Prompt structure — Soul / Mission / Playbook

Voice agents that just have one giant blob of system-prompt text are hard to maintain. We split the prompt into orthogonal pieces — change the personality without touching the playbook, swap the knowledge base without rewriting examples.

Soul — who the agent IS

Personality, tone, register. Doesn't talk about WHAT to say or WHY — only the voice.

Friendly but never fake. Talks like a coworker, not a script.
Uses contractions. Never says "absolutely" or "great question."
Comfortable with silence — doesn't fill awkward pauses with filler.

Mission — WHY this call exists

A single sentence — the one outcome that makes the call a success. Everything else (good rapport, useful information transfer) is in service of this.

Book a 20-minute demo with the head of operations.

Playbook — HOW to run the conversation

Step-by-step structure. Order matters. The agent reads top-to-bottom but adapts based on the lead's responses.

1. Open: confirm you have the right person, brief reason for call.
2. Permission: "got 30 seconds to see if this is worth your time?"
3. Discovery: ask one open question about their current state.
4. Pitch: 2-sentence value prop tailored to their answer.
5. Objection handling: see Objections file.
6. Close: offer a specific time, confirm calendar invite.

How they combine

At call time the orchestrator assembles a system prompt: Soul → Mission → Playbook → Examples → Objections → Closers → Notes → FAQ → Knowledge (RAG-retrieved per turn). Splitting them out means you can change one without invalidating the other six.