style-extractor

name	style-extractor
description	Extract personal writing style from samples to create a style profile. Analyzes sentence patterns, vocabulary preferences, argumentation structure, punctuation habits, tone markers, and AI-distinguishability signals across Chinese and English. Use when "提炼我的写作风格", "extract my style", "create style profile", "analyze my writing", "风格画像", "个性化风格提取", or when setting up petfish-style-rewriter for the first time.
compatibility	opencode
metadata	{"version":"1.0.0","author":"Petfish"}

Style Extractor

Purpose

Transform a user's writing samples into a structured, actionable style profile. The profile captures the user's recognizably-human voice across sentence-level, paragraph-level, vocabulary-level, and argumentation-level signals. The output is consumed by petfish-style-rewriter so that rewrites match the user's style rather than a generic target.

This skill is extraction only. It does not detect AI-generated text and it does not rewrite content.

Triggers / Activation

Use this skill when the user asks for any of the following:

• 提炼我的写作风格 / 提炼写作风格
• extract my style / extract my writing style
• create style profile / build a style profile
• analyze my writing / analyze my style
• 风格画像 / 我的风格画像
• 个性化风格提取 / 个性化设置
• 第一次设置 petfish-style-rewriter
• first-time setup for style-rewriter

Do not use this skill for:

• Direct rewriting or polishing → route to petfish-style-rewriter
• Generic AI-text detection → not supported
• Document conversion only → route to doc-reader

Prerequisites

1. The user must provide 3 or more writing samples.
2. More samples produce a better extraction.
3. Minimum bilingual requirement: if the user writes in both Chinese and English, provide at least 1 Chinese + 1 English sample.
4. Accepted formats: Markdown (.md), DOCX (.docx), PDF (.pdf).
   • DOCX and PDF must first be converted to Markdown via doc-reader (scripts/doc_to_markdown.py) before analysis.

Domain Rules

• This skill never writes, edits, or rewrites the user's text. It only reads and measures.
• The final style profile is written to .petfish/style-profile.md in the active project, not inside the skill pack. Style profiles are private user data.
• Quantitative metrics are produced by a stdlib-only Python script (scripts/analyze_style.py). Qualitative synthesis is performed by the LLM using those metrics as evidence.
• Chinese and English are analyzed separately and then cross-compared. Do not collapse them into one undifferentiated profile.
• Signature phrases, pronoun preferences, and register patterns must be derived from the data, not invented.
• "AI-distinguishability signals" are positive human markers (burstiness, idiosyncrasy, imperfection) that should be preserved, not eliminated.

Extraction Dimensions

Based on nuwa-skill's multi-dimensional extraction methodology:

1. Sentence statistics — length, variance, complexity, burstiness (coefficient of variation).
2. Vocabulary profile — density, formality tier, signature phrases, technical density.
3. Argumentation pattern — how the user builds and closes arguments.
4. Transition style — connector preferences and logical-flow markers.
5. Paragraph organization — length, structure, topic-sentence position.
6. Punctuation habits — comma-heavy vs period-heavy, semicolon/colon/dash usage.
7. Title/heading conventions — heading depth, numbering style, capitalization.
8. Opening/closing patterns — how pieces begin and end.
9. Person pronoun usage — 笔者/我们/本文/我 for Chinese; we/I/this paper for English.
10. Register awareness — formal vs casual, code-switching, domain tone.
11. AI distinguishability — what makes the writing recognizably human and should be preserved.

See references/extraction-dimensions.md for detailed definitions, interpretation guidance, and example profiles.

Two-Stage Workflow

Stage 0: Convert non-Markdown samples

For each PDF or DOCX sample, run:

uv run scripts/doc_to_markdown.py sample.docx --output sample.md
uv run scripts/doc_to_markdown.py sample.pdf --output sample.md

Stage 1: Quantitative analysis

Run the analyzer on the directory of Markdown samples:

uv run .opencode/skills/style-extractor/scripts/analyze_style.py --samples ./my-samples/ --output style-metrics.json

This produces style-metrics.json with per-file and aggregated metrics.

Stage 2: Qualitative synthesis

The agent reads style-metrics.json and each converted Markdown sample, then generates a style profile using the template at references/style-profile-template.md.

Save location (3-tier, global-first):

Priority	Path	When to use
Global (default)	~/.petfish/style-profile.md	Always save here by default — writing style is personal, not project-specific
Project-level	.petfish/style-profile.md	Only when user explicitly says "for this project only" or a project has a different style register

After saving, tell the user which location was used:

风格画像已保存到 ~/.petfish/style-profile.md（全局，所有项目共享）

The synthesis must:

• Cite concrete metrics from Stage 1.
• Quote representative short passages from the samples.
• Keep each section short and actionable for the rewriter.
• Flag uncertainty when sample size is small.

Decision Points

1. Enough samples? If fewer than 3 samples are provided, stop and ask for more. If the user is bilingual but supplies only one language, note the limitation in the profile.
2. Language split: Classify each sample as zh, en, or mixed. Chinese-only, English-only, and bilingual sample sets are all supported; the output structure adapts accordingly.
3. Metrics outlier handling: If one sample is very different from the others, report it as a possible register switch or outlier rather than averaging it away.
4. Profile confidence: Label the profile as high, medium, or low confidence based on sample count, consistency, and language coverage.

Execution Modes

Mode	Behavior
interactive	Ask the user for sample files/locations, confirm language coverage, explain the output location, and show a preview of the metrics before writing the final profile.
auto	Accept the provided sample paths, run Stage 0–2, and write the profile to ~/.petfish/style-profile.md (global) with minimal explanation.

Default to interactive when the user only says "extract my style" without providing paths. Default to auto when paths are explicit.

Output Contract

Required deliverables:

1. style-metrics.json — structured quantitative metrics (per-file + aggregated + cross-lingual).
2. .petfish/style-profile.md — qualitative style profile following references/style-profile-template.md.

The profile must contain:

• Extracted date and sample count
• Chinese fingerprint (if applicable)
• English fingerprint (if applicable)
• Cross-lingual consistency notes
• AI-distinguishability signals to preserve
• Rewriting guidelines for petfish-style-rewriter

Output Location

The profile is saved to global by default (your writing style is personal, not project-specific):

~/.petfish/style-profile.md                   ← global default (saved here by default)
.petfish/style-profile.md                     ← project-level override (use --project flag)

Default behavior: Save to ~/.petfish/style-profile.md so all projects share your personal style.

Project-level override: When a project needs a different style (e.g., ghost-writing for a client), save to .petfish/style-profile.md instead. Tell the user:

默认保存到全局 (~/.petfish/style-profile.md)，所有项目共享。 如果需要项目级覆盖，说"保存到这个项目"。

Migration helper: If the user already has .petfish/style-profile.md from a previous project and wants to promote it to global:

Copy-Item .petfish/style-profile.md ~/.petfish/style-profile.md

This file is private to the user. It is the contract between style-extractor and petfish-style-rewriter.

Integration with petfish-style-rewriter

petfish-style-rewriter loads the style profile at runtime using a 3-tier lookup:

1. .petfish/style-profile.md (project-level override — highest priority)
2. ~/.petfish/style-profile.md (global personal default)
3. Pack built-in rules (fallback)

If any profile exists, the rewriter:

• Uses the user's preferred sentence-length target and burstiness level.
• Preserves signature phrases and pronoun choices.
• Mirrors the user's argumentation structure and transition style.
• Respects the register map.
• Avoids adding patterns listed in the profile's DON'T add section.

If no profile exists, petfish-style-rewriter falls back to the default Petfish style.

Anti-patterns

• Inventing style features without evidence: claiming the user prefers a phrase that does not appear in the samples.
• Averaging away register variation: collapsing formal and casual samples into one undifferentiated profile.
• Treating human imperfection as a bug: burstiness, minor asymmetry, and idiosyncratic connectors are human signals to preserve.
• Omitting the rewriting guidelines section: the profile is useless to the rewriter without explicit DO/DON'T rules.
• Hardcoding a persona: the template must remain generic; only values derived from the samples are filled in.
• Skipping Stage 1: relying only on LLM impression without quantitative anchors produces inconsistent profiles.

Handoff & Boundaries

This skill owns

• Collecting or confirming writing sample paths.
• Converting PDF/DOCX samples to Markdown via doc-reader.
• Computing quantitative style metrics.
• Synthesizing metrics and sample evidence into .petfish/style-profile.md.
• Updating an existing profile when new samples are provided.

This skill does not own

• Rewriting the user's text → petfish-style-rewriter
• Detecting AI-generated text → not supported
• General document conversion without style analysis → doc-reader
• Cross-document series-style governance → series-style-governor

Must Do

• Verify at least 3 samples before extraction.
• Convert PDF/DOCX to Markdown before analysis.
• Run scripts/analyze_style.py and include its metrics in the final profile.
• Write the profile to .petfish/style-profile.md.
• Keep the profile generic in structure but specific in values.
• Preserve human imperfections as positive signals.

Must Not Do

• Do not rewrite, edit, or polish the samples.
• Do not hardcode a specific user's style into the template.
• Do not produce detection or rewriting capabilities inside this skill.
• Do not place the generated profile inside the skill pack directory.
• Do not skip the quantitative stage.

References

• references/extraction-dimensions.md — dimension definitions and interpretation
• references/style-profile-template.md — output contract template
• scripts/analyze_style.py — stdlib-only quantitative analyzer