> ## Documentation Index
> Fetch the complete documentation index at: https://docs.threetone.in/llms.txt
> Use this file to discover all available pages before exploring further.

# Write an agent prompt

> Define an agent's task, boundaries, first message, and response rules, then review prompt changes before publication.

A prompt defines the agent's instructions. Keep factual source material in the [knowledge base](/agents-platform/configure/knowledge-base/overview) and external actions in [tools](/agents-platform/configure/tools/overview). A prompt alone does not grant access to either.

<Frame caption="Recording pending: Write and review an agent prompt. Follow the written guide below.">
  <img src="https://mintcdn.com/threetonelabs-c2efc9c8/RV32i5440UZIaFHA/images/tutorials/recording-pending.webp?fit=max&auto=format&n=RV32i5440UZIaFHA&q=85&s=55f66d2eef49d8c92c9ac46f84bace35" alt="Recording pending for Write and review an agent prompt. No playable video is available yet." width="1600" height="900" className="w-full h-auto aspect-video" loading="lazy" noZoom data-path="images/tutorials/recording-pending.webp" />
</Frame>

## Define the task

Write the desired outcome in one sentence before editing the prompt. For example: help customers understand a support process using approved information, and direct account-specific requests to support.

<Steps>
  <Step title="Set the role and scope">
    State which requests the agent handles and which require a person or another system. Avoid a broad instruction to handle every customer request.
  </Step>

  <Step title="Specify how to use information">
    Tell the agent to use approved sources and acknowledge missing information. Do not invent prices, policies, account state, or actions that it cannot verify.
  </Step>

  <Step title="Define action boundaries">
    Explain when an action requires confirmation. An instruction to check an order or contact a team is useful only when the corresponding integration exists and the caller has access.
  </Step>

  <Step title="Set the response style">
    Describe the preferred length, tone, and follow-up behavior. Set the response language through the configuration's language policy rather than conflicting instructions scattered through the prompt.
  </Step>

  <Step title="Review the complete configuration">
    Read the prompt with the first message, knowledge attachments, tools, and language settings. Review the version diff before publishing the changes.
  </Step>
</Steps>

## Example instructions

This is an example of prompt structure, not a promise of deployed behavior. Adapt it to your approved support process.

```text theme={null}
Role
You help customers understand our published support process.

Task
Answer process questions using approved information. Ask a brief
clarifying question when the request is ambiguous.

Boundaries
Do not invent policies, prices, account details, or completed actions.
If information is missing, say what is unknown and explain the next step.
Do not request passwords, API keys, or payment-card details.

Responses
Use short, clear sentences. Give the next useful step before extra detail.
For account-specific changes, direct the customer to the approved support path.
```

## Prompt and first message

The prompt lives at `conversation_config.agent.prompt.prompt`; the first message lives at `conversation_config.agent.first_message`. The first message opens the interaction. It should match the task and avoid implying access to private account information.

For the example above, a suitable first message is: “Hello. I can help explain our support process. What do you need help with?”

## Review scenarios

Before publication, write expected outcomes for a normal question, an ambiguous request, missing information, an out-of-scope task, and a request involving sensitive data. Use these cases to review the prompt and attachments now, and retain them for runtime testing when execution is available.

<Note>
  Automated response-test invocation is currently gated in the documented implementation. A stored test definition or a published prompt does not prove that a test ran successfully. Check [current availability](/agents-platform/overview#current-availability).
</Note>

## Continue

Read [agent configuration](/agents-platform/configure/agent-behavior/overview) for the field structure and [versions and drafts](/agents-platform/deploy/versioning) for reviewing changes. The [Update agent reference](/api-reference/agents-platform/update-agent) defines the API request shape.

## Frequently asked questions

<AccordionGroup>
  <Accordion title="Should I put all source documents in the prompt?" id="faq-prompt-knowledge">Keep instructions focused. Use reviewed knowledge attachments for factual material and inspect their stored content before relying on them.</Accordion>
  <Accordion title="Does an instruction to take an action enable it?" id="faq-prompt-actions">No. The appropriate tool, resource access, and execution path must exist. The prompt should explain the action boundary without claiming a completed action.</Accordion>
</AccordionGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.