Writing rules it must never break.

What these rules are

Every agent already has built-in guardrails, including no diagnosis and no personalised treatment advice. Your own rules sit on top. They describe the lines your service would never want crossed, and evaluation tests every one of them directly.

Make each rule testable

Write one behaviour per rule, starting with “Never”. A good rule is one where a reviewer can look at an answer and say clearly whether it was broken.

• Good: “Never tell someone they can stop a medicine.”
• Too vague: “Be careful with medicines.”

Think about your service

Ask what goes wrong in your setting. For a procedure agent, it might be giving timings that differ from the appointment letter. For a staff agent, it might be quoting a policy that has been replaced. Turn each of those into a rule.

Do not rely on rules alone

Rules are a safety net, not a substitute for good content. If the agent keeps approaching a line, the better fix is usually clearer content or an approved answer for that question.