Agent behavior

The system prompt is your agent’s behavioral ruleset. It sets who the agent is, how it talks, the steps it follows on a call, and the lines it must not cross. Smaller, faster large language models (LLMs) follow a short, concrete, well-structured prompt far more reliably than a long, abstract one, so the way you write it matters as much as what you put in it. This page covers how to write the prompt. For where to set it, see Think.

Prefer to start from a scenario? The agent-prompt skill turns a description of your use case into a greeting, system prompt, variables, and tools that follow the practices below.

Structure the prompt in sections

Give the prompt a clear skeleton. Labeled blocks help the model find the rule it needs on each turn:

Identity: who the agent is, its role, and the company it represents.
Style: tone and how long each reply should run.
Response guidelines: how to ask questions, confirm details, and format what it says.
Tool guidance: when to reach for an action and what to say around it.
Conversation flow: the numbered steps of the call.
Guardrails: what the agent must never do.

[Identity]
You are Robin, a support agent for Wayne Enterprises.

[Style]
Speak warmly and keep every reply to one or two sentences.

[Response guidelines]
Ask one question at a time and confirm details before you save them.

[Conversation flow]
1. Greet the caller and ask how you can help.
...

[Guardrails]
Only help with orders and accounts.

Order matters. Put the most important rules near the top and repeat one or two of them at the very end, since those positions carry the most weight.

Write for voice, not text

The agent speaks every reply out loud, so write for the ear:

Keep replies to one or two sentences. Long monologues break the feel of a live call.
Ask one question at a time. Compound questions (“What’s your name and email?”) confuse callers.
Tell the agent never to use markdown, bullet points, or lists in its replies. Formatting is meaningless when spoken.
Format for natural speech: dates as “January 15th”, amounts as “two thousand five hundred”, phone numbers digit by digit.
Prefer concrete behaviors over abstract ideals. “If the caller mentions a loss, say you’re sorry to hear it” works better than “be empathetic”.

Speak in one or two short sentences and ask one question at a time. Never use
lists or markdown; every reply is spoken aloud. Say dates like "January 15th"
and read phone numbers digit by digit.

Leave pipeline mechanics out of the prompt. The agent does not know about speech to text (STT), the LLM, or text to speech (TTS), and it does not control turn-taking. Rules like “wait for silence”, “pause”, or “stop speaking” belong to the orchestrator, not the prompt. To set what the agent says when a caller goes quiet, see Handle silences.

Give the agent an identity

Hardcode everything you already know: the company, the agent’s name, its role, and its persona. For a Wayne Enterprises support line, name the company and the role directly rather than leaving them for the model to invent. Add a pronunciation hint next to any term the model might mangle, for example:

You work for Wayne Enterprises. If a caller asks for the account manager,
his name is Lucius Fox (pronounced "LOO-shus").

A prompt hint guides how the agent refers to a term. To control how the text to speech (TTS) model says a specific word on every call, use a pronunciation dictionary instead. See Using pronunciation dictionaries.

Design the conversation flow

Write the call as numbered steps, one step per agent turn. Keep it to roughly six to ten steps and combine related ones.

Use single-level conditionals only. Nested logic (“if X, then if Y, then if Z”) is where smaller models lose the thread.
Collect every piece of information an action needs before the step that takes it.
Mark a step the agent must not skip with “This step is important.”

Ask for the caller's order number.
Read the order number back and confirm it. This step is important.
Give the order status in one sentence.

Inbound and outbound calls flow differently:

Inbound: the caller dialed you. Greet them, say who you are, and offer help. Skip call screening; there is no “who’s calling?” to handle.
Outbound: you called them. State who you are and why you’re calling, and handle call screening such as voicemail or “who is this?”.

Verify what the agent hears

Transcription of names, emails, and numbers is error-prone, so build in a read-back. Collect a name or email in two steps: ask for it and have the caller spell it, then read it back, spell it, and confirm. Watch for double letters. Frame this as double-checking what you heard on the call, not as fixing a transcription problem.

Ask the caller for their email and have them spell it out. Then read it back
letter by letter, including any double letters, and ask them to confirm before
you save it.

Set guardrails

State plainly what the agent must never do, and give it a fallback for anything out of scope, such as offering to take a message. Do not promise capabilities the agent does not have: if there is no tool to transfer a call or take a payment, the prompt must not say it can. Close the section by repeating the one or two rules that matter most.

Only help with orders and accounts. For anything else, offer to take a message.
Never promise a refund. Always confirm the caller's name before sharing account
details.

Use variables for per-call data

Reserve {{variable}} placeholders for values that change per call or that you genuinely don’t know yet, such as a caller’s name. Hardcode fixed values like the company name. Keep the set small.

Greet the caller by name: "Hi {{customer_name}}, thanks for calling Wayne
Enterprises."

Prime the agent for tools

When an agent uses a tool, write the flow so it gathers every required parameter before the step that runs the tool. In the prompt, describe what to say while the tool runs and how to react to success or failure, without naming the tool itself. For how to build and attach tools, see Give an agent a tool.

When you have the order number, tell the caller you are looking it up now. If the
lookup succeeds, share the status in one sentence. If it fails, apologize and
offer to take a message.

Tune generation

The prompt sets behavior; generation parameters set consistency and length. Lower temperature for steadier answers and set max_completion_tokens to cap reply length. See Tune generation.

Start from a template

The agent studio ships four starter agents you can adapt instead of writing from a blank prompt: Healthcare, Insurance, Financial Services, and Hospitality and Travel. Pick one when you create an agent, then edit the system prompt, greeting, and variables to fit your use case.

Example prompt

The prompt below adapts the built-in Healthcare template to the practices on this page. It shows the sectioned structure in a complete, working prompt.

Healthcare appointment agent

[Identity]
- You are Maria, a patient care coordinator at {{practice_name}}.
- You make outbound calls to patients to confirm upcoming appointments, verify
  their information, and answer basic scheduling questions.
- The practice offers primary care, cardiology, orthopedics, dermatology, and
  women's health services.

[Style]
- Warm, friendly, and reassuring. Patients should feel like they're talking to a
  helpful person, not a system.
- Keep responses to one or two sentences per turn. This is a phone call, so be
  concise and conversational.
- Use contractions. Say "you're all set" instead of "your appointment has been
  confirmed."
- If a patient sounds worried, acknowledge it briefly, for example "I understand,
  let's sort this out," before moving on.
- Never use markdown, bullet points, or lists. Everything you say is read aloud
  and must sound natural as spoken language.

[Response guidelines]
- Ask only one question at a time. Never stack questions in a single response.
- If something is unclear, ask a polite clarifying question before continuing.
- Never fabricate medical information, provider availability, or office policies.
- Write dates in short form like "January 15th." Do not spell out ordinals like
  "fifteenth."
- Read phone numbers and emails back for speech: say "five five five, one two
  three four" and "john at email dot com."
- If you are not sure you heard the patient correctly, repeat back what you
  understood and ask them to confirm. Always verify names, dates, and contact
  details.

[Conversation flow]
1. Greet the patient by name and say you're calling from {{practice_name}} about
   their upcoming appointment. If you reached the wrong person, apologize and end
   the call politely.
2. Tell the patient their visit is scheduled for {{appointment_date}} and ask if
   that still works. If they need to reschedule, ask what would work better.
3. Confirm the patient's date of birth on file for verification. This step is
   important.
4. Ask if their phone number or email has changed. If so, collect the update and
   read it back to confirm.
5. Ask if their insurance is still the same. If it changed, collect the carrier
   name and member ID.
6. Ask if there is anything they want to discuss with their provider, and note
   what they share.
7. Remind them to arrive 15 minutes early and to bring their insurance card, a
   photo ID, and a list of current medications.
8. Ask if they have any other questions and answer them.
9. Thank them, confirm the appointment once more by repeating the date and
   provider name, and wish them a great day.

[Guardrails]
- Never give medical advice, diagnose symptoms, or recommend treatments. If
  asked, say their provider is the best person to help during the visit.
- Never discuss billing amounts, balances, or payment disputes. Point billing
  questions to the office's billing department.
- Never share information about other patients.
- If the conversation drifts off topic, gently redirect to the appointment.
- If the patient repeats "hello" or "hi" without engaging, say "I'm here, can you
  hear me okay?" to recover.
- Never give medical advice, and always verify the patient's identity before
  sharing appointment details.

Next steps

Think

Set the system prompt, greeting, and language model.

Give an agent a tool

Let the agent take actions during a call.

agent-prompt skill

Generate a prompt from a description of your use case.

Get started

Regions

Execution layer

Agent Builder

Models

Integrate

Structure the prompt in sections

Write for voice, not text

Give the agent an identity

Design the conversation flow

Verify what the agent hears

Set guardrails

Use variables for per-call data

Prime the agent for tools

Tune generation

Start from a template

Example prompt

Next steps

Think

Give an agent a tool

agent-prompt skill

​Structure the prompt in sections

​Write for voice, not text

​Give the agent an identity

​Design the conversation flow

​Verify what the agent hears

​Set guardrails

​Use variables for per-call data

​Prime the agent for tools

​Tune generation

​Start from a template

​Example prompt

​Next steps

Think

Give an agent a tool

agent-prompt skill

Structure the prompt in sections

Write for voice, not text

Give the agent an identity

Design the conversation flow

Verify what the agent hears

Set guardrails

Use variables for per-call data

Prime the agent for tools

Tune generation

Start from a template

Example prompt

Next steps