voice-create

voice-create

Generate custom voice profiles from natural language descriptions. Every generated profile includes AI tells avoidance rules by default.

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

• "make me sound like [reference]" -> reference-based voice creation
• "voice fingerprint" -> voice profile extraction from text

Behavior

When triggered, this skill:

1. Parses the description to identify:

   • Target audience (developers, executives, general public)
   • Tone characteristics (formal/casual, confident/tentative, warm/clinical)
   • Domain context (technical, marketing, academic, conversational)
   • Any specific constraints or preferences mentioned

2. Maps description to voice dimensions:

   • Formality (0-1): casual <-> formal
   • Confidence (0-1): hedging <-> assertive
   • Warmth (0-1): clinical <-> friendly
   • Energy (0-1): calm <-> enthusiastic
   • Complexity (0-1): simple <-> sophisticated

3. Generates vocabulary guidance:

   • Preferred terms based on domain
   • Terms to avoid based on tone
   • Signature phrases that match the voice
   • Always includes AI tells avoidance: every generated profile's vocabulary.avoid list includes the full set of banned AI vocabulary from ../voice-apply/references/ai-tells.md

4. Creates structure patterns:

   • Sentence length preferences
   • Paragraph structure
   • Use of lists, examples, analogies
   • Always enforces human-like rhythm: high sentence length variance, varied paragraph lengths, em dash avoidance, no Rule of Three abuse

5. Adds anti-AI-tells rules:

   • punctuation.em_dashes: avoid (default for all generated profiles)
   • punctuation.semicolons: avoid (unless user specifies formal/academic)
   • structure.sentence_length_variance: high (enforces burstiness)
   • structure.paragraph_length_variance: high (prevents uniform blocks)
   • structure.rule_of_three: break-pattern (vary item counts)
   • banned_patterns: negation flips, self-posed questions, anaphora, trailing participles, hedge-stacking

6. Outputs valid YAML conforming to voice-profile schema

Usage Examples

Technical Documentation Voice

User: "Create a voice for API documentation - precise, no-nonsense, assumes developer knowledge"

Output: technical-api-docs.yaml
- formality: 0.6
- confidence: 0.9
- warmth: 0.2
- energy: 0.3
- complexity: 0.8
- vocabulary.prefer: technical terms, code references, precise metrics
- vocabulary.avoid: [all AI tells] + marketing language
- structure: varied sentence length, no em dashes

Friendly Tutorial Voice

User: "Make me a voice for beginner tutorials - encouraging, patient, uses lots of analogies"

Output: beginner-tutorial.yaml
- formality: 0.2
- confidence: 0.7
- warmth: 0.9
- energy: 0.7
- complexity: 0.3
- vocabulary.prefer: everyday language, encouraging phrases, analogies
- vocabulary.avoid: [all AI tells] + jargon
- structure: short sentences mixed with longer explanations, no "let's dive in" openers

Executive Summary Voice

User: "Generate a voice profile for board presentations - authoritative but accessible"

Output: board-presentation.yaml
- formality: 0.8
- confidence: 0.9
- warmth: 0.4
- energy: 0.5
- complexity: 0.6
- vocabulary.prefer: business metrics, strategic language, clear conclusions
- vocabulary.avoid: [all AI tells] + buzzwords like "synergy", "paradigm", "stakeholders"
- structure: varied paragraph lengths, direct openers, no "in conclusion" closers

Output Location

Generated profiles are saved to:

1. .aiwg/voices/{name}.yaml (project-specific, default)
2. ~/.config/aiwg/voices/{name}.yaml (user-wide, with --global flag)

Voice Generation Process

Step 1: Dimension Calibration

Parse natural language for dimension indicators:

Description Keywords	Dimension	Value Range
casual, relaxed, conversational	formality	0.1-0.3
professional, business	formality	0.5-0.7
formal, academic, official	formality	0.8-1.0
tentative, careful, hedging	confidence	0.2-0.4
balanced, measured	confidence	0.5-0.7
assertive, authoritative, direct	confidence	0.8-1.0
clinical, detached, objective	warmth	0.1-0.3
neutral, professional	warmth	0.4-0.6
friendly, warm, personable	warmth	0.7-0.9
calm, measured, understated	energy	0.1-0.3
balanced, engaged	energy	0.4-0.6
enthusiastic, dynamic, energetic	energy	0.7-0.9
simple, accessible, plain	complexity	0.1-0.3
clear, moderate	complexity	0.4-0.6
sophisticated, detailed, nuanced	complexity	0.7-0.9

Step 2: Domain Detection

Identify domain from context:

• Technical: API, code, system, architecture, implementation
• Marketing: brand, campaign, audience, engagement, conversion
• Academic: research, methodology, analysis, findings, literature
• Executive: strategy, ROI, decision, outcome
• Support: help, issue, solution, troubleshoot, resolve

Step 3: Vocabulary Generation

Based on domain and tone, generate:

• 5-10 preferred terms
• 3-5 domain-specific terms to avoid
• 2-4 signature phrases
• Mandatory: merge the full AI tells banned vocabulary list into vocabulary.avoid. This list includes 50+ banned verbs, 40+ banned adjectives, 40+ banned nouns, 30+ banned adverbs, and all banned connectors. See ../voice-apply/references/ai-tells.md for the complete lists.

Step 4: Structure Selection

Map tone to structure patterns:

• High formality -> longer sentences, structured paragraphs
• Low formality -> shorter sentences, varied structure
• High confidence -> direct statements, conclusions first
• High warmth -> questions, inclusive language ("we", "let's")

Always enforce:

• sentence_length_variance: high (mix 5-word punches with 30-word flows)
• paragraph_length_variance: high (mix 1-sentence paragraphs with 5-sentence paragraphs)
• em_dash_usage: avoid (use commas, periods, parentheses, colons instead)
• rule_of_three: break-pattern (use 2 items or 4 items to break the AI fingerprint)

Step 5: Anti-Pattern Rules

Every generated profile includes these structure rules:

anti_ai_tells:
  # Sentence patterns to never use
  banned_patterns:
    - negation_flip          # "It's not X, it's Y"
    - self_posed_questions   # "The best part? It's this."
    - anaphora_abuse         # Same opener 3+ times
    - trailing_participles   # "highlighting its importance"
    - false_ranges           # "From innovation to implementation"
    - hedge_stacking         # Multiple qualifiers per sentence
    - consecutive_adverbials # Starting sentences with "Furthermore," "Moreover," etc.
  # Punctuation rules
  punctuation:
    em_dashes: avoid
    semicolons: avoid
    oxford_commas: avoid
    unicode_ellipsis: never
  # Formatting rules
  formatting:
    bold_first_bullets: never
    excessive_headers: avoid
    erratic_bolding: never
    emoji_in_professional: never
  # Rhythm rules
  rhythm:
    sentence_length_variance: high    # std dev > 6 words
    paragraph_length_variance: high   # mix short and long
    uniform_subsections: never        # vary section lengths
  # Tone rules
  tone_guards:
    relentless_positivity: avoid      # be critical when warranted
    false_balance: avoid              # make bold claims
    emotional_flatness: avoid         # show genuine reactions
    elevated_register_for_simple: avoid  # match register to content

Integration

Works with other voice-framework skills:

• Created voices can be applied via voice-apply
• Created voices can be inputs to voice-blend
• voice-analyze can create base profiles that voice-create refines
• All created voices automatically include AI tells avoidance

References

• AI Writing Tells: ../voice-apply/references/ai-tells.md