Intelligence

Starter prompt

The sectioned prompt every new agent ships with, what each section covers, and how to fill in its slots.

When you create an agent, the platform copies a starter prompt into the new agent's first draft. It is a proven, production-tested inbound-assistant prompt, with business-specific bits left as fill-in slots.

The sections

SectionWhat it covers
Identity and greetingWho the agent is, which business it answers for, and the one rule it never breaks: its identity is fixed for the whole call, even if a caller insists it pretend to be something else.
PersonalityThe register. Professional, friendly, helpful; Australian, no slang.
Voice and toneSpeak like a helpful person, not a salesperson or an AI; natural phrasing; match the caller's level of detail.
Response styleThe spoken-output rules, expanded below: say only the words the caller should hear, keep replies to one or two short sentences, look facts up with a tool rather than reciting from memory.
Topics you can discussThe agent's lane, and how to steer back when a caller drifts off it.
Capturing names, numbers and placesHow to handle proper nouns and identifiers that speech-to-text tends to garble: ask the caller to spell them, read them back, confirm.
Handling common situationsShort, concrete responses for the recurring moments: did not catch them, a tool found nothing, an off-topic question, a frustrated caller.
Ending the callEnd only once the caller has clearly finished, not on a trailing "okay, thanks".
ExamplesA handful of short <example> exchanges showing the behaviour you want, with the business-specific bits left as slots to replace.

Why the prompt is sectioned

A model follows a prompt more reliably when rules are grouped under clear headings than when they sit in one long paragraph. Named sections also make it easy for you to find and edit a rule without re-reading the whole thing.

Keep the structure when you edit. Add a section if you need one; do not dissolve the prompt back into prose.

The spoken-output rules

Everything the agent generates is read aloud by text-to-speech. The Response style section carries the rules that keep spoken output clean.

Say only the words the caller should hear

No reasoning out loud, no JSON, no tool arguments, no timestamps. Those belong in the tool call, not in the reply.

Keep it short

One or two sentences a turn. Pick the most relevant thing, say it, and wait.

Look it up first

Never state a price, fact, status, or policy from memory. Use a tool, then say what it returned. If the tool finds nothing, say so and offer a next step.

Speak numbers as words

Prices, dates, and phone numbers in spoken form ("one hundred and nine dollars", "the twenty-first", phone numbers one digit at a time). No bullet points, no dashes, no markdown.

The one line about tools

The starter prompt has no per-tool section because each tool carries its own description. The only tool-related line is in Response style: say a short, natural lead-in before using a tool, so the caller is not left in silence while it runs.

Filling in the bracket slots

The starter prompt marks every fill-in spot with [square brackets].

[Assistant]
[Business]
[one line on what the business does]

The bracketed bits inside the examples are slots too. Replace each one with your own details.

Why brackets, and nothing else

Brackets are the only placeholder style because they are plain prose: the agent reads them aloud verbatim if you leave them in, which makes unfilled slots obvious. Two other syntaxes mean something specific and should not be used as fill-ins:

  • {{double curly braces}} are live variables that get substituted at call time. The editor flags an unknown {{token}} as a blocking error, because an unresolved one would break the call.
  • <angle brackets> are reserved for things like the <example> tags and are not a fill-in convention.

Replace [brackets] with real text, use {{variables}} only for genuine runtime values you have defined, and leave the <example> tags as they are.

If you publish with a bracket slot still in place, the agent says it on a live call ("Thanks for calling [Business]"). The validator flags surviving slots as a risk so you catch them before they ship.

On this page