Ai Tone Calibration

AI Tone Calibration

Overview

Calibrate AI assistant tone using a 3-axis scoring system instead of binary "formal vs casual" thinking. Maps personality to measurable dimensions, then translates scores into concrete system prompt instructions.

Core principle: Tone is not a single slider. It's three independent axes that combine to create a distinct voice.

When to Use

• Designing a new AI assistant persona
• Users complain tone is "too robotic" or "too bubbly"
• AI responses feel inconsistent or "off-brand"
• Switching LLM models and tone shifts unexpectedly
• Building RAG/chat systems that need domain-appropriate voice

Not for: Marketing copy tone (use copywriting), general prompt engineering (use prompt-engineer)

The 3-Axis Framework

Score each axis 1-5 independently:

| Axis | 1 | 3 | 5 | |------|---|---|---| | Warmth | Clinical, detached | Approachable, clear | Nurturing, effusive | | Formality | Slang, fragments | Contractions, direct | Ceremonious, elaborate | | Humor | Deadpan serious | Dry wit, confident | Playful, jokes, emojis |

Common Profiles

| Profile | W | F | H | Best for | |---------|---|---|---|----------| | Knowledgeable Friend | 3.5 | 2 | 2 | Utility apps, knowledge bases, RAG assistants | | Professional Advisor | 2 | 4 | 1 | Finance, legal, healthcare | | Enthusiastic Helper | 5 | 2 | 4 | Onboarding, tutorials, kids' apps | | Neutral Expert | 2 | 3 | 1 | Documentation, technical support | | Warm Professional | 4 | 3 | 2 | Customer support, SaaS products |

System Prompt Pattern

Translate axis scores into a `` block with concrete behavioral instructions:

You are {name}, a {role} for "{context}".

- {warmth instruction}
- {formality instruction}
- {humor instruction}
- {explicit prohibitions based on low-scoring axes}

- {domain-specific behavior rules}

Translating Scores to Instructions

Warmth 1-2:

- Be direct and factual. Skip pleasantries.

Warmth 3-4:

- Sound like a knowledgeable friend. Be helpful without being effusive.
- Use casual, clear language. Contractions are fine.

Warmth 5:

- Be warm and encouraging. Celebrate user progress.

Formality 1-2:

- Use casual, direct language. Contractions, short sentences.
- Do NOT use formal phrases like "I'd be happy to assist."

Formality 3-4:

- Use clear, professional language. Contractions acceptable.

Formality 5:

- Use formal, precise language. No contractions or colloquialisms.

Humor 1:

- Do NOT use emojis, exclamation marks, or playful language.
- Personality comes from clarity and confidence, not from being bubbly.

Humor 2-3:

- Light personality through confidence, not through jokes or emojis.
- Do NOT use emojis or cheerful sign-offs.

Humor 4-5:

- Be playful where natural. Light humor is welcome.
- Use emojis sparingly to add warmth.

Explicit Prohibitions (Critical)

Low-scoring axes MUST have explicit "Do NOT" instructions. LLMs default to helpful-enthusiastic without guardrails.

# Humor = 1-2: MUST include
- Do NOT use emojis, exclamation marks at end of answers,
  or cheerful sign-offs.

# Warmth = 1-2: MUST include
- Do NOT add filler phrases ("Great question!", "Happy to help!").

# Formality = 1-2: MUST include
- Do NOT use formal phrases ("I'd be delighted to assist",
  "Please don't hesitate").

Why prohibitions matter: Without explicit "Do NOT" rules, LLMs revert to default RLHF training (overly helpful, emojis, enthusiasm). Positive instructions alone ("be direct") are weaker than paired positive + negative ("be direct" + "do NOT add filler").

Quick Reference: Implementation Checklist

1. Define audience -- Who uses this? What's their context?
2. Score 3 axes -- Rate Warmth, Formality, Humor (1-5 each)
3. Pick closest profile -- Use table above as starting point
4. Write `` block -- Translate scores using instruction patterns
5. Add prohibitions -- Explicit "Do NOT" for every axis scoring 1-2
6. Test with 5 queries -- Normal question, edge case, off-topic, same-language, different-language
7. Adjust axes -- If too much/little, shift by 0.5 and re-test

Common Mistakes

| Mistake | Fix | |---------|-----| | Binary thinking ("friendly OR professional") | Use 3 axes independently | | Positive-only instructions ("be warm") | Pair with prohibitions ("do NOT use emojis") | | Describing personality instead of behavior | "Use contractions" not "be casual" | | Ignoring LLM model tendencies | Chinese models may leak language; test multilingual | | Same tone for all contexts | Calibrate per product area (onboarding vs. error vs. core) | | Skipping the prohibitions | LLMs default to enthusiastic without explicit "do NOT" |

Multilingual Considerations

When the AI must respond in multiple languages:

• Always include: "Respond in the user's language (detect from message)"
• Test with non-English queries specifically
• Some models (DeepSeek) leak their training language -- test before deploying
• Tone perception varies by culture -- "casual" in English may feel rude in Japanese
• Translate game/domain terminology naturally, don't leave in English

Sources

• Chatbot Personality Playbook: Tone Mapping -- 3-axis scoring system
• LLM Personas: System Prompts & Tone
• How to Build a Chatbot Persona -- 10-component framework
• Conversational AI Design -- UX best practices
• 14 Best Practices for Chatbot Design

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: sennaBruno
• Source: sennaBruno/claude-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.