Writer Skill

Generate content in your authentic voice by analyzing writing samples and applying learned voice patterns through adaptive interviewing and voice-matched generation.

Philosophy

Key Innovation: This isn't a generic AI writer - it's YOUR voice, learned from YOUR writing.

Unlike generic AI content generation, this skill:

• Learns from 6 distinct voice profiles (work email, personal email, blog, LinkedIn, Twitter, reports)
• Interviews you like a Pulitzer-winning journalist to deeply understand your message
• Validates output against your authentic patterns (not generic "good writing")
• Saves versioned drafts to Obsidian for iteration and history

Simplicity: Geoffrey's native LLM capabilities analyze your voice. Scripts just fetch samples. No complex NLP libraries - clean, maintainable architecture.

When to Activate

Use this skill when you need to:

• Write emails that sound like you (not generic AI)
• Draft blog posts in your established voice
• Create social media content consistent with your style
• Generate reports matching your professional tone
• Ensure content authentically represents your perspective and voice

DON'T use for:

• Quick factual responses (use native Claude)
• Content in someone else's voice
• Generic templates

Voice Profiles

Six distinct profiles capture different contexts:

Confidence Levels:

• High (0.85+): 50+ samples - voice well established
• Moderate (0.70-0.84): 20-49 samples - patterns emerging
• Low (<0.70): <20 samples - insufficient data, warn user

6-Phase Workflow

Phase 1: Voice Profile Selection & Validation

Goal: Load the right voice profile and verify it's ready

Process:

1. Detect content type from user request:
   • "write email" → email_work or email_personal (ask which)
   • "blog post" → blog_technical
   • "LinkedIn post" → social_linkedin
   • "tweet" → social_twitter
   • "report" → report_formal
2. Load voice profile:

   cat ~/Library/Mobile\ Documents/com~apple~CloudDocs/Geoffrey/knowledge/writing-voice.json

3. Check confidence and sample count:
   • If confidence ≥ 0.85: Proceed with confidence
   • If confidence 0.70-0.84: Proceed with note
   • If confidence < 0.70: Warn user, offer options

Low Confidence Response:

Only {sample_count} {content_type} samples (confidence: {confidence}).
Target: {target_samples}+ samples for 0.85+ confidence.

Options:
1. Gather more samples (recommended)
2. Use {alternative_profile} voice as proxy (confidence: {alt_confidence})
3. Proceed anyway (may not fully match your voice)

Preference?

Checkpoint 1: User confirms profile or gathers more samples

Phase 2: Deep Interview (Adaptive)

Goal: Deeply understand what you want to communicate

Process:

1. Load interview questions from templates/interview-questions.json
2. Adapt depth to content type:
   • Tweet/Social: 3-5 questions (~2-3 min)
   • Email Quick: 4 questions (~3 min)
   • Email Complex: 8 questions (~5 min)
   • Blog: 8-12 questions (~5-8 min)
   • Report: 15+ questions (~10-15 min)
3. Interview like a journalist:
   • Ask follow-up questions for vague answers
   • Probe for examples, data, evidence
   • Clarify ambiguities
   • Build structured understanding

Email Interview Example (complex):

1. Who are the recipient(s) and what's the context?
   → [User answers]

2. What's the primary goal?
   → [User answers]
   [If vague: "Can you be more specific about the outcome you want?"]

3. What background do they need?
   → [User answers]

4. What are the key points? (Max 3-5)
   → [User answers]
   [If >5: "Let's focus on the most critical 3-5 points."]

... continue through 8 questions

Blog Interview Example:

1. What's your central message?
2. Why does this matter right now?
3. What should readers do/think/feel after?
4. What evidence supports this?
5. Any stories or case studies?
6. Main counterarguments?
7. How should it open?
8. Key sections (3-5)?
9. How should it close?
10. Primary audience and technical level?
11. Desired tone?

Checkpoint 2: Review interview summary, add missing details

Phase 3: Content Planning & Outline

Goal: Create a clear structure before writing

Process:

1. Create outline based on:
   • Interview responses
   • Content type structure (email: intro + points + ask, blog: hook + body + conclusion)
   • Voice profile preferences (e.g., blog uses chronological flow, staff quotes)
2. Apply voice-specific patterns:
   • Email: Opening pattern, bullet points for key points, clear ask
   • Blog: Subheadings for scannability, narrative flow, reflective close
   • Social: Hook first line, call-to-action
3. Suggest length from profile:
   • Email avg: Check profile examples
   • Blog avg: Check profile examples
   • Social: Platform limits (LinkedIn 1300 chars, Twitter 280 chars)

Email Outline Example:

## Email Plan

**Subject**: [Based on purpose]

**Opening**: [Use opening pattern from profile, e.g., "I've reviewed {topic}..."]

**Body**:
- Point 1: [From interview]
- Point 2: [From interview]
- Point 3: [From interview]

**Ask**: [Specific call-to-action from interview]

**Closing**: [Use closing pattern from profile, e.g., "Let me know if..."]

**Tone**: Direct, professional, action-oriented
**Length**: ~200 words (typical for email_work)

Blog Outline Example:

## Blog Plan

**Title**: [Derived from central message]

**Structure**:
1. Opening: [Hook strategy - context-setting or reflective quote]
2. Background & Context
3. [Main section 1]
4. [Main section 2]
5. [Main section 3]
6. Impact & Results
7. Closing: [Forward-looking reflection]

**Voice elements**:
- Staff quotes (use throughout)
- Chronological narrative flow
- Subheadings for scannability
- Accessible but authoritative tone

**Length**: ~850 words (typical for blog_technical)

Checkpoint 3: User approves outline

Phase 4: Draft Generation

Goal: Write content that authentically sounds like you

Process:

1. Load identity context:

   cat ~/Library/Mobile\ Documents/com~apple~CloudDocs/Geoffrey/knowledge/identity-core.json

   Use for personality alignment:

   • communication_preferences.style → Direct, no-nonsense, concise
   • decision_framework → Analytical, evidence-based
   • telos_summary.top_3_values → Equity, excellence, empathy
2. Apply voice characteristics as generation guidelines:
   • Tone: From voice_characteristics.tone
   • Sentence structure: From voice_characteristics.sentence_structure
   • Vocabulary: Common terms, avoided terms
   • Formatting: Bullet points, markdown, emphasis patterns
   • Opening: Use patterns from opening_patterns
   • Closing: Use patterns from closing_patterns
   • Argument structure: Lead with conclusion, use evidence
3. Write content naturally:
   • Use Geoffrey's native writing capabilities
   • Guided by voice profile (not rigid pattern matching)
   • Weave in interview content (examples, data, stories)
   • Match typical length from profile examples
4. Reference example emails/posts from profile for inspiration

Generation Approach (Hybrid):

• Write naturally with rich voice context
• Don't mechanically apply patterns
• Let voice characteristics inform style, not constrain creativity

Phase 5: Validation & Refinement

Goal: Ensure draft authentically matches your voice

Native LLM Validation (no script needed):

1. Geoffrey self-validates by comparing draft to voice profile:
   • Read the draft I just generated
   • Read the voice profile (voice_characteristics + examples)
   • Compare tone, structure, vocabulary, formatting
   • Score alignment 0-100
2. Scoring criteria (weighted):
   • Tone match (20%)
   • Sentence structure - length, complexity (15%)
   • Vocabulary - common/avoided terms (15%)
   • Formatting - bullets, markdown, emphasis (10%)
   • Opening pattern match (10%)
   • Closing pattern match (10%)
   • Argument structure - conclusion placement, evidence use (10%)
   • Stylistic markers - contractions, punctuation, framing (10%)
3. Identify specific deviations:
   • "Sentences 15% longer than typical (avg 16 words vs typical 14)"
   • "Missing characteristic bullet points (you use them 67% of time)"
   • "Tone slightly more formal than usual"
4. Scoring thresholds:
   • 85-100: Excellent match, present draft
   • 70-84: Good match with minor deviations, note them
   • <70: Significant deviations, propose refinements

Refinement (if score < 85):

Voice Validation: {score}/100

Issues:
- {Issue 1 with specific evidence}
- {Issue 2 with specific evidence}
- {Issue 3 with specific evidence}

Should I refine to better match your voice?

If user approves refinement:

• Apply specific fixes identified
• Re-validate
• Present refined draft

Checkpoint 4: User accepts, requests refinement, or manually edits

Phase 6: Storage & Versioning

Goal: Save draft to Obsidian with proper versioning

Process:

1. Determine file path:

   ~/Library/Mobile Documents/iCloud~md~obsidian/Documents/Personal_Notes/Geoffrey/Writing/{ContentType}/{date}-{slug}.md

   Content types:

   • Blog → Geoffrey/Writing/Blog/
   • Email → Geoffrey/Writing/Email/
   • Social → Geoffrey/Writing/Social/
   • Reports → Geoffrey/Writing/Reports/
2. Generate frontmatter:

   ---
   created: 2025-12-06
   type: blog
   profile: blog_technical
   topic: AI implementation
   status: draft
   version: 1.0
   validation_score: 87
   word_count: 847
   tags: [geoffrey, writing, ai]
   source: geoffrey-writer
   interview_date: 2025-12-06
   ---

3. Save file using Obsidian MCP tools:

   mcp__obsidian-vault__create_vault_file

4. Also return text directly (copy-paste ready)
5. Offer to open in Obsidian:

   mcp__obsidian-vault__show_file_in_obsidian

Versioning:

• Initial: version: 1.0
• Refinements: version: 1.1, 1.2
• Major rewrites: version: 2.0

Output to user:

## Draft Complete

**Validation Score**: 87/100 ✓
**Word Count**: 847
**Profile**: blog_technical

**Saved to**: Geoffrey/Writing/Blog/2025-12-06-ai-implementation.md

[Full draft text here]

Actions:
1. Open in Obsidian
2. Request refinements
3. Mark as final

Voice Learning Process

Initial Setup (Before First Use)

Required: Gather writing samples for each profile you'll use

Step 1: Email Samples (3 accounts)

# PSD work emails
bun scripts/extract-email-samples.js \
  --account psd \
  --date-range "2024-06-01:2025-12-06" \
  --max-samples 100 \
  --output "/tmp/email-samples-psd.json"

# Personal emails
bun scripts/extract-email-samples.js \
  --account kh \
  --date-range "2024-06-01:2025-12-06" \
  --max-samples 50 \
  --output "/tmp/email-samples-kh.json"

# Business/consulting emails
bun scripts/extract-email-samples.js \
  --account hrg \
  --date-range "2024-06-01:2025-12-06" \
  --max-samples 50 \
  --output "/tmp/email-samples-hrg.json"

Step 2: Blog Samples

# Launch Geoffrey Chrome (for personal blog access)
cd ~/non-ic-code/geoffrey/skills/browser-control
./scripts/launch-chrome.sh

# Extract blog posts
cd ~/non-ic-code/geoffrey/skills/writer
bun scripts/extract-blog-samples.js \
  --urls "https://psd401.ai/blog/stanford-ai-tinkery-2025,https://psd401.ai/blog/why-this-website,https://blog.krishagel.com/post1,https://blog.krishagel.com/post2" \
  --output "/tmp/blog-samples.json"

Step 3: Social Media Samples (requires login)

# IMPORTANT: Manually log into LinkedIn and X in Geoffrey Chrome first

# LinkedIn
bun scripts/extract-social-samples.js \
  --platform linkedin \
  --profile-url "https://linkedin.com/in/krishagel" \
  --max-posts 50 \
  --output "/tmp/social-linkedin.json"

# Twitter/X
bun scripts/extract-social-samples.js \
  --platform twitter \
  --profile-url "https://x.com/KrisHagel" \
  --max-posts 50 \
  --output "/tmp/social-twitter.json"

Step 4: Geoffrey Analyzes

Tell Geoffrey:

Analyze my writing samples and create voice profiles.

Email samples: /tmp/email-samples-psd.json, /tmp/email-samples-kh.json, /tmp/email-samples-hrg.json
Blog samples: /tmp/blog-samples.json
Social samples: /tmp/social-linkedin.json, /tmp/social-twitter.json

Geoffrey will:

1. Read all sample files
2. Analyze each profile's:
   • Tone and voice
   • Sentence structure patterns
   • Vocabulary (common/avoided terms)
   • Formatting preferences
   • Opening/closing patterns
   • Argument structure
   • Distinctive markers
3. Select representative examples
4. Calculate confidence scores
5. Write to writing-voice.json:

   ~/Library/Mobile Documents/com~apple~CloudDocs/Geoffrey/knowledge/writing-voice.json

Estimated time: 30-45 min initial setup

Integration Points

knowledge-manager → Identity Context

When: Phase 4 (Draft Generation)

Purpose: Ensure voice reflects personality (analytical, direct, impact-driven)

How:

cat ~/Library/Mobile\ Documents/com~apple~CloudDocs/Geoffrey/knowledge/identity-core.json

Use:

• communication_preferences → Tone, style, dislikes
• decision_framework → Analytical, evidence-based approach
• telos_summary → Values (equity, excellence, empathy), mission
• strengths_essence → Input (collector), Significance (impact), Analytical

google-workspace → Email Samples

When: Initial voice creation + quarterly refresh

Scripts: search_messages.js, read_message.js

Process:

1. Search sent emails
2. Filter automated, short, quoted
3. Clean signatures, HTML
4. Extract plain text + metadata
5. Save to JSON for Geoffrey to analyze

browser-control → Blog/Social Samples

When: Initial voice creation + ad-hoc updates

Prerequisites: Geoffrey Chrome running with logins

Process:

1. Navigate to blog/profile URL
2. Extract article/post content
3. Handle scrolling/lazy-loading (social)
4. Parse and clean content
5. Save to JSON

obsidian-manager → Draft Storage

When: Phase 6 (Storage)

Folder structure:

Geoffrey/Writing/
├── Blog/
├── Email/
├── Social/
└── Reports/

File format: YYYY-MM-DD-{slug}.md with frontmatter

Edge Cases & Handling

Insufficient Samples

Scenario: User requests email, but only 12 samples (confidence 0.62)

Response:

Only 12 email_work samples (0.62 confidence). Target: 50+ for 0.85+ confidence.

Options:
1. Gather ~40 more email samples from recent sent (recommended)
2. Use email_personal as proxy (if similar, confidence: 0.88)
3. Proceed with low confidence (may not fully match your voice)

Note: Voice profiles improve with more samples. I can still write, but authenticity may vary.

Preference?

Voice Drift (Outdated Profile)

Scenario: Profile last updated 18 months ago

Detection: Check last_updated in writing-voice.json

Response:

Your email_work profile is 18 months old (last updated: 2024-06-15).
Writing styles evolve over time.

Recommendation: Refresh with recent samples (past 6 months)

Run:
bun scripts/extract-email-samples.js --account psd --date-range "2024-06-01:2025-12-06" --max-samples 100 --output "/tmp/email-refresh.json"

Then ask me to analyze and update the profile.

Proceed with current profile or refresh first?

Contradictory Samples

Scenario: Email samples vary significantly (45% formal, 35% casual, 20% mixed)

Geoffrey analyzes and responds:

Your email_work voice varies significantly across samples:
- 45% formal (board, superintendent, external)
- 35% casual (staff, peers, project teams)
- 20% mixed

Options:
1. Split into email_work_formal and email_work_casual profiles
2. Choose dominant style (formal) as default
3. Let me dynamically adjust based on recipient (you specify each time)

Preference?

Low Validation Score

Scenario: Draft scores 62/100

Response:

Voice Validation: 62/100 ⚠️

Issues:
- Tone too formal (target: direct, action-oriented)
  Example: "It would be appreciated if you could consider..." vs typical "Please review and confirm."

- Sentences 35% longer (avg 19 words vs typical 14)
  Example: Draft paragraph 2 avg 21 words, your emails avg 14 words.

- Missing bullet points (you use bullets 67% of the time for key points)
  Draft uses prose for 5 key points, you typically bullet them.

Should I refine to better match your voice?

New Content Type

Scenario: User requests "write a press release"

Response:

I don't have a press_release voice profile yet.

Options:
1. Use report_formal profile as starting point (confidence: 0.83)
2. Provide 5-10 past press releases for me to analyze
3. Write without voice matching (generic professional press release style)

Note: If you provide samples, I can create a new press_release profile.

Preference?

Usage Examples

Example 1: Quick Work Email

User: "Write email to my team about next week's leadership meeting"

Geoffrey:

I'll use your email_work profile (confidence: 0.92, 156 samples).

Quick interview (4 questions):

1. Who's the team and what's your relationship?
   → [User: Direct reports - IT leadership team]

2. Purpose? (inform, request, respond, update)
   → [User: Inform and request prep]

3. Key information?
   → [User: Meeting Tuesday 2pm, agenda items: budget, staffing, AI roadmap. Need their input on AI priorities.]

4. Tone?
   → [User: Professional but casual - these are my directs]

[Generates outline] → User approves

[Generates draft, validates (score: 89), presents to user]

Example 2: Blog Post

User: "Write blog post about our AI implementation journey"

Geoffrey:

I'll use your blog_technical profile (confidence: 0.78, 4 samples).

Note: Only 4 blog samples. Recommend 6+ more for 0.85+ confidence.
Proceed or gather more samples first?

→ [User: Proceed]

Standard interview (11 questions):

1. Central message?
   → [User: ...]

2. Why now?
   → [User: ...]

[... continues through 11 questions ...]

[Creates detailed outline] → User approves

[Generates ~850 word draft with staff quotes, chronological flow, reflective close]

[Validates (score: 87), presents draft]

[Saves to Geoffrey/Writing/Blog/2025-12-06-ai-implementation-journey.md]

Example 3: LinkedIn Post

User: "LinkedIn post about Stanford AI Tinkery visit"

Geoffrey:

I'll use your social_linkedin profile (confidence: 0.91, 48 posts).

Quick interview (5 questions):

1. Main point (1-2 sentences)?
   → [User: Visited Stanford's AI Tinkery with district team. Collaborative learning happening everywhere, not just at elite institutions.]

2. Target audience?
   → [User: Ed tech leaders, superintendents]

3. Tone?
   → [User: Inspirational, reflective]

4. Call-to-action?
   → [User: Encourage others to build their own learning networks]

5. Link or visual?
   → [User: No link, maybe photo but I'll add later]

[Generates ~180 word LinkedIn post]

[Validates (score: 92), presents]

Limitations

Known Constraints

1. Sample Dependency: Voice quality directly tied to sample quantity and quality
   • Low samples (<20) = low accuracy
   • Poor quality samples = poor voice matching
2. Voice Drift: Styles evolve over time
   • Recommend quarterly refresh (every 3 months)
   • Auto-warn if profile >12 months old
3. New Contexts: No profile for content types you haven't written before
   • Press releases, technical docs, grant proposals require new samples
   • Can adapt existing profiles but may not match perfectly
4. Platform Nuances: Social media scraping limitations
   • LinkedIn/X may change selectors
   • Requires manual login
   • May violate ToS (user accepted risk)
5. Multilingual: Currently English only
   • No Spanish email profile (even though you write Spanish emails)
   • Could add with Spanish samples
6. Collaborative Writing: Single voice only
   • Can't blend your voice + co-author
   • Future enhancement

Validation & Quality

Success Metrics

Quantitative:

• Voice alignment: 85+ for established profiles (0.85+ confidence)
• Sample size: 50+ emails, 10+ blogs, 30+ social per profile
• User acceptance: >80% drafts accepted with minor/no edits
• Refinement iterations: <2 average

Qualitative:

• User feedback: "This sounds like me"
• Consistent voice across topics within profile
• Authentic, not mechanical or formulaic

Quality Checklist

Before presenting draft:

• ✓ Matches tone from voice profile
• ✓ Sentence structure within 20% of typical length
• ✓ Uses common vocabulary, avoids avoided terms
• ✓ Follows formatting preferences (bullets, markdown)
• ✓ Opening/closing match typical patterns
• ✓ Argument structure aligned (conclusion placement, evidence use)
• ✓ Validation score ≥85 (or explained deviations if <85)
• ✓ Word count within typical range for content type

Future Enhancements (Out of Scope)

1. Real-time Learning: Auto-analyze new sent emails monthly
2. A/B Testing: Generate 2 versions, learn from user choices
3. Collaborative Voice: Blend user + co-author voices
4. Multi-language: Spanish voice profiles
5. Performance Tracking: Voice drift detection over time
6. Audience-Aware: Sub-profiles (email_work_superintendent vs email_work_staff)
7. Feedback Loop: User rates content, Geoffrey learns from scores

Troubleshooting

"No voice profile found"

Cause: writing-voice.json doesn't exist or profile type missing

Solution:

1. Check if file exists:

   cat ~/Library/Mobile\ Documents/com~apple~CloudDocs/Geoffrey/knowledge/writing-voice.json

2. If missing: Run initial setup (extract samples + ask Geoffrey to analyze)
3. If file exists but profile missing: Add samples for that content type

"Low confidence warning"

Cause: Fewer than 20 samples for profile

Solution:

1. Gather more samples using extraction scripts
2. Ask Geoffrey to re-analyze with new samples
3. Or: Proceed with caveat that voice may not fully match

"Geoffrey Chrome not running"

Cause: Social/blog extraction requires Geoffrey Chrome with remote debugging

Solution:

cd ~/non-ic-code/geoffrey/skills/browser-control
./scripts/launch-chrome.sh

"Failed to extract social posts"

Cause: Not logged into LinkedIn/X, or selectors changed

Solution:

1. Open Geoffrey Chrome manually
2. Navigate to LinkedIn/X and log in
3. Verify posts load
4. Re-run extraction script
5. If still fails: Platform may have changed selectors (update script)

Notes

• Progressive Disclosure: Load only the profile needed for current task (not all 6)
• Identity Alignment: Always cross-reference identity-core.json for personality consistency
• Checkpoints: 4 user approval points prevent wasted effort
• Versioning: All drafts saved to Obsidian with version numbers
• Native Analysis: Geoffrey's LLM capabilities analyze voice - no complex libraries needed