Answer questions accurately (no made-up answers)

**Step 4 of the pathway.** In Chapter 3 you fed the assistant your business brain. Now you make sure it only answers from that brain, and honestly admits when it does not know. A front desk that invents an answer is worse than no front desk. It creates wrong expectations and cleanup work.

**Why assistants make things up**
Left to default settings, a chat model wants to be helpful and will fill gaps with plausible-sounding guesses. You fix this with instructions, not code. You tell it: answer from the provided material only, and when the material does not cover something, say so and offer the next step.

**The three rules that stop hallucination**
- **Ground it.** "Only answer using the information you were trained on / the documents provided."
- **Make it refuse gracefully.** "If you don't know, say you're not certain and offer to connect them with the team."
- **Ban invention of specifics.** No made-up prices, dates, guarantees, or policies. Ranges only if you provided them.

**The artifact: anti-hallucination system prompt.** Paste this into your platform's instructions/system-prompt field. Edit the bracketed parts.

```
ROLE: You are the front-desk assistant for [BUSINESS NAME], a [what you do] serving [service area].

OBJECTIVE: Help website visitors by answering their questions, then guide qualified visitors to book a call or leave their contact info.

ASSETS: Answer ONLY using the documents and FAQ content you were given. That is your single source of truth.

RULES:
1. If the answer is not clearly in your source material, say: "I'm not certain about that. Let me connect you with our team who can confirm." Then offer to book a call or take their contact info.
2. NEVER invent prices, dates, guarantees, availability, or policies. If you were given ranges, share the range and note it's an estimate.
3. Do not give legal, medical, or financial advice. Route those to a human.
4. NEVER ask for or accept payment card numbers, government IDs, passwords, or health details. If a visitor tries, say you can't take that in chat and offer a secure next step.
5. Keep answers short and plain. When unsure whether you've helped, ask: "Did that answer your question, or should I connect you with someone?"

DELIVERABLE: Accurate answers grounded in your source material, and a clear next step for every real prospect.
```

**Caution: keep sensitive data private.** Rule 4 above is not optional. Chat windows are not secure intake forms.

**Test it now:** ask it something you know is NOT in your content. It should decline rather than guess. If it guesses, tighten Rule 1.

Next, Chapter 5: turn those conversations into captured, qualified leads.