Style Extraction

Extract writing style characteristics from articles to establish and maintain personal style profiles.

Core Concepts

Style is not an abstract "good" or "bad", but a set of quantifiable and replicable specific characteristics. This skill does two things:

Analyze and Store: Break down the style of an article into specific dimensions and store it in the style material library
Merge and Update: Select features from the style material library to synthesize or update the main style profile

The main style profile (

my_style.json

) is the recipe actually called by creation and polishing skills. The style material library is the raw material accumulated over the long term.

Data Directory

All style data is stored in

./writing-workspace/styles/

writing-workspace/styles/
├── my_style.json       # 主力风格档案（创作/润色读取的那一份）
├── index.jsonl          # 风格素材索引（JSONL 格式，每行一条 {id, source_title, author, analyzed_at}）
└── entries/            # 每篇分析过的文章一条记录
    └── sty_xxx.json

The directory structure is automatically created on first use.

Mode Judgment

Automatically select the mode based on user intent:

User provides an article and says "analyze style" or "look at the writing style of this article" → Mode A: Analyze and Store
User says "absorb this", "update my style", "merge these" → Mode B: Merge and Update
User provides their own article and says "analyze my writing style" → Mode A, and mark
```
is_self: true
```

Mode A: Analyze and Store

Analysis Dimensions

Analyze the following dimensions one by one, with each dimension summarized in one sentence:

Language Dimensions:

Dimension	Focus
vocabulary_level	Colloquial / formal / mixed? Are words more daily or professional?
sentence_rhythm	Dominated by short sentences or long sentences? Is the rhythm rapid or slow?
favorite_expressions	Repeated catchphrases, transition words, modal particles (extract 3-5)
punctuation_habits	More dashes or ellipses? Comma density?
person_perspective	First person / second person / third person? Is there any switch?
emotion_intensity	Calm and restrained / moderate / passionate and sensational? Is the attitude clear or gentle?

Structure Dimensions:

Dimension	Focus
opening_pattern	Story introduction / question introduction / data impact / scene description / direct presentation of viewpoint?
paragraph_rhythm	Regular alternation of paragraph lengths? Are key sentences independent into paragraphs?
argument_logic	Inductive (case → viewpoint) / deductive (viewpoint → argument) / parallel?
transition_style	Transition with conjunctions / colloquial transition / direct jump?
closing_pattern	Closing with golden sentences / open question / call to action / story loop?
title_pattern	Counter-intuitive / numeric / pain point / suspense / direct?

Refine Highlights

In addition to dimension analysis, extract 2-3 most worthy specific skills from this article for reference. Don't talk in general terms, be specific about "how to do it".

Good examples of highlights:

"Use three consecutive rhetorical questions to create rhythm, then wrap up with a declarative sentence to form the impact of 'ask-ask-ask → answer'"
"End each large paragraph with a short sentence on a separate line as a summary, nailing the viewpoint like a nail"

Execution Process

Ensure the
```
./writing-workspace/styles/
```
directory exists; create it (including the
```
entries/
```
subdirectory) if it doesn't
If
```
index.jsonl
```
doesn't exist, create an empty file
Read the full text and understand the overall style tone
Analyze the 12 dimensions one by one, with each dimension summarized in one sentence
Extract 2-3 most worthy specific skills
Generate a style material entry JSON, with ID format:
```
sty_YYYYMMDD_NNN
```
. How to determine NNN: Take the ID serial number of the last line in
index.jsonl
, then list the
sty_YYYYMMDD_*.json
files already existing in the
entries/
directory on the same day, use the larger serial number +1 from the two as the starting number. If the index is empty and there are no existing files on the day, start from 001.
Write to
```
styles/entries/{id}.json
```
Update
styles/index.jsonl
immediately——append a line of JSON
```
{id, source_title, author, analyzed_at}
```
at the end of the file. This step cannot be skipped, as the index is the entry for other skills to retrieve.
Show the analysis results to the user (in a readable format, not raw JSON)
Ask the user: "Would you like to absorb some of these features into your main style?"

Style Material JSON Structure

json

{
  "id": "sty_20260327_001",
  "source": {
    "title": "Original Title",
    "author": "Author",
    "url": "Source Link (if any)",
    "is_self": false
  },
  "language": {
    "vocabulary_level": "",
    "sentence_rhythm": "",
    "favorite_expressions": [],
    "punctuation_habits": "",
    "person_perspective": "",
    "emotion_intensity": ""
  },
  "structure": {
    "opening_pattern": "",
    "paragraph_rhythm": "",
    "argument_logic": "",
    "transition_style": "",
    "closing_pattern": "",
    "title_pattern": ""
  },
  "highlights": "Natural language description of 2-3 specific skills",
  "analyzed_at": "ISO 8601 timestamp"
}

Notes

If only one article is provided, after analysis, remind the user: "The style analysis of a single article may not be comprehensive. It is recommended to provide 3-5 articles by the same author or with the same style for more accurate feature extraction."
Focus on "human-like" features during analysis——those elements that make the article not read like it was written by AI. These features are particularly valuable in subsequent creation and polishing.
Must correctly escape special characters when writing JSON: Double quotes
```
"
```
in text fields such as
```
highlights
```
,
```
overall_tone
```
,
```
anti_ai_notes
```
,
```
favorite_expressions
```
must be escaped as
```
\"
```
, backslashes
```
\
```
as
```
\\
```
, and line breaks as
```
\n
```
. When using the Write tool to write JSON, ensure that string values do not contain unescaped double quotes.

Mode B: Merge/Update Main Style

Execution Process

Read the current main style profile
```
styles/my_style.json
```
- If it doesn't exist, enter the initialization process (see below)
Read the style materials specified by the user (can be the newly analyzed ones or existing ones in entries/)
Compare the feature differences between the current main style and the new materials dimension by dimension
For each dimension with differences, show the comparison to the user and ask:
- "Your current
```
opening_pattern
```
  is 'story introduction', this article uses 'data impact'. Would you like to replace, keep, or combine both?"
Update
```
my_style.json
```
according to the user's choice:
- Increase
```
version
```
  by 1
- Update
```
last_updated
```
  to the current time
- Record the source of each feature in
```
trait_sources
```
  (which style material entry)
Show the updated summary of the main style

Main Style Initialization

When

my_style.json

does not exist:

Case 1: User provides their own article

Directly use the analysis result as the initial version of the main style
Set version to 1, and mark all
```
trait_sources
```
as "my own article"

Case 2: User provides someone else's article

After analysis, guide the user to select the desired features
Assemble into the first version of the main style

Main Style JSON Structure

json

{
  "name": "My Main Style",
  "version": 1,
  "last_updated": "ISO 8601 timestamp",
  "language": {
    "vocabulary_level": "",
    "sentence_rhythm": "",
    "favorite_expressions": [],
    "punctuation_habits": "",
    "person_perspective": "",
    "emotion_intensity": "",
    "forbidden_words": ["值得注意的是", "总而言之", "综上所述", "让我们", "在当今社会"]
  },
  "structure": {
    "opening_pattern": "",
    "paragraph_rhythm": "",
    "argument_logic": "",
    "transition_style": "",
    "closing_pattern": "",
    "title_pattern": ""
  },
  "overall_tone": "A natural language description of the overall style tone",
  "anti_ai_notes": "Avoid parallel sentences, avoid mechanical structures like 'first, second, last', avoid paragraphs of similar length...",
  "trait_sources": {
    "opening_pattern": "From sty_xxx",
    "sentence_rhythm": "My own article"
  }
}

forbidden_words

initially contains common AI phrases, and users can supplement it at any time. Creation and polishing skills will read this list.

Output Format

Present the analysis results in readable Chinese, do not directly output JSON. Example format:

## 风格分析报告

**来源：** 《文章标题》 by 作者

### 语言特征
- **词汇风格：** 口语化为主，偶尔蹦出书面用语制造反差
- **句式节奏：** 短句密集，长句用来收尾，形成"碎碎碎——收"的节奏
- **常用表达：** "说白了""你想想看""本质上就是"
- ...

### 结构特征
- **开头方式：** 故事切入——用一个具体场景把读者拉进来
- ...

### 最值得借鉴的技巧
1. ...
2. ...

---
风格素材已保存为 sty_20260327_001
要把其中某些特征吸收到你的主力风格里吗？

style-extract

NPX Install

Tags

SKILL.md Content (Chinese)

Style Extraction

Core Concepts

Data Directory

Mode Judgment

Mode A: Analyze and Store

Analysis Dimensions

Refine Highlights

Execution Process

Style Material JSON Structure

Notes

Mode B: Merge/Update Main Style

Execution Process

Main Style Initialization

Main Style JSON Structure

Output Format