voice-architect

Voice Architect

Build an author's voice profile through structured interview. The profile captures how they argue, their sentence rhythm, vocabulary, emotional patterns, and what they'd never say — everything needed to generate prose that sounds like THEM, not like AI.

Core Principle

Voice is not style. Style is mechanics (sentence length, punctuation, word choice). Voice is the person — their origin story, their patterns of thought, how they fight, what they refuse to say. This interview extracts both.

The Interview: 8 Questions

Ask ONE question at a time. Wait for the full answer before proceeding. These questions may take hours or days to answer — save state after each.

Question 1: Origin Story

"How did you come to this work? Not your resume — the moment you knew this was your thing."

What to extract: The emotional entry point. Where conviction comes from. Whether they lead with stories or credentials.

Question 2: The Pattern

"What do you see that others miss? What pattern have you noticed over years that most people in your field haven't caught yet?"

What to extract: Intellectual signature. How they synthesize. Whether they think in systems, narratives, or data.

Question 3: How You Argue

"When you're making a case to someone who disagrees, how do you do it? Stories first? Data first? Do you get loud or get quiet?"

What to extract: Argument structure. Rhetorical instincts. Whether they're deductive, narrative, Socratic, or confrontational.

Question 4: What You Hate

"What's the thing in your field that makes you genuinely angry? The practice, the phrase, the mindset that you'd burn if you could?"

What to extract: Emotional triggers. Where conviction turns to heat. The forbidden territory they patrol.

Question 5: How You Explain

"Pick something complex in your domain. Explain it to me like I'm smart but from a different field."

What to extract: Teaching style. Metaphor preferences. How they simplify without dumbing down. Technical vocabulary comfort.

Question 6: Your Reader

"When you picture someone reading your work and thinking 'they get it' — who is that person and what specifically did you say that landed?"

What to extract: Who they write FOR. What "landing" feels like to them. Their theory of impact.

Question 7: The Line You Won't Cross

"What's the popular advice in your space that you refuse to give? Why?"

What to extract: Intellectual boundaries. Where they diverge from consensus. The contrarian positions they hold with conviction.

Question 8: Your Natural Register

"Do you write like you talk? More formal? Less? What's the closest published thing to how you actually sound?"

What to extract: Self-awareness about register. The gap (or lack of gap) between spoken and written voice. Reference points for calibration.

After the Interview

Analysis

After all 8 answers are collected, analyze for:

1. Argument style: Deductive (premise → conclusion), Narrative (story → insight), Socratic (question → discovery), Confrontational (challenge → proof), or hybrid
2. Sentence rhythm: Short and punchy? Long and layered? Alternating? Fragments for emphasis?
3. Vocabulary register: Technical depth, jargon comfort, metaphor density
4. Emotional patterns: Where they get heated, where they pull back, what triggers conviction vs. caution
5. Teaching mode: Analogies? Step-by-step? Examples first? Theory first?
6. Forbidden territory: Phrases, structures, tones that would never come from this person

Voice Profile Generation

Generate the voice profile document:

# Voice Profile: [Author Name]
## Created: [date]
## Last validated: [date]
## Project: [project name]

### How They Argue
[Analysis of argument style with examples from interview answers]

### Sentence Rhythm
[Observed patterns: length, variation, fragment usage, paragraph style]

### Vocabulary & Register
[Technical depth, jargon comfort, metaphor preferences, words they use vs. avoid]

### Emotional Patterns
[Where they get heated, where they pull back, what triggers conviction]

### Teaching Mode
[How they explain complex things, metaphor style, simplification approach]

### Forbidden Patterns
[Phrases, structures, or tones that would never come from this author]
[Include specific examples: "This author would never say 'It's worth noting that...'"]

### "Sounds Like Me" Samples
[3-5 short passages the author confirmed as authentic to their voice]
[Drawn from interview answers — the parts where they were most themselves]

### Anti-Detection Markers
[Specific checklist items calibrated to THIS author's natural style]
- [ ] Contraction rate: [X]% (based on observed patterns)
- [ ] Sentence length range: [X]-[Y] words typical
- [ ] Fragment usage: [frequency and purpose]
- [ ] Paragraph length: [typical range]
- [ ] Technical vocabulary depth: [level]
- [ ] Emotional vocabulary: [patterns]
- [ ] Conjunction starters: [frequency]
- [ ] Metaphor density: [low/medium/high]

Validation

Present the voice profile to the author: "Does this sound like you? What's wrong? What's missing?"

Iterate until the author confirms. Log each iteration in the authorship record.

Importing Voice Profiles

When a voice profile file is found in a project's voice-profiles/ directory that wasn't created during this project's interview:

1. Detect the import (file exists but no corresponding voice_interview_answer decisions in the authorship record)
2. Ask: "I found an existing voice profile for [name]. Use it as-is, or re-interview to update?"
3. If as-is: Log a voice_profile_imported decision in the authorship record
4. If re-interview: Run the full 8-question interview, noting where answers diverge from the imported profile
5. Either way, validate the profile against the current project's content type

State Persistence

After each interview answer:

1. Write the answer to voice-profiles/[author-name]-raw-qa.md
2. Log a voice_interview_answer decision in the authorship record
3. Update the manifest: pending_question with the next question and context
4. Update progress.voice_profile to in_progress

After profile generation and validation:

1. Write the profile to voice-profiles/[author-name].md
2. Log voice_profile_generated and voice_profile_validated
3. Update progress.voice_profile to complete
4. Clear pending_question

What Voice Architect Does NOT Do

• Does not apply any preset voice (MoxyWolf, brand guidelines, etc.) unless the author explicitly requests it
• Does not generate prose — it only builds the profile. Prose generation happens in Editorial Forge Phase 4.
• Does not skip questions or batch them. Each question builds on previous answers.
• Does not assume voice is static. If an author's voice evolves, re-interview and update the profile.