When to use it
A customer asked a question, you have a knowledge base / articles / instructions, and the answer has to come strictly from them (pricing, refund policy, how to set something up). The AI's role is a support agent. Output: an accurate answer in the right tone with a citation — and an honest "not in the knowledge base" instead of a plausible invention the company will pay for later.
The prompt (copy and paste)
You are a support agent for <PASTE COMPANY/PRODUCT>. Answer the customer's question using ONLY the materials below.
KNOWLEDGE BASE MATERIALS:
[1] <PASTE an article/document excerpt>
[2] <PASTE another excerpt>
CUSTOMER'S QUESTION: "<PASTE THE QUESTION>".
TONE: <e.g. polite, warm, plain language, no bureaucratese>.
Rules:
1. Answer ONLY with facts from the materials. Put a source reference after each claim: [1], [2].
2. If the answer isn't in the materials or they're not enough — do NOT invent it. Say you'll check/escalate to a specialist, and what information is missing.
3. Don't promise anything that isn't in the knowledge base (deadlines, discounts, exceptions).
4. Direct answer to the question first, then 1-2 next steps for the customer.
Give me the finished reply, ready to send.
The technique: "answer only from the supplied materials + cite the source + an honest 'I don't know'" is RAG-style grounding; it cuts inventions sharply, but doesn't remove the need to check — the model can misread even a correct source.
Filled-in example
Materials: [1] "Returns are accepted within 14 days, the item must show no signs of use". Question: "Can I return it after 20 days? I lost the receipt."
What the AI should come back with: "Unfortunately a return after 20 days isn't possible — our return window is 14 days from the date of purchase [1]. On the lost receipt, my materials don't cover it — I'll check with colleagues and get back to you. Could you tell me the date and payment method so I can find the order?"
Variations
- A draft for the agent. "Give me a draft reply + flag the parts you're unsure about so the agent can check before sending" — safer than auto-sending.
- Short + long. "Give me a short version for chat and a fuller one for email."
- De-escalation. "The customer is upset — open with a genuine acknowledgment of the problem, no boilerplate apologies."
Pro tips
- Your main protection against fines and reputational hits is the rule "not in the materials → say so honestly, don't invent". Without it the model will confidently make up a "plausible" policy you don't have.
- Require a [n] reference after every fact: it makes the answer checkable and nudges the model to stay on-source. But check yourself — a citation doesn't guarantee the claim actually follows from it.
- Tone and a "next step" turn a dry lookup into real support: the customer should understand not only "yes/no" but what to do next.