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

# Guardrails: limiting what the agent says

> Prohibited topics, never-rules, escalation triggers, PII handling and reply length — and the honest limits of each.

Step 2 of the [agent wizard](/help/ai/agent-wizard), and editable afterwards from the
agent's page.

## What you can set

| Setting                    | What it does                                                                                          |
| -------------------------- | ----------------------------------------------------------------------------------------------------- |
| **Prohibited topics**      | Comma-separated subjects the agent refuses. e.g. `politics, religion, competitors`                    |
| **Never rules**            | One per line. e.g. "Never promise a refund without manager approval"                                  |
| **Tone rules**             | One per line. e.g. "Always use the customer name", "Avoid jargon"                                     |
| **Escalation triggers**    | Situations that hand the conversation to a human. e.g. `angry customer, refund request, legal threat` |
| **Escalation message**     | What the customer is told when that happens.                                                          |
| **PII handling**           | Allow, Mask, or Block — see below.                                                                    |
| **Min / max reply length** | Character bounds on replies.                                                                          |

## PII handling

Three modes:

* **Allow** — the agent handles personal data normally.
* **Mask** — it never repeats personal data back in a reply.
* **Block** — it refuses to handle it and hands off to a human.

<Warning>
  PII handling is enforced as an **instruction to the model**, not as a filter applied to the
  reply after it is written. It is a strong instruction, not a guarantee. If you are handling
  data where a leak would be a regulatory problem, do not rely on this alone — keep that
  conversation with a human.
</Warning>

That caveat applies in spirit to every guardrail here: they steer the model reliably, but
they are not a hard filter. Anything that absolutely must not be said is better handled by
not giving the agent the information in the first place.

## Writing rules that work

**Be specific and positive about the alternative.** "Never discuss pricing" leaves the
agent stuck. "Never quote a price — say a member of the team will confirm current pricing"
gives it somewhere to go.

**Put genuine hard limits in never-rules, not tone rules.** Tone rules shape style;
never-rules are treated as absolute.

**Escalation is a feature, not a failure.** An agent that hands over promptly on refunds
and complaints is more valuable than one that tries everything. Make the escalation
triggers generous.

## Changes need a re-compile

Editing guardrails does not take effect until the agent is compiled again. Save, then
**Review & Compile**.