The output becomes infrastructure -- a reference document used every time you work with AI to maintain your authentic voice instead of producing generic content.
<quick_start> Provide 3-5 writing samples where your voice feels strongest (500-2000 words each). The skill will:
- Analyze patterns across all samples
- Identify your distinctive voice markers
- Generate a VOICE.md style guide
- Create a forbidden phrases list specific to your anti-patterns
- Provide testing prompts to validate the guide
Ideal samples: Newsletter issues, blog posts, emails you're proud of, social media threads that felt "like you"
Avoid: Heavily edited corporate content, collaborative pieces, anything that felt forced </quick_start>
<sample_gathering_guidance> <what_makes_good_samples> Include samples that:
- You wrote when feeling confident and natural
- Received feedback like "this sounds just like you"
- You'd be happy to write again
- Show your voice across different contexts (casual, professional, explanatory)
- Are at least 500 words (longer is better for pattern detection)
Avoid samples that:
- Were heavily edited by others
- Feel generic or corporate even to you
- Were written under heavy constraints
- Don't represent how you want to sound going forward </what_makes_good_samples>
<minimum_requirements>
- Minimum: 3 samples, 500+ words each
- Ideal: 5-7 samples, 1000+ words each
- Advanced: 10+ samples including different formats (email, long-form, social)
More samples = more accurate pattern detection, but diminishing returns after 10. </minimum_requirements>
<alternative_sources> If you don't have strong writing samples:
Voice memo transcription: Record yourself explaining something you know well. Transcribe it. Your speaking patterns often reveal writing voice better than forced writing.
Email archaeology: Search sent mail for messages where you were being genuinely helpful, not performing professionalism.
Chat/Slack messages: Longer explanatory messages often show natural voice patterns.
Old social media: Before you started "performing" for algorithms. </alternative_sources> </sample_gathering_guidance>
<analysis_framework> When analyzing samples, examine these dimensions:
<output_format> Generate a VOICE.md file with this structure:
# Voice Profile: [Name]
Generated: [Date]
Based on: [X] writing samples ([total word count] words)
## Voice Summary
[2-3 sentence description of overall voice character]
## Core Voice Characteristics
### Sentence Patterns
- Average length: [X] words
- Variation: [Low/Medium/High]
- Notable patterns: [specific observations]
### Vocabulary Fingerprint
- Formality: [Casual/Conversational/Professional/Formal]
- Characteristic phrases: [list]
- Words to use freely: [list]
### Rhythm and Flow
- Paragraph style: [description]
- Transition patterns: [description]
- Pacing notes: [description]
### Tone Markers
- Primary tone: [description]
- Humor style: [description]
- Reader relationship: [description]
### Structural Habits
- List vs. prose preference: [description]
- Formatting patterns: [description]
### Opinion Expression
- Directness level: [1-10]
- Qualification style: [description]
- Authority stance: [description]
## The Forbidden List
### Never Use (These kill your voice)
- [phrase 1]
- [phrase 2]
- [etc.]
### Use Sparingly (Context-dependent)
- [phrase 1] - only when [context]
- [etc.]
### Watch for Clusters (OK alone, problematic together)
- [pattern description]
## Voice Maintenance
### Monthly Check
- Read 3 recent pieces aloud
- Do they still sound like you?
- Update this guide if voice has evolved
### Quarterly Refresh
- Gather new strong samples
- Re-run analysis
- Compare to this guide
- Update patterns that have changed
## Testing Prompts
Use these to validate the guide works:
**Test 1 - Short form:**
"Using my voice profile, write a 3-sentence response to [common scenario in your field]"
**Test 2 - Long form:**
"Using my voice profile, write the opening 2 paragraphs for a piece about [topic you know well]"
**Test 3 - Edge case:**
"Using my voice profile, write about [topic outside your usual content]"
Compare outputs to your natural writing. If they feel off, update the guide.
</output_format>
<analysis_process> Read all samples without analyzing. Get a feel for the overall voice impression. Note your gut reaction: what makes this writing distinctive?
Key findings:
- Sentences vary wildly (5-40 words)
- Heavy use of "Here's the thing" and "The reality is"
- Self-deprecating humor about own mistakes
- One-sentence paragraphs for emphasis
- Direct address to reader ("you")
- Specific numbers and timeframes ("three hours later, no hyperbole")
Forbidden list generated:
- "It's worth noting that"
- "In today's digital landscape"
- "Leverage" as a verb
- Paired adjectives ("comprehensive and thorough")
- Any sentence starting with "Furthermore" or "Moreover"
Key findings:
- Longer sentences with technical precision
- Code-switches between casual and technical
- Humor through understatement
- Heavy use of concrete examples
- "Actually" and "technically" as qualifiers
- Lists for technical content, prose for opinions
Forbidden list generated:
- Vague claims without specifics
- Marketing superlatives ("revolutionary," "game-changing")
- Excessive hedging ("it might be argued")
- Emoji in technical sections
- "Simply" before complex explanations
<integration_notes> Where to save VOICE.md:
- Project-specific:
./.claude/VOICE.md(for this project only) - User-wide:
~/.claude/voice/VOICE.md(for all projects) - Multiple profiles:
~/.claude/voice/[context]-voice.md(professional, casual, technical)
Using with other skills:
slop-detectorwill reference VOICE.md for personalized pattern detectionvoice-editoruses VOICE.md to guide rewrites- Include path to VOICE.md in project's CLAUDE.md for automatic loading
Maintenance schedule:
- Monthly: Quick read-aloud check (15 min)
- Quarterly: Full refresh with new samples (1 hour)
- After major voice shifts: Complete re-analysis </integration_notes>
<success_criteria> The voice analysis is complete when:
- All provided samples have been analyzed
- Each dimension in the analysis framework has findings
- VOICE.md file is generated with all sections
- Forbidden list contains at least 10 specific items
- 3 validation prompts are tailored to the user's context
- User has been shown where to save the file
- Testing process has been explained
Quality check: The generated guide should allow someone unfamiliar with the writer to produce content that readers would recognize as authentic. </success_criteria>
<core_principle> You're not trying to achieve perfection on attempt one. You're building infrastructure that improves with use. The first guide will be good but not perfect -- that's normal. Each piece written with this guide makes it more precise.
Voice doesn't live in first drafts. It lives in editing choices. The guide gives AI direction; your editing gives the work your actual voice. </core_principle>