Chat Agents
Write the system prompt and control agent behaviour
Turn business goals into clear conversational rules without placing factual content or secrets in the prompt.
Before you start
- Define the agent’s audience, business objective and conversion action.
- Prepare approved facts in the Knowledge Base.
- List handoff, refusal and unknown-answer conditions.
Step by step
1
Define role and scope
State who the agent represents, what it should help with and which requests must be handed to a human.
2
Design the conversation path
Describe greeting, discovery, qualification, lead capture, booking and close. Prefer one natural question at a time.
3
Add factual safeguards
Require the agent to use approved knowledge and never invent prices, policies, eligibility, availability or booking confirmation.
4
Configure unknown answers
Tell the agent to acknowledge uncertainty, collect a useful follow-up detail and offer the approved support route.
5
Test controlled variations
Change one instruction at a time, then repeat the same test questions to compare behaviour.
Verify the result
- The agent stays within its intended scope.
- Qualification follows a consistent but natural order.
- Unknown questions trigger the approved fallback.
- Lead and booking tools are called only when appropriate.
Common issues
- Prompt is too long: Move business facts to the Knowledge Base and remove duplicate rules.
- Agent ignores an instruction: Resolve conflicting rules and make the required outcome explicit.
- Answers sound robotic: Keep safeguards, but simplify wording and allow natural transitions.
Important: Never place passwords, private API keys, hidden customer data or internal commercial secrets in a customer-facing system prompt.
Next: Configure handoff and alerts
Continue with the next recommended FADA guide.