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

# Guardrails

> Set hard safety boundaries that prevent your AI from crossing privacy, legal, or escalation limits.

Guardrails are the rules an agent must never break, whatever a customer asks and whatever sits in Knowledge, Skills or a tool result. Every message passes through them before it is sent, and a triggered guardrail blocks or escalates the reply.

## Where to find them

Guardrails sit in the [Persona](/build-your-agent/what-goes-where/style) tab, below Instructions. Open **Agents** from the workspace sidebar, select an agent, open **Persona**, then scroll down to Guardrails. Each guardrail has a name and a single rule, and you add a new one with the plus button on the right.

## What goes in a guardrail

* Privacy limits, such as withholding account details until identity is confirmed.
* Legal limits, such as not making price comparisons.
* Escalation limits, such as not promising a refund above 30 percent without approval.
* Safety limits, such as never sharing a customer's identity with a third party.

## Examples

These are real guardrails from a live workspace:

* **Address Sharing** stops the agent giving out the exact address unless the guest has a confirmed reservation.
* **Pre-Stay Message Restriction** stops it telling guests to enjoy their stay while the dates are still in the future.
* **Discount and Price Change Restrictions** stop it offering a discount or changing the reservation price unless something else explicitly authorises it.
* **Recommendation Availability Rule** stops it recommending places that are not in the listing information.

Write them in absolute language. A rule that only applies sometimes belongs in a Skill.

## Skill or guardrail?

A Skill fires when a scenario matches, such as explaining that identity verification is required when a customer asks for account details.

A guardrail applies to every reply, such as never sharing account details without verified identity.

## What belongs somewhere else

* Product and property facts belong in [Knowledge](/build-your-agent/what-goes-where/knowledge).
* Conditional instructions belong in [Skills](/build-your-agent/what-goes-where/procedures).
* Tone and character belong in [Persona](/build-your-agent/what-goes-where/style).


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