---

name: humanize description: Transform AI-generated text to match user's voice and communication style. Use when drafting messages, emails, documents, or any external communication that should sound like the user wrote it. user-invocable: true

Humanize

Purpose

Make AI-generated content indistinguishable from what the user would write themselves. Every executive has a distinctive voice - word choices, sentence structure, level of formality, use of specific phrases. This skill learns and applies that voice.

Data Locations

• reference/system/patterns.md - Learned communication patterns
• me/writing-samples/ - Example communications from user (if provided)
• Memory graph - Accumulated style observations

Workflows

Rewrite

When: User asks to humanize text, draft communications, or make AI output "sound like me"

Input:

• Text to humanize
• Target audience (optional - affects formality calibration)
• Purpose/context (optional - affects tone)
• User's style patterns from reference/system/patterns.md

Process:

1. Analyse the input text for AI-typical patterns:

   • Overly formal transitions ("Furthermore", "Additionally")
   • Excessive hedging ("It might be worth considering")
   • Generic corporate language
   • Unnaturally balanced structure

2. Load user's known patterns from memory/patterns.md:

   • Preferred phrases and expressions
   • Typical sentence length
   • Formality level
   • Common openers/closers

3. Rewrite applying user's voice:

   • Replace AI phrases with user equivalents
   • Adjust sentence structure to match user's rhythm
   • Add/remove formality as appropriate
   • Include user's characteristic expressions

4. Verify it sounds human:

   • Read aloud mentally - does it flow naturally?
   • Would someone who knows the user recognise this as theirs?
   • Are there any "tells" remaining?

Output: Humanized text that sounds like the user wrote it, with brief note on changes made (if helpful).

Verification:

• No obvious AI-isms remaining
• Matches user's typical formality level
• Sentence structure varied naturally
• Characteristic expressions included where appropriate
• Length appropriate for user's style

Next: User reviews, provides feedback if adjustments needed

---

Extract Style

When: User provides writing samples, or corrects AI output to "sound more like me"

Input:

• Sample text written by user
• OR correction/feedback on AI-generated text

Process:

1. Analyse the sample for distinctive patterns:

   • Vocabulary: What words does user favour? Avoid?
   • Sentence structure: Short and punchy? Complex with clauses?
   • Formality: Casual, professional, varies by context?
   • Openings/closings: How do they start/end messages?
   • Expressions: Signature phrases, idioms, preferences?

2. Compare to AI defaults:

   • What would AI typically say vs what user says?
   • Identify specific transformations

3. Document patterns:

   • Add to reference/system/patterns.md
   • Add observations to memory graph

4. Confirm with user:

   • "I noticed you prefer X over Y - is that right?"

Output: Updated patterns documentation with newly extracted style elements.

Verification:

• At least one new pattern identified
• Pattern documented with example
• User confirmed accuracy (if asked)

Next: Apply patterns in future humanize requests

---

Draft Message

When: User asks to draft an email, Slack message, or other communication

Input:

• Key points to communicate
• Recipient (person entity if in memory)
• Context/purpose
• Urgency level

Process:

1. Check recipient in memory graph for:

   • Communication preferences
   • Relationship context
   • Previous interaction patterns

2. Determine appropriate tone:

   • Direct report: supportive but direct
   • Manager: concise, outcome-focused
   • Peer: collaborative, casual-professional
   • External: more formal, clear context

3. Draft applying user's voice:

   • Use appropriate greeting for relationship
   • Structure for user's typical message format
   • Include user's characteristic expressions
   • Match formality to recipient

4. Keep it appropriately brief:

   • Most executives prefer shorter messages
   • Lead with the ask/point
   • Background only if needed

Output: Ready-to-send message draft in user's voice.

Verification:

• Appropriate for recipient relationship
• Key points all covered
• Length appropriate (not overlong)
• Sounds like user, not AI
• Clear call-to-action if relevant

Next: User reviews, sends or requests adjustments

Related Files

• reference/system/patterns.md - Learned style patterns
• people/{person}.md - Recipient context
• Memory graph - Relationship and preference data

Best Practices

• Capture corrections: When user says "I wouldn't say it like that," extract the pattern
• Context matters: Same user might be casual with team, formal with board
• Less is more: Executives usually prefer concise over comprehensive
• Personality over polish: A few rough edges can sound more authentic
• Test with feedback: Periodically ask "did that sound like you?"

Common AI-isms to Avoid

Style Pattern Template

When documenting a pattern in reference/system/patterns.md:

Copy### {Pattern Name}

**Instead of:** {AI default}
**User prefers:** {User's version}
**Context:** {When this applies}
**Example:** {Real example if available}

Example:

Copy### Email Sign-offs

**Instead of:** "Best regards," or "Please let me know if you have any questions."
**User prefers:** "Thanks," or just name
**Context:** All internal emails, most external
**Example:** "Thanks, Sarah" or just "- S" for close colleagues