Prompt wizard
The prompt is where nearly all agent behaviour comes from, and most voice platforms hand you a blank textarea for it. That works if you already know what a phone agent's prompt has to contain. If you don't, you write the three or four sections you thought of, deploy, and discover the rest from real callers — the ones who ask something out of scope, who ask whether you're a real person, who get angry, who trail off mid-sentence.
The prompt wizard interviews you instead. It works through six focus areas one at a time, asks about the situations you wouldn't have thought to cover, and assembles a 17-section prompt written for a voice call rather than a chat window.
If you'd rather write the prompt yourself, writing the system prompt covers the hand-written path and the rules a voice prompt has to follow. The wizard produces the same kind of artefact — you can edit anything it writes, at any point.
Where it lives
Prompts are their own library, separate from any one agent, because the same prompt is often worth reusing across several.
- Open the dashboard.
- Left sidebar → Configure → Prompts.
- New prompt to start the wizard, or click an existing prompt to reopen it.
How the interview works
The wizard asks one question at a time. Every question belongs to one of six focus areas, and it works through them in a fixed order:
| # | Focus area | Sections of the prompt it writes |
|---|---|---|
| 1 | Persona & Voice | Role, How You Sound |
| 2 | Capabilities & Boundaries | What You Can Actually Do, Guardrails |
| 3 | Discovery & Conversation Flow | Discovery Approach, Context to Hold, Conversation Arc, Opening Lines, Discovery Questions |
| 4 | Product Knowledge & Callers | Product/Domain Truth, Caller Types, Answering Questions |
| 5 | Escalation & Honesty | Escalation & Next Steps, Honesty Rules |
| 6 | Tone Pitfalls & Closing | What Sounding Like a Bot Looks Like, Closers, Example Conversations |
The first question is always the same and costs nothing: "What does this agent do, and what business or product is it for?" Everything after that is generated from your answers.
At most three questions per focus area, and 18 in total. Those two caps are enforced in code, not requested of the model — once an area has had three questions the wizard is told outright to move on, whatever it would otherwise have done. The totals are deliberately equal (3 × 6 = 18): if the overall cap were lower, the early areas could consume the whole budget and the last ones would never get asked about at all.
Progress moves in one direction only. The interview's position is derived from which area your last answer was filed under, so it cannot drift backwards into an area you've already finished.
Answering
Each question comes with three suggested answers — concrete, specific to what's already known about your agent, and fully editable. They exist because "what should the agent say if a caller asks for something it can't do?" is much easier to answer by editing a plausible example than from a blank box.
- Next submits your answer and generates the following question.
- Skip section jumps straight to the next focus area. The skipped question is recorded as deliberately skipped, so the wizard fills that area with general best practice rather than leaving it thin and circling back to it later.
- The ‹ › arrows page back through earlier questions and answers, read-only. Only one question is ever on screen — the interview is not a scrolling transcript.
- Restart clears the questions and starts the interview over. It deliberately keeps the draft — the prompt you've built so far is the floor, not something to be thrown away.
The draft pane
The prompt builds up on the left as you answer, and it is always directly editable. You are never locked out of it waiting for the interview to finish.
- Changed lines are shaded after each turn, so you can see what actually moved instead of re-reading the whole prompt. The shading clears as soon as you edit by hand.
{{variables}}are highlighted. Type{{for autocomplete, or use the Variables menu to browse. Insertions go through the browser's native undo stack, so Ctrl+Z reverses them.- A token count sits above the box. Hover it for the limits and what they cost — see prompt length and cost.
- While a turn is running, the draft is overlaid with a sweep and a label naming the sections being rewritten. Reading and scrolling stay available throughout.
Hand edits are never overwritten
The wizard writes the prompt one focus area at a time, merging into the sections that area owns rather than regenerating the document. If your draft no longer matches what those sections assemble to — because you edited it directly — your text is re-read as the new baseline before the next turn. Edit as much as you like, whenever you like.
This structure is also what protects the draft from a bad generation. A turn can only ever rewrite the two-to-five sections its own focus area owns, and a turn that fails outright changes nothing at all. A single malformed response cannot take the rest of the prompt with it.
Attaching knowledge base documents
Attach knowledge next to the answer box pulls content from your existing knowledge base documents into the wizard's context, so the prompt it writes reflects your real hours, policies and procedures instead of placeholders.
This is background reading for the drafting model only. It is not the same mechanism as the live retrieval your agent performs during a real call — attaching a document here does not attach it to any agent. Up to 6,000 characters are pulled in, split evenly across the documents you select.
Choosing the model
Model settings below the answer box selects which model runs the interview, and its temperature. The picker lists active OpenAI and Anthropic models; the default is gpt-4o-mini. Each option shows its estimated cost per message, because you are billed per message rather than per minute here.
This is the model that writes your prompt. It has nothing to do with the model your agent uses on calls — that is set per agent, and it is the one the prompt test runs against.
Free messages and billing
Every message that reaches the model counts: each question asked, each polish pass, and each prompt test reply. The hardcoded opening question does not, because no model call happens for it.
- 50 free messages per calendar month, per user. The counter resets the first time you use the wizard in a new month.
- Past that, each message is charged to your account balance at the selected model's rate for the real duration of that call. The badge in the header shows where you are; hover it for the model and the per-message estimate.
- The first billable message of a session asks you to confirm before anything is charged. Confirm once and you won't be asked again for the rest of that browser session.
- If the free messages are gone and your balance is empty, the wizard stops before calling the model rather than failing partway.
Prompt length and cost
Prompt length is measured in tokens, not characters, because tokens are what actually drive cost and latency.
| Threshold | Value | What happens |
|---|---|---|
| Free limit | 5,000 tokens | No warning, no surcharge |
| Hard limit | 32,768 tokens | The prompt is rejected above this |
Past 5,000 tokens the LLM portion of your per-minute call rate rises in steps — one extra multiple of the LLM rate per additional 5,000-token block, reaching roughly 7× at the hard limit. Only the LLM component is affected; text-to-speech and transcription are unchanged, since prompt length only affects LLM input.
A thorough wizard-built prompt typically lands in the 4,000–6,000 token range.
Saving and using the prompt
Save prompt stores it in the library. Save & test saves it and opens the prompt test against one of your agents.
To put a prompt on an agent:
- Left sidebar → Configure → Agents, and open the agent.
- Top tab → Agent Overview.
- Import from library next to the system prompt.
- Pick the prompt and choose Use.
Importing copies the text. Editing the library prompt afterwards does not change agents that already imported it, and editing an agent's prompt does not change the library entry. The link is recorded only so the library can show you how many agents use a prompt. This is deliberate: a prompt that silently rewrote itself across every agent using it would make a small library edit an unreviewed change to live call behaviour.
Next
- Prompt test — rehearse the prompt against a real agent before deploying it.
- Writing the system prompt — the rules a voice prompt has to follow.
- Testing your agent — end-to-end testing once the prompt is on an agent.