Extract Voice

Voice Extractor Skill Guide

Purpose

Extract voice patterns from writing samples and generate production-ready prompts that replicate the author's unique style, tone, and structural patterns.

When to Use

• User provides writing samples and wants to replicate that voice
• User needs to create social media content prompts matching their style
• User wants to analyze what makes their writing distinctive
• User needs to refine existing prompts to produce more authentic output

Process

Phase 1: Voice Analysis

1. Collect Samples: Request 5-10 writing samples from the user (more variety = better extraction)

2. Analyze Patterns: For each sample, identify:

   Structural Patterns

   • Opening styles (personal story, direct address, observation, question, analogy)
   • Paragraph flow and transitions
   • Closing patterns (call to action, reflection, open question, statement)
   • Average length and length variation

   Tonal Patterns

   • Primary tones (inspirational, conversational, instructional, philosophical, humorous)
   • Emotional arc within posts
   • Formality level

   Linguistic Fingerprints

   • Signature phrases or sentence structures
   • Punctuation habits (em dashes, ellipses, question chains)
   • Arrow syntax usage (->)
   • List formatting preferences
   • Emoji/symbol usage (or absence)

   Content Patterns

   • How topics are introduced
   • Personal experience integration
   • Technical vs emotional balance
   • Multi-topic handling strategies

3. Create Voice Analysis Document: Generate a structured analysis file containing:

   ## VOICE DNA
   Core traits that appear in 80%+ of samples
   
   ## STRUCTURAL FORMATS
   List distinct formats with examples
   
   ## OPENING ROTATION
   All unique opening styles identified
   
   ## TONAL PALETTE
   Tones and when each appears
   
   ## LINGUISTIC MARKERS
   Signature elements to preserve
   
   ## ANTI-PATTERNS
   What this voice NEVER does
   

Phase 2: Prompt Generation

1. Draft Initial Prompt with sections:

   • Role/persona definition
   • Voice DNA (immutable traits)
   • Structural toolkit (multiple format options, NOT rigid formula)
   • Tone matching guidance
   • Length flexibility ranges
   • Anti-patterns (what to avoid)
   • 3-5 voice examples directly from samples

2. Key Principles:

   • Provide OPTIONS not requirements for structure
   • Include actual examples from the source material
   • Specify what the voice NEVER does
   • Allow length variation based on content
   • Handle multi-topic inputs gracefully

Phase 3: Iterative Refinement

1. Run Experiments: Use codex skill with gpt-5 to test prompts

2. Create Experiment Folders: example-1/, example-2/, etc.

3. Each Folder Contains:

   • PROMPT.md - The prompt version being tested
   • INPUT.txt - Test input(s)
   • OUTPUT.txt - Generated output
   • NOTES.md - What worked/didn't work

4. Iterate Until:

   • Outputs show structural variety (not all same format)
   • Opening styles rotate naturally
   • Multi-topic handling feels cohesive
   • Tone matches content appropriately
   • Length adapts to complexity

Output Artifacts

Required Deliverables

1. VOICE_ANALYSIS.md - Complete voice pattern breakdown
2. FINAL_PROMPT.md - Production-ready prompt
3. experiment-N/ folders - All iteration history

Optional Deliverables

• PREPROCESSING_REQUIREMENTS.md - If input cleaning needed
• PROBLEMS.md - Issues identified and solved
• EXAMPLES.md - Additional voice examples for reference

Common Issues and Solutions

Quick Start Template

When user provides samples, respond with:

I'll analyze your writing samples to extract your voice patterns.

**Analyzing:**
1. Structural patterns (openings, flow, closings)
2. Tonal patterns (primary tones, emotional arc)
3. Linguistic fingerprints (signature phrases, punctuation)
4. Content patterns (topic handling, personal integration)

After analysis, I'll create:
- Voice Analysis document
- Production-ready prompt
- Experiment iterations to verify quality

Shall I proceed with the analysis?

Integration with Codex

When testing prompts:

1. Use codex skill with gpt-5 model, medium reasoning effort
2. Test with varied inputs (single topic, multi-topic, different lengths)
3. Verify variety across multiple generations
4. Store all experiments for reference

Resume syntax for continued refinement:

echo "Generate post about [topic]" | codex exec --skip-git-repo-check resume --last 2>/dev/null