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

# Instructions and guardrails

> Give your agent rules that apply to every turn, and set guardrails that check responses and end or transfer the conversation when a rule is broken.

The script tells the agent *what* to cover. **Agent Instructions** tell it *how* to behave throughout the conversation: tone, persona, facts to keep in mind, things it must never do. **Guardrails** go a step further: they check the agent's responses against rules you define and act when a response breaks one, before the contact hears it.

| | Agent Instructions | Guardrails |
| - | - | - |
| What it is | Extra guidance added to the agent's prompt | An independent check on each response |
| Scope | One agent | Organization-wide, plus per-agent custom rules |
| When it applies | Every turn, as the agent writes its reply | After the reply is written (or while it streams) |
| What happens on a violation | Nothing extra; the agent is simply guided | The conversation ends or is transferred to a human |
| Channels | Voice, Text, Email | Voice |

## Agent Instructions

**1. Open General Settings**

In the agent editor, click **General Settings** in the header.

**2. Expand Agent Instructions**

Open the **Agent Instructions** section and type your guidance (**Type instructions here**). You can insert [variables](https://docs.callers.ai/documentation/build-your-agent/agent-variables.md) from the **Variables** tab.

**3. Save**

Click **Save** (and **Publish**, if your account uses [versions](https://docs.callers.ai/documentation/build-your-agent/versions-and-split-traffic.md)). Instructions apply to new conversations.

### What to put in instructions

> **Persona and tone**
>
> `You are Maya, a friendly scheduling assistant for Bright Smile Dental. Speak casually and keep answers under two sentences.`

> **Business facts the agent always needs**
>
> `Our office is open Monday to Friday, 8am–6pm. We don't accept walk-ins. The contact's current plan is {{planName}}.`
>
> Longer reference material (price lists, policies, product specs) belongs in the [knowledge base](https://docs.callers.ai/documentation/using-callers/knowledge-base.md).

> **Rules and limits**
>
> `Never promise a discount. If asked about refunds, explain that the billing team will follow up by email. Don't ask for credit card details.`

> **How to use other sources**
>
> `When quoting a fee, use the exact figure from the knowledge base. If you can't find it, say you'll have someone follow up.`

> **Language and formatting**
>
> `Say dates as "Tuesday, March 3rd", not "03/03". Spell out phone numbers digit by digit.`

> [!TIP]
>
> Write instructions as short, direct rules. Put "always" and "never" rules first. Avoid repeating the script; instructions that contradict a step make the agent inconsistent. See [Prompting voice agents](https://docs.callers.ai/documentation/best-practices/prompting-voice-agents.md#agent-instructions-and-guardrails) for more.

Callers can also run an agent on instructions alone, without following the script. Contact support to adjust this setting.

## Guardrails

Guardrails are rules such as "Never give medical advice" that Callers checks the agent's responses against. If a response breaks a rule, the agent doesn't say it; instead it ends the conversation or hands the call to a person.

### Create organization guardrails

Organization guardrails are shared by every agent that uses them. Only admins can manage them.

1. In the sidebar, click **Guardrails**.
2. Click **Add Custom** (or **Create** on an empty page).
3. Enter a **name** (**Guardrail name**) and describe what to block (**Describe what this guardrail should block**). Be specific, for example: "Block any content that provides personalized financial or tax advice, recommends specific investments, guarantees returns, or describes investments as risk-free."
4. Click **Save**. New guardrails are on by default; use the toggle on each row to turn one off.
5. Click **Settings** to choose **What happens when the guardrail is triggered** for every agent that uses organization settings.

Changes on this page save immediately.

> [!NOTE]
>
> Guardrails are enabled per account. If you don't see **Guardrails** in the sidebar, contact [support@callers.ai](mailto:support@callers.ai) to turn them on.

### Turn on guardrails for an agent

Open the agent, click **Agent** in the header, then click **Guardrails** ("Set clear boundaries for your agent. Limit what it can say and do in production to reduce risk and maintain consistency.").

| Setting | What it does | Default |
| - | - | - |
| Guardrails toggle | Turns guardrail checks on for this agent. | Off |
| **Use organization guardrails** | "Apply your organization's guardrails to this campaign." Expands to list them. | On |
| **When to check the agent's response** | **Before speaking**: check the full response before it's spoken. **Streaming**: check each part of the response as it's generated. | Before speaking |
| **What happens when the guardrail is triggered** | **End conversation** ("The agent ends the call.") or **Transfer to human** ("The call is transferred to a human operator."). Locked to the organization setting while organization guardrails are in use. | End conversation |
| Custom guardrails | Click **Add** to create rules that apply to this agent only. | — |

**Before speaking** is the safer choice: nothing reaches the contact until it passes. **Streaming** responds faster but checks in pieces.

### Transfer to a human

When the action is **Transfer to human** (and the agent isn't using organization guardrails), set:

| Field | Notes |
| - | - |
| **Transfer message** | Spoken before the handoff. Example: "Let me connect you with a member of our team." |
| Destination | A phone number or SIP URI, or a [human operator](https://docs.callers.ai/documentation/human-operators/overview.md) queue. Required. |
| **Caller ID** | **Agent's number** (default) or **Contact's number**. |
| **Warm transfer** | "Play a summary of the conversation to the receiver." Add a **Summary prompt**, choose **Who hears the summary** (**Receiver only** or **Both parties**), and **When to play the summary** (**When call is connected** or **When ⁕ is pressed**). |

See [Call transfers](https://docs.callers.ai/documentation/voice/call-transfers.md) for more on warm transfers.

### Writing good guardrails

* **One topic per guardrail.** "No medical advice" and "No pricing commitments" should be separate rules so you can see which one fired.
* **Describe the behavior, not keywords.** "Block any statement that diagnoses a condition or recommends a medication or dosage" works better than a list of words.
* **Leave room for safe answers.** Say what's allowed too: "General information about office hours and booking is fine."
* **Test the edges.** Place test calls that try to push the agent over the line, and check that legitimate questions still get answered.

Common organization guardrails include: no medical advice, no legal or financial advice, no pricing or contract commitments, and staying on topic.

## FAQ

> **Do guardrails slow the agent down?**
>
> Checking adds a short delay before the agent speaks, mainly with **Before speaking**. **Streaming** reduces the delay.

> **Do guardrails apply to Text and Email agents?**
>
> Guardrail settings are available on Voice agents. For Text and Email agents, put your rules in **Agent Instructions**.

> **Can I see when a guardrail fired?**
>
> The call ends or transfers right after the blocked response. Review the transcript of that call to see where it happened.

## Related

> **[Agent settings](https://docs.callers.ai/documentation/build-your-agent/agent-setup.md)**
>
> Where every agent setting lives.

> **[Knowledge base](https://docs.callers.ai/documentation/using-callers/knowledge-base.md)**
>
> Reference material the agent can look up.

> **[Human operators](https://docs.callers.ai/documentation/human-operators/overview.md)**
>
> Queues that can receive guardrail transfers.

> **[Scenarios](https://docs.callers.ai/documentation/build-your-agent/agent-scenarios.md)**
>
> Prepared responses for specific situations.


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