Brand Voice

1. Progress bar — based on how many core voice dimensions have been covered:

   • Sentence structure (length, fragments, rhythm)
   • Vocabulary and word choice
   • Slang and colloquialisms
   • Tone and energy level
   • Punctuation habits
   • Perspective and point of view (I/we/third person)
   • Tense preferences
   • Humor style
   • How they handle descriptions (sparse vs vivid)
   • Emotional register (hype, understated, raw, polished)

   Show: [=========> ] 70% — 7/10 dimensions covered

2. Gaps — list which dimensions need more data

3. Word count — current size vs target range

4. Strongest patterns — the 3 most confident observations in the profile

Voice Session

1. Read the current profile

Read .claude/brand-voice.md. If it doesn't exist, this is the first session — create it with a header and empty structure.

2. Design prompts based on profile state

If the profile is empty or thin (< 500 words): Use broad, open-ended writing prompts that cast a wide net. These should feel natural and fun, not like a quiz. Examples:

• "Describe the best night out you've had recently"
• "Tell me about a track or album that changed how you listen to music"
• "Write a caption for a photo of a packed dance floor"
• "Rant about something that annoys you about the music scene"
• "How would you introduce a friend to someone at a party?"

If the profile is developing (500-1,500 words): Get more targeted. Look at which dimensions are weak and craft prompts that specifically draw those out:

• If missing humor data: "Describe someone doing something ridiculous at an event"
• If missing descriptive style: "Paint a picture of a venue right before the doors open"
• If missing opinion voice: "What's overrated right now?"
• If missing instructional voice: "Explain to someone how to find good music"

If the profile is strong (1,500+ words): Test and refine. Use prompts that explore edge cases and push boundaries:

• Present two versions of the same sentence and ask which sounds more like them
• Ask them to write in a context they haven't covered (formal announcement, sad moment, technical explanation)
• Challenge established patterns: "Write this same idea but in a completely different way than you normally would"
• Use contrasts: "How would you describe this to a close friend vs to a stranger?"

3. Present prompts

Present 3-5 prompts per session. Show them one at a time. After each response:

1. Acknowledge briefly — don't over-praise, don't critique. Just a quick nod so they know you received it.
2. Move to the next prompt — keep momentum. Don't analyze out loud between prompts.

After all prompts are answered, proceed to analysis.

4. Analyze and update the profile

After the session, analyze all responses together. Look for:

• Sentence patterns — average length, use of fragments, run-ons, rhythm
• Word choice — preferred vocabulary, words they reach for, words they avoid
• Slang and colloquialisms — specific terms, frequency, context
• Tone — energy level, confidence, warmth, edge
• Punctuation — em dashes, ellipses, periods vs no periods, exclamation marks (or lack of)
• Perspective — first person, plural, how they refer to themselves
• Tense — present vs past, when they switch
• Descriptions — sparse/punchy or vivid/detailed
• Humor — dry, playful, sarcastic, absurd
• Emotional register — how they express excitement, frustration, nostalgia

Update the voice profile file:

• Add new patterns discovered in this session
• Reinforce existing patterns that were confirmed
• Note any contradictions or context-dependent shifts
• Include 2-3 direct quotes from the session as examples (these are the most valuable — real voice samples)

Important: Write the profile as a style guide, not a conversation log. Distill patterns, don't dump transcripts.

5. Trim if needed

If the profile exceeds 3,000 words after updating:

1. Merge redundant observations
2. Remove the weakest/least confident patterns
3. Keep the strongest example quotes, drop weaker ones
4. Consolidate similar dimensions
5. Never drop a dimension entirely — just tighten the description

6. Show session summary

After updating, show:

• Progress bar (updated)
• What was learned this session (2-3 bullet points)
• Gaps remaining
• How many sessions until the profile is strong (estimate)

Voice Profile Structure

The .claude/brand-voice.md file should follow this structure:

# Brand Voice Profile — Audeos

> Last updated: YYYY-MM-DD
> Sessions completed: N
> Coverage: N/10 dimensions

## Overview
[2-3 sentence summary of the overall voice]

## Sentence Structure
[Patterns about length, fragments, rhythm]

## Vocabulary & Word Choice
[Preferred words, avoided words, register]

## Slang & Colloquialisms
[Specific terms with usage context]

## Tone & Energy
[Overall energy, confidence level, warmth]

## Punctuation & Formatting
[Habits with dashes, periods, caps, etc.]

## Perspective & Point of View
[I/we/they, how they refer to themselves and others]

## Tense & Temporality
[Present vs past, when they shift]

## Descriptive Style
[Sparse vs vivid, what they focus on]

## Humor
[Type, frequency, context]

## Emotional Register
[How they express different emotions]

## Example Quotes
[Direct quotes from sessions that best capture the voice]

Important Rules

• Never ask personal or biographical questions — this repo is open source. Focus only on writing style, opinions, and creative expression. No questions about family, relationships, identity, location history, etc.
• Never analyze out loud between prompts — keep the session flowing. Save all analysis for the end.
• Use the profile to evolve prompts — every session should feel different. Never repeat the same prompt twice. Read the profile before designing prompts.
• Distill, don't dump — the profile is a style guide, not a transcript. Extract patterns, include select quotes, discard the rest.
• Respect the size budget — 2,000-3,000 words. Quality over quantity. A tight, accurate profile beats a bloated one.
• Direct quotes are gold — always include a few real quotes from each session. They are the most reliable voice reference.
• No privacy-sensitive content — if the user's response contains personal details, extract only the stylistic patterns, not the personal facts.