dna.md first; where they conflict, the distilled
artifacts win.

It rewrites from a whitelist, touching only the listed problems; sentences matching no rule are preserved verbatim, and the article's structure is left intact. Every rule has a locatable trigger and was measured against 300 AI outputs and 329 human articles, rather than resting on "this reads like AI". Full methodology, data, and the fifteen hypotheses that failed are documented in lieflat-less-ai-tone.

Key point: where the distilled artifacts conflict with the AI-tone rules, the artifacts win. That is how the target author actually writes, not a machine tell. If an author genuinely favors em dashes, a generic rule should not strip them.

Why It Works

The skill separates writing style into layers that an agent can inspect, execute, and audit. Instead of relying on a vague impression of similarity, it records explicit style rules and directs the agent's attention back to them before each writing task.

The result is a reusable set of instructions rather than a one-time imitation prompt.

Important Boundaries

This project is intended for learning, analysis, style research, and building personal writing assets. Do not use it to impersonate an author, mislead readers, or violate copyright.

Do not place unauthorized source articles in a public repository. Public releases should contain directory templates, field schemas, documentation, and examples you have permission to share.

Repository Structure

writing-dna-skill/
├── SKILL.md
├── README.md
├── README.en.md
├── LICENSE
├── agents/
│   └── openai.yaml
├── references/
│   └── workflow.en.md
├── docs/
│   ├── release-checklist.md
│   └── usage-boundaries.md
├── templates/
│   └── author-corpus/
│       ├── zh/                  # Chinese templates and filenames
│       └── en/                  # English templates and filenames
├── skills/
│   └── lieflat-less-ai-tone/    # AI-tone rules (installed with this repo)
└── examples/
    └── format-only/

License

Use writing-dna-skill on MoxtHub →

MIT