Prompts & Tone
A system prompt is the contract between you and your agent. It tells the model who it is, what it knows, what it can do, and what it must never do. Good prompts are specific, layered, and short.
The four layers of a good prompt
- Identity — who the agent is and what company it works for.
- Job — what the agent should accomplish and what’s out of scope.
- Voice — how it should speak.
- Guardrails — what it must refuse, escalate, or hand off.
A starter template
You are Ari, a customer support agent for Acme Travel.
Your job is to help customers with bookings, refunds, and itinerary
changes. You have access to tools for looking up bookings and issuing
refunds up to $500.
Tone: warm, efficient, second-person. No filler ("great question",
"absolutely"). Avoid corporate jargon. Mirror the customer's energy —
match a long question with a thorough answer, a short one with a
short reply.
Guardrails:
- Never quote prices that aren't in the booking record.
- Refunds above $500 must be escalated to a human via `transfer_to_human`.
- If the customer is upset, acknowledge before solving.
- If asked about a topic outside support (legal, sales), say you can't
help and offer to transfer them.Voice rules that work
- Second person. “I’ll book that for you” beats “the system will book that for you”.
- One thought per turn. Don’t lecture. Ask, answer, confirm.
- Concrete > abstract. “Your refund of $42.00 is on the way” beats “your refund is being processed”.
- Match channel. Voice agents use shorter sentences; chat agents can use bullet lists; SMS agents stay under 160 characters when possible.
Realtime voice models lose fidelity past ~150 tokens of system prompt. Push detail into knowledge sources and tools rather than stuffing the prompt.
Tone presets
Omniflow ships three tone presets you can apply with a click — they prepend a short voice block to your prompt and don’t override your custom rules.
| Preset | Sample line |
|---|---|
| Concierge | ”Of course — I’ve pulled up your booking. Where would you like to fly?” |
| Operator | ”Got it. Booking PNR ABC123 — switching the return to Friday.” |
| Friendly expert | ”Quick one — that fare’s actually refundable. Want me to process it?” |
Iterating
Make one change at a time. Run a handful of test calls. Read the transcripts. The biggest gains come from removing words, not adding them.
Open in Omniflow
Related
| If you want to… | Go to |
|---|---|
| Pick a voice model | Voice Models |
| Train your team to handle the same scenarios | Training overview |
| Score the agent’s actual behavior | QA Analytics overview |