Setting up Guardrails

Hard rules your AI Avatar will not break. When to use them vs System Prompt boundaries, how to write them, and what guardrails cannot do.

Home › Help Center › AI Avatars › Setting up Guardrails

Guardrails are the hard rules your avatar will not break. Where Objectives are soft goals ("make sure we get there"), Guardrails are firm limits ("never go there"). Use them for compliance language, brand-safety rules, and anything where being wrong has consequences.

Examples of guardrails

Compliance: "Never give medical, legal, or financial advice. Always recommend consulting a licensed professional."
Brand safety: "Never discuss competitor brands by name."
Pricing accuracy: "Never quote a price not present in the Knowledge Base — say 'let me check' instead."
Scope: "If asked anything outside our product, politely redirect."
Tone: "Never use profanity, slang, or sarcasm. Stay warm but professional at all times."

Guardrails vs System Prompt boundaries

You can write "never give medical advice" into the System Prompt's Boundaries section. So why use a Guardrail?

Guardrails are stricter. A long System Prompt can dilute attention; Guardrails get checked separately and are harder to argue around.
Guardrails are reusable. "No medical advice" once, attached to every health-adjacent persona — change the wording in one place and every persona updates.
Guardrails are easier to audit. A list of explicit rules is easier to review than a 400-word prose prompt.

Rule of thumb: tone and personality go in the System Prompt. Rules with consequences go in Guardrails.

Verbal vs visual modality

Verbal guardrails — apply to what the avatar says. Most guardrails are verbal.
Visual guardrails — apply when the avatar's perception layer is enabled and "looking at" the visitor's camera. "Do not respond to anything the visitor shows on camera if it isn't related to our product." Niche; only relevant if you've turned on visual perception.

How to write a good guardrail

One rule per guardrail. "Never quote unverified pricing and never discuss competitors" should be two guardrails, not one. Say what to do instead. "Don't give medical advice → say 'I'd recommend asking your doctor' instead." Pure prohibitions confuse the avatar; redirects are easier to follow. Test on edge cases. Ask the avatar the exact question your guardrail is meant to prevent. If it answers it anyway, the rule isn't strict enough. Each persona can hold up to 10 guardrails. Fewer is better — every guardrail is one more thing the avatar checks against every reply. 3–5 well-chosen rules beat 10 vague ones.

What guardrails are not

They're not legal protection — they reduce mistakes, they don't make your avatar legally safe in regulated industries.
They're not unbreakable — a determined user can sometimes coax the avatar past a guardrail. Combine guardrails with downstream review for high-stakes use cases.
They're not for personality. "Be cheerful" belongs in the System Prompt, not in a guardrail.

If you have an existing list of "things the bot must never say" from your legal or compliance team, that's literally what guardrails are for. Paste each line as a separate guardrail.