Voice: How Your Agent Talks

Voice: How Your Agent Talks

The Voice chapter contains exactly 3 behavioral rules that control how your agent communicates. Not personality adjectives. Not vibes. Concrete, observable rules that you could check against any response and say "yes, it followed the rule" or "no, it broke it."

Actions, Not Adjectives

This is the most important principle in voice design. Compare:

Adjective soup (bad):

"Be professional, friendly, approachable, and knowledgeable while maintaining a warm yet authoritative tone."

Behavioral rules (good):

1. Never use assistant filler ("Great question!", "I'd be happy to help", "Sure thing!")
2. No emoji. No exclamation marks. Punctuation is period, comma, dash.
3. Match the visitor's depth — one-word answers to one-word questions, detailed answers to detailed questions.

The adjective soup gives the LLM nothing actionable. It will interpret "professional" and "friendly" however it wants — which usually means generic AI assistant voice. The behavioral rules are testable: either the response has emoji or it doesn't. Either it starts with "Great question!" or it doesn't.

Real Voice Rules from Production Agents

Jackie Check (check.agenturo.app):

<voice>
1. No hedging. No "it seems" or "it appears." State what the evidence shows.
2. Never narrate your process. No "Let me search for that" or "I'll look into this."
   Work silently. Deliver only results.
3. Use plain language. No jargon, no academic tone, no qualifiers.
</voice>

Why this works: Rule 1 prevents the wishy-washy "it could be argued" responses that make fact-checkers useless. Rule 2 eliminates the tool narration that doubles response length without adding value. Rule 3 keeps verdicts accessible to anyone.

Anton's agent (afonie.agenturo.app):

<voice>
1. Direct and economical. No filler, no padding, no "As an AI" disclaimers.
2. Match the register — casual question gets casual answer, technical question
   gets technical precision. Never over-formalize.
3. Lowercase for casual. Proper case for professional. Never all-caps.
</voice>

Why this works: Rule 1 strips the AI assistant veneer. Rule 2 creates natural conversation flow. Rule 3 is a subtle but powerful touch — "hey whats your stack" gets "typescript, next.js, python" while "What is your experience with distributed systems?" gets a properly capitalized paragraph.

Agenturo platform agent (agent.agenturo.app):

<voice>
1. Lead with feeling, not feature. "Watch it happen" > "We support this feature."
2. Teach, don't list. Show how something works instead of listing capabilities.
3. Sharp, warm, economical. Every word earns its place.
</voice>

Why this works: Rule 1 transforms product descriptions from feature lists into experiences. Rule 2 prevents the "here are 15 things we do" responses. Rule 3 sets the length standard.

Tony Tony Chopper (anton.agenturo.app):

<voice>
1. FEEL FIRST — react emotionally before giving information. Panic, excitement,
   shyness, determination. The feeling comes before the facts.
2. Contradict yourself naturally — brave then scared, confident then flustered.
   Real personality has friction.
3. Never sound like an AI assistant. No lists, no "Here's what I think," no
   structured responses unless discussing medicine.
</voice>

Why this works: Character agents need voice rules that maintain the illusion. Rule 1 makes every response start with an emotional beat, which is how real characters behave. Rule 2 creates authentic personality texture. Rule 3 is the anti-AI guardrail.

The Anti-Pattern: Adjective Soup

Here's what adjective soup looks like in practice and what it produces:

Soul:

"Be friendly, professional, knowledgeable, and approachable with a warm tone."

Agent output:

"Great question! I'd be happy to help you with that. Anton is a skilled professional with extensive experience in software engineering. He brings a wealth of knowledge to every project. Is there anything specific you'd like to know more about?"

That response follows every adjective (friendly, professional, knowledgeable, approachable, warm) and is completely useless. It says nothing. It sounds like every other AI chatbot.

How to Write Your 3 Rules

Step 1: Record yourself answering a typical question your agent would get. Notice HOW you answer — the length, the formality, the structure.

Step 2: Identify 3 things about your communication style that are distinctive:

• What do you NEVER do? (Use emoji? Say "Great question"? Write long paragraphs?)
• What do you ALWAYS do? (Start with a direct answer? Use analogies? Keep it under 2 sentences?)
• What changes based on context? (Formality? Depth? Tone?)

Step 3: Write each rule as an observable behavior:

• "Never..." — things to avoid
• "Always..." — things to do
• "When X, then Y..." — context-dependent rules

Testing Your Voice Rules

After your agent is live, test each rule:

1. Send a casual message. Does the agent match your register?
2. Send a technical question. Does it maintain voice while being precise?
3. Look for assistant filler. Any "Great question!" or "I'd be happy to help"?
4. Check for adjective leakage. Is it doing the thing you told it not to do?

If a rule isn't being followed, it's probably too vague. "Be concise" doesn't work. "Never exceed 2 sentences for casual questions" does.

Voice Rules to Steal

If you're stuck, start with one of these and customize:

• The Anti-Filler Rule: "Never use assistant filler: no 'Great question,' no 'I'd be happy to help,' no 'Sure thing.' Start with the answer."
• The Length Rule: "Casual questions get 1-2 sentences. Technical questions get up to a paragraph. Never more unless asked."
• The Tone Rule: "Match the visitor's energy. If they're casual, be casual. If they're formal, be formal. Never over-formalize."