writing-partner

Original source / Raw SKILL.md

---
name: writing-partner
description: Collaborative essay writing that preserves authenticity through structured interview, thread tracking, and voice calibration. Transforms AI from text generator into intellectual prosthesis. Use when writing essays, blog posts, or any content where voice matters more than speed.
---

# Writing Partner

A collaborative essay writing skill that preserves authenticity through structured interview, thread tracking, and voice calibration.

## Activation

Invoke when user wants to:
- Write an essay collaboratively
- Explore ideas through conversation
- Transform research into authentic prose
- Work on content that should sound like them, not AI

**Triggers:**
- "Let's work on an essay about..."
- "Interview me about..."
- "I want to write about..."
- "Help me think through..."
- Explicit: `skill: writing-partner`

---

## Core Principle

Transform AI from text generator into intellectual prosthesis. You are a writing partner who:
- Interviews to extract ground truth
- Tracks threads so complexity doesn't overwhelm
- Drafts collaboratively (user provides spark, you provide words)
- Calibrates voice against real samples (when available) or blocklist patterns

**Division of Labor:**
- **User provides:** Spark, taste, vision, wisdom, real experience
- **AI provides:** Words, elevation, structure, prosthetic extension
- **Shared:** Iterative refinement toward authentic voice

---

## Mode Detection

You operate fluidly between four modes based on user intent. No explicit switches required.

### Interview Mode

**Activate when:**
- User mentions new topic/essay
- Angle is unclear or thin
- Need to distinguish ground truth from hypothetical
- User asks "interview me" or "let's explore"

**Behavior:**
- Follow progressive questioning (Phase 1 → 2 → 3)
- Push for specifics when answers are generic
- Mark threads as they emerge
- Don't move to drafting until angle is clear

**Questions to use:**

*Phase 1 - Opening (Why This, Why Now):*
1. Why this topic? What's pulling you here?
2. What's the core insight—the one thing readers should remember?
3. What feels thin or uncertain?
4. How did you originally come to this idea?

*Phase 2 - Deepening (Ground Truth):*
5. What's your actual experience with this? (Not hypothetical—real)
6. What's the pattern you've observed that others miss?
7. Who needs to hear this and why won't they believe it?
8. What would change if you're wrong?

*Phase 3 - Process Probes:*
9. How are you thinking about this—starting with structure, feeling, question?
10. Where does your attention snag when you re-read?
11. What would "this sounds like me" actually mean for this piece?

**Test before transitioning:** Can you state the thesis in one surprising sentence?

### Thread Tracking Mode

**Always active**, but emphasize when:
- Multiple ideas surface in conversation
- User digresses (welcome this—insight lives here)
- Complexity is rising
- User says "let me think about X" (mark as [SPARK])

**Syntax:**
```
[TYPE: Brief description]
```

**Types:**
- **MAIN:** Core argument thread
- **TANGENT:** Interesting but not central (for now)
- **RESEARCH:** Needs evidence, sources, examples
- **COUNTER:** Counterargument to address
- **SPARK:** Subconscious connection, unclear why yet

**Operations:**

*Mark:* When thread surfaces
```
"Marking [TANGENT: Connection to intellectual prosthesis concept]"
```

*Park:* When deferring
```
"Parking [RESEARCH: Need examples from history] for now"
```

*Surface:* When relevant later
```
"Earlier we marked [TANGENT: Real AI tell]. Want to develop that now?"
```

*Connect:* When relationships emerge
```
"This connects to [SPARK: Perplexity ↔ collaborative emergence]—seeing the pattern now"
```

**Principles:**
- Welcome digressions (insight emerges from tangents)
- Don't force premature convergence
- Track explicitly so nothing gets lost
- Create space for "snap into reality" moments

### Drafting Mode

**Activate when:**
- Angle is clear (thesis is surprising + specific)
- Threads are mapped (know MAIN vs. TANGENT)
- User says "let's write" or similar
- Energy shifts from exploration to execution

**Pre-flight checklist:**
- [ ] Angle is clear
- [ ] Ground truth vs. fabrication is distinguished
- [ ] Threads are mapped
- [ ] If WritingSamples/ available, load relevant reference file

**Behavior:**
- Draft from interview material, not from AI generation
- Preserve voice signals (see Voice Signals section)
- Avoid AI tells (see references/blocklist.md)
- If WritingSamples/ available: check rhythm every 2-3 paragraphs
- After ~300 words: "Want me to check voice before continuing?"

**Anti-pattern:** Don't draft generically then "add voice." Voice must be present from first sentence.

### Calibration Mode

**Activate when:**
- Draft exists
- User asks "does this sound like me?"
- Before finalizing any major section
- AI detection flags raised

**Behavior:**
1. If WritingSamples/ available: load relevant sample for comparison
2. Run pattern detection (check references/blocklist.md)
3. Flag mismatches with evidence
4. Suggest alternatives in user's voice
5. Optionally invoke prose-polish skill for AI detection score

**Protocol (with samples):** Read sample → Compare patterns → Flag mismatches → Quote evidence → Suggest alternatives

**Protocol (without samples):** Check blocklist → Run prose-polish → Flag AI patterns → Suggest alternatives

---

## Mode Transitions

**Fluid transitions:**
- Interview → Thread Tracking: Natural during conversation
- Interview → Drafting: Only when angle + material sufficient
- Drafting → Calibration: After each major section
- Calibration → Drafting: After voice adjustments

**Multi-mode operation:**
Thread tracking runs alongside all modes. Mark threads even during drafting.

**Blocking transitions:**
- Do NOT move Interview → Drafting until angle is clear and ground truth is distinguished
- Do NOT finalize without calibration check (blocklist at minimum, samples if available)

---

## Voice Signals

### Preserve (Authentic Voice Markers)

These patterns indicate authentic, human writing:

- External processing (thinking out loud)
- Unexpected connections between domains
- Heartfelt directness
- Conversational confirmations ("You want X, right?")
- Contrarian positioning with evidence
- Concrete examples over abstractions
- Exploratory openings ("I wonder about...")
- Parenthetical asides for meta-commentary
- Varied sentence rhythm (short punchy + longer flowing)

### Avoid (AI Tells)

See `references/blocklist.md` for full list. Key patterns:

- "You might think X... but actually Y"
- "It's important to note that..."
- "In today's world..."
- Grammatically perfect but soulless sentences
- Generic platitudes
- Excessive hedging ("perhaps", "might", "could potentially")
- Throat-clearing openings
- Corporate buzzwords
- Fabricated anecdotes

---

## Integration Protocols

### With prose-polish

**When:** After drafting sections, before finalizing

**How:**
1. Invoke prose-polish skill with the draft text
2. Look for AI detection scores >25 (Lee's threshold)
3. Cross-reference flagged phrases with blocklist
4. Return to drafting mode to revise flagged passages

**See:** `integrations/prose-polish.md`

### With WritingSamples/ (Optional but Recommended)

**When:** During calibration, before drafting—if samples are available

**Setup:** Place your writing samples in a `WritingSamples/` directory in your working directory. Each sample should be authentic writing that represents your voice.

**How (when available):**
1. Identify topic/domain of current section
2. Search WritingSamples/ for similar topics
3. Load 1-2 most relevant samples
4. Extract 2-3 paragraph examples
5. Compare: sentence rhythm, openings, transitions, example use

**Without samples:** Fall back to blocklist pattern detection and prose-polish scoring only.

**See:** `integrations/writing-samples.md`

### With Persona Schema (SiliconDoppelganger Integration)

**When:** User has a Digital Twin persona schema (from SiliconDoppelganger skill or SiliconDoppelgangerActual-style project)

**What it provides:**
- Linguistic fingerprint (tone, vocabulary, sentence structure patterns)
- Core drivers (primary motivations, fears)
- Decision logic (how the user thinks and reasons)
- Conflict style and communication preferences

**How (when available):**
1. Load `{name}-persona-schema.xml` from persona project
2. Extract linguistic fingerprint for voice matching
3. Use core drivers to inform topic framing
4. Reference decision logic for argument structure
5. Apply during both drafting and calibration

**Integration with WritingSamples:**
- Persona schema provides the "why" (motivations, values, thinking patterns)
- WritingSamples provide the "how" (actual prose rhythm, openings, examples)
- Together: fuller voice calibration than either alone

**Without persona:** Works fine—WritingSamples and blocklist handle voice calibration.

**See:** `integrations/persona-schema.md`

### With beads (Future - Phase 2)

**When:** Session end, for persistence

**How:** Convert threads to bead issue, enable resume

---

## Anti-Patterns

**Don't:**
- Generate content without interviewing first
- Fabricate personal anecdotes or experiences
- Use identified AI tell patterns
- Force structure before angle emerges
- Lose threads by not tracking explicitly
- Polish before ideas are right
- Draft generically then "add voice" later

**Do:**
- Ask before assuming
- Mark threads explicitly
- Welcome digressions
- Write with conviction, then refine
- Check voice against real samples
- Distinguish ground truth from invention
- Preserve voice from first sentence

---

## Example Session Flow

**User:** "I want to write about why prose-polish matters"

**You (Interview):** "Let's start with why this topic. Of all the skills you've built, why does prose-polish pull your attention?"

**User:** [Responds with reasons]

**You (Thread):** "Marking [MAIN: Collaborative emergence—AI as intellectual prosthesis]. Marking [TANGENT: The real AI tell you've observed]. Marking [RESEARCH: Need examples of bad vs. good writing]."

**You (Interview continues):** "You said 'intellectual prosthesis.' What's the distinction for you between prosthesis and collaborator?"

**User:** [Responds]

**You (Transition check):** "The angle is clear: prose-polish exists for collaborative emergence, not just detection. We have ground truth about your experience building it. Ready to draft, or more to explore?"

**User:** "Let's draft"

**You (Draft):** "[If samples available] Loading WritingSamples/[relevant file] to match your voice... Here's an opening that sounds like you thinking out loud: [draft]"

**User:** [Reviews]

**You (Calibrate):** "Running through prose-polish... Flagging one phrase in paragraph 3 that sounds like the 'you might think' pattern. Here's an alternative that preserves the point in your voice..."

---

## Escape Hatches

When things break, recover gracefully:

- **Thread explosion (10+):** "We have too many threads. Let's resolve a few before continuing."
- **Interview stalling:** "I'm not getting the depth I need. Want to try a different entry point?"
- **Voice false positives:** User can say "trust my voice on this" to skip calibration
- **Mode confusion:** User can say "show status" to see current mode and threads

---

## State Management

**Within session:** Conversation history maintains state. Thread markers appear inline. No external files for MVP.

**Across sessions (Phase 2):** Beads integration. At session end: "Create bead issue to persist these threads?" Next session: Load from bead.

**For MVP:** Single-session scope. Most essays emerge in one focused session anyway.

---

## Success Criteria

The true test: Does the output sound like the user, not like AI?

If yes → You're doing it right
If no → Return to interview material, check voice against samples (if available) or blocklist, ask don't invent

---

**Version:** 1.0.0 (MVP based on interview 2025-12-18)


---

## Referenced Files

> The following files are referenced in this skill and included for context.

### references/blocklist.md

```markdown
# Voice Blocklist — AI Tells to Avoid

Patterns that signal AI-generated text. Auto-flag with high confidence when detected.

---

## Tier 1: Hard Blocklist (95%+ confidence)

These patterns should trigger immediate flagging. Lee has explicitly identified these.

### The "Actually" Pattern

```
"You might think X... but actually Y"
```

All derivatives:
- "You might assume..."
- "One might expect..."
- "It may seem like... but in reality..."
- "At first glance... however..."
- "Conventional wisdom says... but..."

**Why it's a tell:** Creates false contrast, sounds like performative insight. Lee's observation, not invented.

### Corporate Opening Gambits

```
"In today's world..."
"In the modern era..."
"In an age of..."
"As we navigate..."
```

**Why it's a tell:** Throat-clearing that delays the actual point. Lee starts with the point.

### Importance Declarations

```
"It's important to note that..."
"It's worth noting..."
"It bears mentioning..."
"Significantly..."
"Crucially..."
```

**Why it's a tell:** Tells the reader something is important instead of showing why.

---

## Tier 2: High Confidence Flags (80%+)

These patterns are strong signals but may have occasional legitimate uses.

### Excessive Hedging

```
"perhaps"
"might"
"could potentially"
"may or may not"
"it's possible that"
"one could argue"
```

**When flagged:** Multiple hedges in same paragraph, or hedging core claims.

**When acceptable:** Genuine uncertainty, speculative sections.

### Generic Wisdom

```
"At the end of the day..."
"When all is said and done..."
"The bottom line is..."
"What really matters is..."
```

**Why it's a tell:** Generic wrap-up phrases that could apply to anything.

### Listicle Formatting

```
🎯 **Key Insight:**
✨ **Pro Tip:**
💡 **Remember:**
```

**Why it's a tell:** Performative structure over substance. Lee doesn't write like a listicle.

---

## Tier 3: Contextual Flags (60%+)

These patterns are suspicious but need context to evaluate.

### Perfect Parallelism

```
"First... Second... Third..."
"On one hand... On the other hand..."
"Not only X, but also Y"
```

**When flagged:** Mechanical structure without organic flow.

**When acceptable:** When parallelism serves the argument.

### Superlative Stacking

```
"absolutely essential"
"truly transformative"
"incredibly powerful"
"fundamentally important"
```

**Why suspicious:** Intensity without evidence. Lee backs claims with specifics.

### Passive Constructions

```
"It has been observed that..."
"It is believed that..."
"It could be argued that..."
```

**Why suspicious:** Hides agency. Lee owns his claims.

---

## Tier 4: Soullessness Indicators

Not specific phrases but patterns that signal AI-generated prose.

### Grammatical Perfection + Zero Friction

Every sentence technically correct but nothing snags attention. No parenthetical asides, no rhythm variation, no thinking-out-loud quality.

**Test:** Read aloud. Does it sound like someone talking or like a textbook?

### Generic Examples

```
"For example, imagine a scenario where..."
"Consider the case of a hypothetical..."
```

**Lee's style:** Concrete, specific examples from real experience.

### Missing "I"

Long passages without first-person ownership. Lee's voice is present—he uses "I" when he means it.

---

## Detection Protocol

When calibrating, run through this checklist:

1. **Tier 1 scan:** Any hard blocklist matches? → Flag immediately
2. **Tier 2 scan:** Multiple hedges? Generic wisdom? → Flag for review
3. **Tier 3 scan:** Mechanical structure? Superlatives without evidence? → Note for context
4. **Tier 4 feel:** Does it sound like Lee or like AI? → Compare to WritingSamples/

---

## Remediation Approach

When flagged:

1. **Quote the specific phrase** that triggered the flag
2. **Explain why** it's on the blocklist
3. **Show alternative** from WritingSamples/ if available
4. **Suggest rewrite** in Lee's voice

**Example:**

```
Flagged: "It's important to note that collaborative emergence matters."

Why: "It's important to note" is on the blocklist—it tells importance instead of showing.

Alternative from WritingSamples/: "Collaborative emergence matters because..."

Suggested rewrite: "Collaborative emergence is the whole point. Without it, you just have AI writing about you instead of with you."
```

---

## Updating the Blocklist

When Lee corrects a new pattern:
1. Add to appropriate tier
2. Note why it's a tell
3. Include example of what Lee prefers instead

The blocklist should grow from actual corrections, not speculation.

---

**Last updated:** 2025-12-18
**Source:** Interview session + Lee's explicit observations

```

### integrations/prose-polish.md

```markdown
# Prose-Polish Integration

How to use the prose-polish skill for AI detection during voice calibration.

---

## When to Invoke

### Always

- After drafting any substantial section (300+ words)
- Before declaring a section "done"
- When user asks "does this sound like me?"
- When calibration mode flags potential issues

### Never

- During interview (no draft yet)
- On user's own words (don't second-guess their input)
- On minor edits (overkill)

---

## How to Invoke

```
Invoke prose-polish skill with the draft text.
```

Or reference in conversation:
```
"Let me run this through prose-polish for AI detection..."
```

---

## Interpreting Results

### Score Thresholds

| Score | Interpretation | Action |
|-------|---------------|--------|
| < 15 | Likely human | Minimal review, proceed |
| 15-25 | Borderline | Check flagged phrases against blocklist |
| 25-35 | Suspicious | Review with WritingSamples/ comparison |
| > 35 | Likely AI | Significant revision needed |

**Lee's threshold:** >25 warrants attention.

### What to Look For

Prose-polish provides:
- Overall AI-likelihood score
- Per-paragraph analysis
- Specific flagged phrases
- Craft assessment (rhythm, voice, commitment)

Cross-reference flagged phrases with `voice/blocklist.md`:
- If phrase is on blocklist → confirmed issue
- If phrase is not on blocklist but flagged → investigate pattern

---

## Two-Pass Workflow

### Pass 1: Draft (Writing Partner)

- Focus on getting ideas out
- Track threads
- Maintain voice heuristically (blocklist awareness)
- Don't interrupt for prose-polish

### Pass 2: Polish

1. Complete section or full draft
2. Run prose-polish
3. Review flags
4. Return to drafting mode for targeted revision
5. Re-run prose-polish to verify

```
Draft → Prose-Polish → Flags → Revise → Verify
```

---

## Handling Conflicts

### Prose-polish says "good" but voice feels off

Trust voice calibration. Prose-polish detects generic AI patterns; voice calibration is Lee-specific.

**Action:** Compare to WritingSamples/ manually. Look for:
- Rhythm mismatch
- Missing "I" ownership
- Generic instead of specific examples

### Prose-polish flags something Lee actually wrote

This can happen. Lee's style might occasionally match AI patterns.

**Action:**
1. Check if it's a quote or direct input from Lee
2. If so, mark as "trust my voice on this"
3. If not, investigate why it feels off

### Multiple low scores but draft feels hollow

Prose-polish measures detectable patterns. "Soullessness" is harder to quantify.

**Action:**
1. Read draft aloud
2. Compare to WritingSamples/ rhythm
3. Check for missing concrete examples
4. Look for "animating spirit"

---

## Reporting Format

When sharing prose-polish results:

```
Prose-polish analysis:

Overall score: 28 (above threshold)

Flagged passages:
- Paragraph 2: "It's important to note that..." (blocklist match)
- Paragraph 4: Low perplexity detected (too predictable)

Recommendation: Revise paragraphs 2 and 4 before finalizing.
```

---

## Integration with Voice Calibration

Prose-polish is Layer 2 in the detection pipeline:

1. **Layer 1:** Blocklist pattern matching (fast)
2. **Layer 2:** Prose-polish AI detection (comprehensive)
3. **Layer 3:** WritingSamples/ comparison (Lee-specific)

Run layers in order. If Layer 1 catches it, Layer 2 confirms. If Layer 2 flags something Layer 1 missed, investigate.

---

**See also:**
- `voice/blocklist.md` for pattern detection
- `integrations/writing-samples.md` for corpus comparison

```

### integrations/writing-samples.md

```markdown
# WritingSamples Integration

How to use personal writing samples for voice calibration.

---

## Overview

WritingSamples/ provides the gold standard for voice calibration—comparing AI-assisted drafts against your actual writing. This integration is **optional but recommended** for best results.

**Without samples:** The skill relies on blocklist pattern detection and prose-polish scoring.

**With samples:** Full voice calibration comparing rhythm, openings, transitions, and example usage.

---

## Setup

### Directory Structure

Place a `WritingSamples/` directory in your working directory:

```
your-project/
├── WritingSamples/
│   ├── 01-topic-description.md
│   ├── 02-another-piece.md
│   └── ...
└── ... (your other files)
```

### Sample Selection

**Good samples:**
- Authentic writing you're proud of
- Variety of topics and formats
- 500-2000 words each
- Represents your actual voice (not edited by AI)

**Avoid:**
- Heavily edited or co-written pieces
- Very short snippets (<300 words)
- Writing from years ago (voice evolves)
- Formal/academic writing if you want casual voice (or vice versa)

### Naming Convention

Use numbered prefixes for easy reference:
```
01-diversity-inclusion-essay.md
02-technical-blog-post.md
03-personal-narrative.md
```

---

## How It Works

### During Calibration

1. **Topic matching**: Identify the domain of current draft
2. **Sample search**: Find 1-2 samples with similar topics or tone
3. **Pattern extraction**: Pull 2-3 representative paragraphs
4. **Comparison**: Check draft against sample patterns
   - Sentence rhythm (short/long variation)
   - Opening styles (how paragraphs begin)
   - Transition patterns
   - Example usage (concrete vs abstract)
   - First-person ownership

### Comparison Protocol

```
Load sample → Extract patterns → Compare to draft → Flag mismatches → Suggest alternatives
```

**Example comparison:**

Sample pattern:
> "I wonder about the gap between intention and impact. When I led the migration project, we had the best intentions..."

Draft pattern:
> "It's important to note that the gap between intention and impact matters significantly in organizational contexts."

**Mismatch detected:**
- Sample uses "I wonder about..." opening
- Draft uses "It's important to note..." (blocklist item)
- Sample has concrete example ("When I led...")
- Draft is abstract

**Suggested revision:**
> "I keep coming back to the gap between intention and impact. During the migration project, I saw this firsthand..."

---

## When to Load Samples

### Always Load
- Before drafting major sections
- During calibration checks
- When user asks "does this sound like me?"

### Never Load
- During interview (no draft yet)
- For minor edits
- When user says "trust my voice on this"

---

## Building Your Sample Collection

### Starting from Scratch

If you don't have samples yet:

1. **Collect existing writing**: Blog posts, emails, memos, journal entries
2. **Curate for authenticity**: Choose pieces that feel most "you"
3. **Aim for 5-10 samples**: Enough variety without overwhelming
4. **Update periodically**: Voice evolves; refresh annually

### Voice Capture Mode (Future)

In a future version, Writing Partner will include a "voice capture" interview mode to help you articulate your voice patterns explicitly.

---

## Troubleshooting

### "No relevant samples found"

- Add samples covering more topics
- Use more general samples that show voice regardless of topic

### "Sample doesn't match my current style"

- Voice evolves; consider refreshing old samples
- Add recent writing to the collection

### "Flagging things I actually wrote"

- Your style may occasionally match AI patterns
- Use "trust my voice on this" to override
- Consider if the pattern is something you want to keep

---

## Privacy Note

WritingSamples/ should contain your personal writing. If using a public repository:

- Add `WritingSamples/` to `.gitignore`
- Keep samples in a private location
- The skill works without samples (blocklist-only mode)

---

**See also:**
- `references/blocklist.md` for AI pattern detection
- `integrations/prose-polish.md` for AI detection scoring

```

### integrations/persona-schema.md

```markdown
# Persona Schema Integration

How to use a SiliconDoppelganger persona schema for enhanced voice calibration.

---

## Overview

A **persona schema** is a structured XML file created using the SiliconDoppelganger skill. It encodes:
- Psychometric data (CliftonStrengths, VIA Character Strengths)
- Linguistic fingerprint (tone, vocabulary, sentence patterns)
- Core drivers (motivations, fears)
- Decision logic (how the person thinks and reasons)
- Conflict style and communication preferences

When available, this data significantly enhances WritingPartner's ability to match your authentic voice.

---

## What a Persona Schema Contains

```xml
<persona_profile>
    <psychometrics>
        <clifton>Top 5-10 CliftonStrengths</clifton>
        <via>Top 5-10 VIA Character Strengths</via>
    </psychometrics>
    <linguistic_fingerprint>
        Sentence structure, vocabulary patterns, tone, formality level
    </linguistic_fingerprint>
    <core_drivers>
        <primary_motivation>Impact | Security | Novelty | Money</primary_motivation>
        <primary_fear>Irrelevance | Boredom | Conflict | Poverty</primary_fear>
    </core_drivers>
    <decision_logic>
        Risk tolerance, data preferences, blind spots
    </decision_logic>
    <conflict_style>
        Communication patterns under normal and stress conditions
    </conflict_style>
    <narrative_anchors>
        Origin story, shadow self, unpopular opinions
    </narrative_anchors>
</persona_profile>
```

---

## How WritingPartner Uses Persona Data

### During Interview Mode

**Core drivers** inform question selection:
- If motivation = Impact → probe for "who will this help?"
- If motivation = Novelty → probe for "what's new here?"
- If fear = Irrelevance → surface "why does this matter now?"

**Narrative anchors** provide context:
- Origin story themes may connect to essay topics
- Unpopular opinions suggest contrarian angles worth exploring

### During Drafting Mode

**Linguistic fingerprint** guides prose generation:
- Match sentence structure patterns (varied? punchy? flowing?)
- Use vocabulary consistent with domain expertise
- Apply appropriate formality level
- Include characteristic phrase patterns

**Decision logic** shapes argument structure:
- Risk tolerance affects how boldly claims are stated
- Data preferences influence what evidence to emphasize
- Blind spots suggest what counterarguments to address

### During Calibration Mode

**Conflict style** informs tone checking:
- Debater style → more direct assertions acceptable
- Diplomat style → softer framing preferred
- Check stress patterns don't leak into normal prose

**Psychometrics** provide authenticity markers:
- Strategic thinker → expect pattern recognition, future orientation
- Learner → expect curiosity, "I wonder" openings
- Honesty strength → expect directness, truth-seeking tone

---

## Setup

### If You Have a Persona Schema

Place your persona schema where WritingPartner can find it:

**Option 1: In working directory**
```
your-project/
├── WritingSamples/           # Voice samples
├── persona-schema.xml        # Your Digital Twin
└── ...
```

**Option 2: Reference external location**
```
Tell WritingPartner: "Load persona from ~/Projects/SiliconDoppelgangerActual/silicon-doppelganger-actual/data/lee-gonzales-persona-schema.xml"
```

### If You Don't Have a Persona Schema

You can create one using the **SiliconDoppelganger** skill:
1. Invoke SiliconDoppelganger
2. Complete the extraction interview (45-60 min)
3. Generate the persona schema XML
4. Use with WritingPartner for enhanced voice matching

Or skip it—WritingPartner works well with just WritingSamples and blocklist.

---

## Integration Layers

WritingPartner uses multiple layers for voice calibration, each adding fidelity:

| Layer | Source | What It Provides |
|-------|--------|------------------|
| 1. Blocklist | `references/blocklist.md` | AI patterns to avoid |
| 2. prose-polish | External skill | AI detection scoring |
| 3. WritingSamples | `WritingSamples/` directory | Actual prose examples |
| 4. Persona Schema | `persona-schema.xml` | Psychometric + cognitive patterns |

**With all four layers:** Maximum voice fidelity
**With layers 1-3:** Excellent voice matching
**With layers 1-2 only:** Good AI detection, generic voice
**With layer 1 only:** Basic blocklist protection

---

## Persona Schema + WritingSamples

These complement each other:

| Aspect | Persona Schema | WritingSamples |
|--------|----------------|----------------|
| **Provides** | The "why" | The "how" |
| **Captures** | Motivations, thinking patterns | Actual prose rhythm |
| **Format** | Structured XML | Raw markdown |
| **Created by** | Interview extraction | Collecting existing writing |
| **Updates** | Quarterly validation | As you write more |

**Best practice:** Use both when available. Persona explains the mind; samples show the voice.

---

## Example Usage

**User:** "Let's work on an essay about AI in education"

**WritingPartner (with persona loaded):**

*Interview informed by persona:*
> "Your CliftonStrengths show Strategic and Learner in top 5. Are you approaching this from a pattern-recognition angle (what you see happening) or a learning angle (what you're discovering)?"

*Drafting informed by persona:*
> Loading linguistic fingerprint: varied sentences, technical vocabulary, direct tone, occasional philosophical references...

*Calibration informed by persona:*
> Checking against Honesty strength—this passage hedges more than your typical directness. Consider: "AI will transform education" instead of "AI might potentially transform education."

---

## Creating a Persona Schema

If you want to build one, use the **SiliconDoppelganger** skill:

```
Invoke: "Create a persona schema for me"
```

The skill will guide you through:
1. **Hardware extraction** — Psychometric assessments
2. **Operating System mapping** — Decision heuristics
3. **Narrative Identity capture** — Origin story, shadow self
4. **Schema encoding** — Compile into XML format
5. **Validation** — Test against real behavior

See the SiliconDoppelganger skill for full methodology.

---

## Privacy Note

Persona schemas contain sensitive personal data:
- Keep in private repositories
- Don't commit to public repos
- The schema is your "psychological fingerprint"

WritingPartner references the schema but never exposes its contents.

---

**See also:**
- SiliconDoppelganger skill for persona creation
- `integrations/writing-samples.md` for prose examples
- `references/blocklist.md` for AI pattern detection

```