brand-voice-architect（品牌声音架构师）

Brand Voice Architect (BVA)

A skill for engineering, documenting, and synthesizing brand-specific voice with quantifiable precision. Brand voice is treated as a Linguistic DNA — a measurable baseline, not an aesthetic preference.

Core Workflow

Phase I: Decomposition — /analyze [corpus]

Run a linguistic audit on provided text samples:

1. Lexical Audit — High-frequency verbs/adjectives, prohibited terms, vocabulary signature
2. Structural Mapping — Average Sentence Length (ASL), syntactic complexity, variance
3. Sentiment Baseline — Emotional temperature on a 0.0–1.0 scale

→ Use scripts/voice_analyzer.py to compute metrics programmatically when a corpus is provided.

Phase II: Architectural Design — /synthesize [pillars]

Build the voice matrix:

1. Pillar Definition — Establish 3 core attributes (e.g., *Authoritative, Wit-driven, Technical*)
2. The Spectrum — Define "This, Not That" logic gates for each pillar
3. Persona Encoding — Translate pillars into LLM system-level instructions

→ Use scripts/prompt_synthesizer.py to generate deployable system prompts.

Phase III: Delivery

1. Artifact Generation — Produce voice guide docs, style reference cards, prompt templates
2. Manual Review — /review [output] provides a qualitative checklist to assess whether output aligns with the established voice pillars (Claude-assisted, not script-automated)
3. Platform Pivot — /pivot [context] adapts voice for specific channels while preserving DNA, using generate_platform_pivot() from prompt_synthesizer.py

Note on prohibited words: The generated system prompt instructs the LLM to replace prohibited words with preferred equivalents. This is a prompt-level instruction — enforcement depends on the model following the system prompt, not on automated script-level filtering.

The 4-Pillar Framework

Map every brand voice across four axes to define its Safe Operating Area:

See references/methodology.md for full framework details including Cadence Analysis and Semantic Salience scoring.

Mandatory Output Components

Every Brand Voice engagement must produce:

1. Metrics Report — Lexical density %, ASL, top keywords, cadence variance
2. Voice Matrix — 3 pillars × "This/Not That" for each
3. System Prompt — Ready-to-deploy LLM persona encoding
4. Platform Pivots — At minimum: formal/informal, long-form/short-form variants
5. Prohibited/Preferred Lexicon — Concrete word lists

Quick Reference Commands

Scripts

• scripts/voice_analyzer.py — Computes lexical density, ASL, cadence variance, sentiment temperature, and top keywords from a corpus
• scripts/prompt_synthesizer.py — Generates deployable LLM system prompts from a BrandConfig object; includes generate_platform_pivot() for channel-specific adaptations

References

• references/methodology.md — Full technical methodology: 4-Pillar Framework, Cadence Analysis, Semantic Salience, Human-AI Collaborative Loop