A Grok Voice 3.0 system prompt works best in the shape xAI's prompting guide recommends: Markdown, written to the agent in the second person, with five ## sections in a fixed order. The generator above fills those sections from a few plain answers about your business, adds the escalation and safety rules most agents need, and keeps the greeting out of the prompt, as the guide asks.
The five sections, in order
| Section | What goes in it |
|---|---|
## Role & Persona | who the agent is, for which business, and its personality |
## Objective | what counts as done, and what to do when it cannot get there |
## Conversation Flow | how a call should go, from the caller's first answer; numbered steps if needed |
## Guardrails & Escalation | scope, refusals, when to hand over to a human, and the safety path |
## Voice & Communication Style | how the words should sound when spoken: length, language, numbers |
xAI's guide says prompts in this shape "sit closest to the training distribution" and behave most predictably. Two optional extras are allowed: a short ## Business Facts block for hours, prices and policies, and a ## CRITICAL INSTRUCTIONS section at the end for rules that must never be broken. Source: xAI speech to speech prompting guide, read 19 Oct 2026.
Rules the generator follows
- Facts go in word for word. Hours, prices and the full website address are pasted into the prompt, and the agent is told to say it does not know rather than guess.
- Tools named = tools enabled. The prompt mentions
web_searchor your transfer function only if you tick them, and the box under the prompt lists the tools to turn on in the session. A mismatch either way makes the agent worse. - No greeting inside the prompt. The opening line is a separate setting; the flow starts from the caller's first answer. In the API, send the greeting as a
force_messageso it is spoken exactly. - Short turns. One or two sentences per turn is the default, which also keeps calls, and the per-minute bill, short.
- A safety path. For any agent that might talk to someone in crisis, the guide treats the empathetic hand-off as required. It is on by default.
What a prompt cannot do
The prompt controls the words, not the voice. To change how the agent sounds, pick another of the 28 Grok voices, set audio.output.speed, or fix a brand name with the replace map; all three are options in the Grok Voice 3.0 session builder. Rules like "sound excited" or "speak slower" belong in the session settings, or need rewriting as rules about the words ("use short sentences", "pause after questions").
Tips from testing a prompt
- Change one thing at a time. Small wording changes can shift behaviour a lot; test each edit on the same few calls.
- Shorter beats longer on Grok Voice 3.0. xAI suggests a much shorter prompt than older models needed, and dropping workaround rules written for other models.
- Use capitals sparingly.
ALWAYS/NEVERhelp a rule stand out; when everything is in capitals, nothing is. - If the agent repeats itself, tick the variety rule. If it switches language on noisy lines, set a language lock.
To hear how the agent will sound before you build anything, try your script in the free Grok text to speech tool, which uses the same voices. For the rest of the build (sessions, tokens, phone numbers), see the Grok voice agent guide.
Last updated: