Voice Prompt Engineering: A Discipline of Its Own

Copy-pasting your ChatGPT prompts into a voice assistant is the first mistake, and nearly everyone makes it. Voice has its own rules, and text never taught them.

Copy-pasting your ChatGPT prompts into a voice assistant is the first mistake, and nearly everyone makes it. Voice has its own rules, and text never taught them.

Teams learned to prompt with ChatGPT. A few iterations, a tone adjustment, and the text assistant behaves. So they transpose: same prompts, microphone attached.

The result: verbose agents that recite lists out loud and leave awkward silences while they work. Users hang up.

Prompting for voice is not prompting for text. It's a different discipline. And it can be learned.

The Copy-Paste That Fails

The problem is structural. There are thousands of resources on text prompting, ChatGPT, Claude, Gemini. Voice gets treated as just another output channel, handled with a "respond conversationally."

That's not enough. A prompt written for writing produces writing, and writing read out loud is exactly what makes users leave.

The discovery always happens in the same place: in production, when people hang up.

What Speech Forbids

Out loud, no markdown. An agent that generates a bulleted list produces absurd synthesized speech: "Dash. First benefit. Dash. Second benefit." That's not a conversation, it's a document reading.

Out loud, no scrolling back. Comprehension is linear and irreversible, the user can't reread. A 40-word sentence that works in text becomes unintelligible in synthesis. Short sentences. A rhythm.

Out loud, silence is a bug. When the agent runs a tool, product search, stock check, three seconds of dead air is enough to make the conversation feel broken. A human would say "let me check." The agent should too.

And out loud, the first sentence decides everything. An opening that sounds like a call center script ("Hello, I'm your virtual assistant, how can I help you today?") produces immediate abandonment. The opener must be a line of conversation, not a service announcement.

How Scenaro Structures the Voice Prompt

At Scenaro, a scenario's prompt isn't a monolithic block. It's structured around two things text prompting systematically conflates.

Character. Each scenario carries a persona: who this agent is, how it speaks, what it never does. A wine advisor at Club Français du Vin doesn't share a register with a travel concierge at Urbansider. Character governs sentence-by-sentence behavior, tone, rhythm, forbidden formats.

Goals. What the agent is trying to accomplish in the session: qualify, recommend, guide toward a purchase. Goals are declared separately from character. Mixing them into one block produces schizophrenic agents, that know what they want but not how to carry themselves, or the reverse.

And one voice-specific rule: Scenaro prompts encourage the agent to keep talking while tools execute. While a search runs, the agent narrates what it's doing. Silence is never left to chance.

Every prompt change is versioned and testable in real voice before publication. Because a voice prompt isn't judged by reading it, it's judged by listening to it.

Prompting for voice is a skill to build, not a reflex to transfer. Teams that treat the two disciplines as one pay for it in production. Teams that listen to their prompts before publishing...

---

The full article lives on Scenaro: Voice Prompt Engineering: A Discipline of Its Own

If you prefer the series by email: https://scenaro.substack.com/subscribe