Stylometric analysis produces measurable patterns: sentence length distributions, vocabulary preferences, punctuation habits, rhetorical moves. These patterns describe a writer's voice with precision that subjective labels like "conversational" or "authoritative" cannot match. Yet translating these measurements into AI-compatible instructions presents its own challenge.
The gap between understanding style and getting AI to replicate it comes down to format. AI models need structured, explicit guidance—not vague instructions. A style specification bridges this gap: a document that translates measurable patterns into instructions AI can actually follow.
This article provides a comprehensive style specification framework and demonstrates implementation with Claude's custom styles feature and ChatGPT's custom instructions.
The Six Components of a Complete Style Specification
A style specification is more than a list of preferences. It's a structured instruction set that captures the mechanical, rhetorical, and tonal elements of writing. The framework functions as a style guide meets operating manual.
1. Voice Attributes (Tonal Foundation)
This section defines the fundamental character of the writing:
Formality Level: Position on the formal-casual spectrum
Example: "Professional but approachable—uses contractions, avoids jargon, occasionally addresses reader directly"
Authority Stance: How expertise is positioned
Example: "Evidence-based authority without personal credentialing—cite research, not résumé"
Reader Relationship: The implied dynamic with the audience
Example: "Peer-to-peer collaboration, not expert-to-novice instruction"
2. Sentence Patterns (Mechanical Structure)
This section documents syntactic preferences from stylometric analysis:
Average Sentence Length: Based on analysis (e.g., 15-20 words)
Sentence Variety: Mix of simple, compound, complex structures
Example: "Favor declarative statements. Use fragments sparingly for emphasis. Avoid question-heavy openings."
Punctuation Style: Em dashes vs. semicolons, Oxford commas, parenthetical preferences
3. Vocabulary Guidelines (Word Choice)
This section specifies lexical territory:
Preferred Terms: Words and phrases consistently used
Example: "Use 'framework' not 'paradigm'; 'approach' not 'methodology'"
Words to Avoid: Buzzwords, clichés, overused phrases
Example: "Never: synergy, leverage (as verb), game-changer, deep dive"
Specificity Level: Concrete vs. abstract language preferences
4. Structural Preferences (Organization)
This section defines compositional habits:
Paragraph Length: Typical range (e.g., 3-5 sentences, 60-100 words)
Section Organization: How longer pieces are structured
Example: "Problem → Evidence → Solution → Application, not abstract-to-specific"
Transition Style: Explicit connectors vs. implicit flow
5. Rhetorical Moves (Argumentative Style)
This section captures how ideas are built and presented:
Concept Introduction: Lead with application or define terms first?
Example: "Start with the reader's problem, not the theoretical background"
Evidence Handling: Citation style, data presentation
Example: "Integrate research seamlessly—'Studies show X' not 'According to Smith (2020)'"
Prohibited Patterns: Rhetorical moves to avoid
Example: "No rhetorical scaffolding—never announce what's coming next"
6. Sample Passages (Annotated Examples)
Include 2-3 representative excerpts with annotations explaining characteristic elements:
[SAMPLE PASSAGE]
"The problem with most AI writing advice? It assumes writers want to sound
like everyone else. But distinctive voice isn't a luxury—it's the
difference between content people skim and content they remember."
[ANNOTATIONS]
- Opening question for engagement (use sparingly)
- Immediate answer (no suspense-building)
- Rhetorical contrast ("not X, but Y")
- Concrete outcome focus (remember vs. generic "engage")
- Sentence variety: 7 words, then 17 wordsImplementation: Claude Custom Styles
Claude's custom styles feature provides clean implementation for style specifications.
Step 1: Prepare the Style Spec
Before uploading, the specification should be:
- Complete: All six sections filled
- Specific: Concrete examples, not vague descriptions
- Concise: 2-3 pages maximum (AI models have context limits)
Step 2: Access Custom Styles
- Open Claude.ai or Claude desktop app
- Navigate to Settings → Custom Styles
- Click "Create New Style"
Step 3: Structure the Entry
Name the style descriptively (e.g., "Professional_Analysis_Voice"). Format the spec for AI comprehension:
VOICE ATTRIBUTES:
- Formality: Professional but approachable; uses contractions
- Authority: Evidence-based, not credential-based
- Reader relationship: Peer collaboration
SENTENCE PATTERNS:
- Average length: 18 words
- Variety: Mix declarative with occasional fragments for emphasis
- Punctuation: Semicolons for related ideas; em-dashes sparingly
VOCABULARY:
- Prefer: framework, approach, patterns, specific
- Avoid: synergy, leverage, paradigm, game-changer
- Level: Educated general audience, minimal jargon
STRUCTURE:
- Paragraphs: 3-5 sentences
- Organization: Problem → Evidence → Solution
- Transitions: Implicit flow over explicit connectors
RHETORICAL MOVES:
- Introductions: Lead with reader's problem
- Evidence: Integrated, not cited formally
- Never: Announce structure, use rhetorical questions excessively
SAMPLE (annotated):
[Include strongest example with annotations]
Step 4: Test and Validate
Generate 3-5 test pieces on different topics. Review each against original writing:
- Does the sentence length distribution match?
- Are vocabulary choices consistent?
- Do structural patterns align?
- Does the output sound like the original writer?
Step 5: Iterate
Based on testing, refine the spec:
- If output is too formal → Add more casual examples
- If vocabulary is off → Expand preferred/avoid lists
- If structure feels wrong → Clarify organizational preferences
Track specific failures. When AI consistently misses something, that element needs more explicit specification.
Implementation: ChatGPT Custom Instructions
ChatGPT uses custom instructions rather than named styles. The approach is similar but requires fitting the spec into two fields:
"What would you like ChatGPT to know about you?"
Include:
- Writing context (what is written, for whom)
- Voice attributes summary
- Key vocabulary preferences
"How would you like ChatGPT to respond?"
Include:
- Sentence pattern requirements
- Structural preferences
- Rhetorical moves and prohibitions
- One annotated sample passage
Limitation: ChatGPT's custom instructions have stricter character limits than Claude's custom styles. Writers may need to prioritize the most distinctive elements of their voice.
Platform Comparison: Claude vs. ChatGPT
Testing style specifications across both platforms reveals different strengths:
| Claude Strengths | ChatGPT Strengths |
|---|---|
| Better at maintaining complex sentence structures | Strong adherence to explicit vocabulary lists |
| More consistent with nuanced tonal requirements | Reliable structural template following |
| Handles longer style specifications without truncation | Better at formulaic pattern replication |
| Custom styles persist across conversations | — |
Both struggle with:
- Subtle voice attributes that resist explicit description
- Context-dependent style shifts
- Maintaining consistency over very long outputs
Writers whose style relies heavily on sentence rhythm and tonal nuance may find Claude's custom styles offer better fidelity. Writers whose style is more defined by specific vocabulary and structural patterns may find ChatGPT's custom instructions work well.
Common Pitfalls
Too Vague vs. Too Restrictive
Too Vague:
"Be conversational and engaging"
Problem: AI interprets differently every time
Fix: "Use contractions 60-70% of the time; address reader directly 2-3x per paragraph"
Too Restrictive:
"Always use exactly 15-word sentences"
Problem: Creates robotic, unnatural output
Fix: "Average 15-20 words per sentence; vary from 8-30 words"
Over-Specifying vs. Under-Specifying
Over-Specifying:
Including preferences for every grammatical choice
Problem: Overwhelming; AI can't prioritize
Fix: Focus on distinctive patterns, not universal rules
Under-Specifying:
Only describing tone without mechanics
Problem: AI has no concrete guidance
Fix: Include both "what" (voice) and "how" (mechanics)
Static Specifications
Writing evolves; specifications should too.
- Review quarterly: Does this still match current writing?
- Update after major style shifts
- Test new outputs against recent work, not archived samples
One-Size-Fits-All
Using the same spec for all contexts:
- Problem: LinkedIn posts sound like academic papers
- Fix: Create context-specific variants (Professional, Social, Academic)
- Maintain core voice across variants; adjust formality and structure
- Gather our analysis: Collect the stylometric patterns we identified in Article 2
- Complete all six sections: Voice attributes, sentence patterns, vocabulary, structure, rhetorical moves, and samples
- Choose a platform: Claude for tonal nuance, ChatGPT for vocabulary adherence
- Test with 3-5 pieces: Generate content on different topics to validate
- Iterate on failures: Track what doesn't match and refine the specification
What's Next
The analysis from Article 2 combined with the specification framework from this article provides the foundation. The final piece: putting it together into a complete workflow that takes writers from idea to polished draft, with AI maintaining voice throughout.
Article 4 focuses on evaluating AI output, identifying failure patterns, and iterating on specifications over time. Testing specifications against real writing samples—noting where AI succeeds and where it fails—prepares us for the evaluation strategies covered next.