| name | style-extractor |
| description | Extract and document writing styles from source texts into reusable style guides. Use when the user wants to analyze an author's voice, create a style guide from a book or document, capture writing patterns for replication, or build a writing style rubric. Triggers on 'extract style', 'analyze writing style', 'capture voice', 'style guide from', or 'writing style analysis'. |
Style Extractor
Extract reusable writing style guides from any source text. Produces four standardized deliverables that capture general patterns — not copied phrases — so the style can be replicated for any topic.
When to Use
- User asks to extract or analyze a writing style from a book, article, or document
- User wants to replicate an author's voice for new content
- User wants a style guide, voice card, or writing rubric from a source text
- User says "extract style from...", "analyze the writing style of...", or "capture this voice"
Prerequisites
MCP tools required:
mcp__fuzzy-search — PDF page counting, outline extraction, and page-level readingmcp__sequential-thinking__sequentialthinking — Structured analysis across dimensions
Output location: writing-styles/<style-name>/ in the project root. See writing-styles/README.md for the collection structure and writing-styles/_template/ for deliverable templates.
Workflow
Phase 1: Source Analysis
Identify the source text and assess its structure before reading.
For PDFs:
- Get page count:
mcp__fuzzy-search__get_pdf_page_count
- Get outline/TOC:
mcp__fuzzy-search__get_pdf_outline
- Understand document structure before sampling
Select 5–7 sampling points distributed across the text:
| Sample | Location | Why |
|---|
| Opening | First 10% | Capture introductory voice, setup patterns |
| Early-middle | 25–35% | Style settles after opening |
| Middle | 45–55% | Core voice, least influenced by beginning/ending effects |
| Late-middle | 65–75% | Check for drift or adaptation |
| End | 85–95% | Closing patterns, evolved voice |
| +1–2 optional | Varied | Specific chapters, asides, or structurally distinct sections |
Read representative passages using mcp__fuzzy-search__extract_pdf_pages or the Read tool. Aim for ~50–80 pages total across all samples.
For non-PDF files: Use the Read tool directly. Apply the same sampling distribution across the document's length.
Phase 2: Dimension Extraction
Analyze 17 style dimensions across all sampled passages. Use mcp__sequential-thinking__sequentialthinking to organize observations systematically.
| # | Dimension | What to look for |
|---|
| 1 | Tone | Formal/informal, serious/playful, authoritative/conversational |
| 2 | Sentence length & structure | Average length, variation, complexity, use of fragments |
| 3 | Paragraph length & density | Short punchy vs. long expository, information density |
| 4 | Vocabulary level & register | Technical depth, word choice patterns, jargon usage |
| 5 | Transitions | How sentences, paragraphs, and sections connect |
| 6 | Humor & personality | Wit, asides, self-deprecation, enthusiasm markers |
| 7 | Directness | Hedging vs. assertion, qualification patterns |
| 8 | Emotional warmth | Distance vs. intimacy, encouragement, empathy |
| 9 | Explanation structure | Top-down vs. bottom-up, example-first vs. definition-first |
| 10 | Opening patterns | How sections/chapters begin, hooks, framing devices |
| 11 | Closing patterns | How sections end, callbacks, forward references |
| 12 | Punctuation habits | Em dashes, semicolons, parentheticals, ellipses, exclamation marks |
| 13 | Rhetorical moves | Analogies, rhetorical questions, repetition, contrast pairs |
| 14 | Reader address | Pronouns (we/you/one), assumed relationship, direct address |
| 15 | Use of examples | Frequency, placement, concrete vs. abstract, real vs. constructed |
| 16 | Historical/cultural references | Allusions, citations, name-dropping, assumed shared knowledge |
| 17 | Adaptive patterns | How style shifts with material difficulty, topic type, or audience |
Focus on patterns, not content. A style dimension describes HOW the author writes, never WHAT they write about.
For detailed guidance on analyzing each dimension, see references/style-dimensions.md.
Phase 2b: Narrative Construction (Fiction Only)
When the source text is fiction or narrative-driven writing, analyze 10 narrative dimensions in addition to the 17 surface style dimensions. Surface style captures how sentences read; narrative construction captures how stories are built. These are orthogonal layers — capturing both produces a significantly more complete style profile.
| # | Dimension | What to look for |
|---|
| 1 | Plot architecture | Single-track vs multi-subplot, causal tightness, resolution pattern |
| 2 | Temporal structure | Linear vs nonlinear, flashbacks, time jumps, anachrony |
| 3 | Character introduction | In-dialogue, in-action, external description, deferred naming |
| 4 | Emotional conveyance | Explicit labels vs embodied metaphors vs behavioral cues |
| 5 | Thematic handling | Explicit moralizing vs implicit, narrator commentary, thematic unity |
| 6 | Revelation & information | What's withheld, pacing, recontextualization depth, foreshadowing |
| 7 | Reader engagement | Fourth-wall status, direct address, audience positioning |
| 8 | Intertextual strategy | Named vs allusive references, pastiche, self-referential gestures |
| 9 | Moral framing & agency | Protagonist morality, resolution agency, authorial judgment |
| 10 | Event escalation | Intensity arc, event-type diversity, ending pattern |
Pay special attention to where the author's choices diverge from AI narrative defaults (e.g., nonlinear time, morally ambivalent protagonists, implicit themes). These divergences are the most valuable and hardest-to-replicate aspects of the style.
For detailed guidance, see references/narrative-dimensions.md.
Phase 3: Synthesis
Generate four deliverables in writing-styles/<style-name>/:
| Deliverable | Path | Target Length | Purpose |
|---|
| Full Style Guide | style/full-style-guide.md | 800–1500 words | Comprehensive analysis of all 17 dimensions |
| Voice Card | style/voice-card.md | ~300 words | One-page cheat sheet for quick reference |
| Do/Don't Checklist | style/do-dont.md | 20–30 items | Actionable guardrails for writing in this style |
| Style Rubric | evals/style-rubric.md | 5 dimensions | 1–5 scoring scale for evaluating style match |
Every deliverable MUST include 2–3 examples of neutral text rewritten in the extracted style. These examples demonstrate the style applied to content unrelated to the source.
For fiction sources: Each deliverable should include a "Narrative Construction" section alongside the surface style analysis. The voice card should list the author's distinctive narrative patterns (e.g., "nonlinear time, morally ambivalent protagonists, implicit themes"). The do/don't list should include narrative-level items (e.g., "Do: use subplots that contrast the main arc" / "Don't: resolve through protagonist epiphany"). The rubric should include at least one narrative construction dimension.
Use writing-styles/_template/ as the structural starting point for each file.
Phase 4: Output
-
Create the folder structure:
writing-styles/<style-name>/
├── style/
│ ├── full-style-guide.md
│ ├── voice-card.md
│ └── do-dont.md
└── evals/
└── style-rubric.md
-
Write all four deliverables.
-
Verify the structure matches the template layout in writing-styles/_template/.
Naming Convention
Use kebab-case: <author-or-title>
| Source | Directory Name |
|---|
| Mathematical Proofs by Chartrand | warm-academic-exposition |
| Hemingway's short stories | hemingway-short-stories |
| Stripe API documentation | stripe-api-docs |
| Paul Graham's essays | paul-graham-essays |
Best Practices
- Sample broadly. Pull from beginning, middle, AND end — styles evolve across a text.
- Prioritize consistency. Features that appear across multiple samples define the real style; one-off features are noise.
- Track adaptation. Note how style shifts for different material types, difficulty levels, or audiences. This is dimension 17 and often the most valuable.
- Extract patterns, not phrases. Never copy distinctive phrases as style rules. Identify the general pattern the phrase exemplifies.
- Use agent teams for large texts. Assign different sampling regions to different researchers, then synthesize.
- Cross-validate. If a pattern appears in only one sample, it may be a local anomaly — flag it but don't center the guide on it.
Anti-Patterns
- Extracting content as style. Topic-specific vocabulary is content, not style. "Uses the word 'eigenvalue'" is content; "introduces technical terms with an informal gloss" is style.
- Projecting absent features. Don't claim humor if there is none. Don't invent warmth in clinical prose. Report what IS there.
- Sampling from one section only. A single chapter gives a biased picture. The five-point sampling is non-negotiable.
- Copying signature phrases. "Always use the phrase 'Let us consider...'" is wrong. "Opens explorations with an inclusive invitation" is right.
- Domain-locked guides. A style guide that only works for the original domain has failed. The deliverables must transfer to any topic.
References
For detailed analysis guidance and deliverable templates:
