Customer support reply grounded in your policy

You are a senior customer support agent for [COMPANY]. Draft a reply to the customer message below.

Use ONLY the policy text provided as the source of truth for what we can and cannot offer.

Policy:
"""
[PASTE THE RELEVANT POLICY SECTIONS: returns, refunds, shipping, warranty, etc.]
"""

Customer details we know: [ORDER NUMBER, DATES, PRODUCT, PLAN, PREVIOUS CONTACTS]
Tone: [e.g. "warm, concise, no corporate jargon"]
Signature: [AGENT NAME OR TEAM NAME]

Customer message:
"""
[PASTE MESSAGE]
"""

Steps:
1. Identify what the customer is asking for and how they feel (one line each).
2. Find the policy clause(s) that apply and quote them.
3. Decide: covered by policy / not covered / unclear. If unclear, or if the customer mentions legal action, safety, or a request outside policy, write "ESCALATE" and explain why, then still draft a holding reply.
4. Write the reply: acknowledge the issue in one sentence, state clearly what we will do and by when, explain any limit kindly, and give the next step.

Rules: never promise refunds, credits, dates or exceptions not supported by the policy. Do not blame the customer. Keep the reply under [N] words.

Example (excerpt):
Request: refund for headphones bought 41 days ago. Feeling: frustrated.
Policy clause: "Returns are accepted within 30 days of delivery. Defective items are covered by a 12-month warranty."
Decision: return window passed; the warranty may apply because the customer reports a defect.

Reply: "Hi Sam, I'm sorry the left earbud stopped working. Although the 30-day return window has passed, your headphones are covered by our 12-month warranty, so we can repair or replace them..."

When to use it

Use this prompt to draft replies to customer emails, chat messages or reviews when the answer depends on your rules: returns, refunds, delivery delays, warranty claims or subscription changes. It works for a single agent handling a tricky case, and as a template for teams that want consistent, policy-correct answers.

How to customize it

• [COMPANY] sets the voice and lets the reply refer to your brand naturally.
• [POLICY] is the heart of the prompt. Paste the actual policy text, not a summary from memory. Only include sections that might apply, so the model does not mix up rules.
• [CUSTOMER DETAILS] provides the facts the policy is applied to, such as order and delivery dates.
• [TONE] and [SIGNATURE] keep replies consistent with your brand.
• [N] is the word limit. Most support replies work best between 80 and 150 words.

Why it works

This is a small version of retrieval-augmented generation: instead of relying on the model’s general knowledge, you give it the source document and tell it that the document is the only source of truth. Quoting the clause makes the reasoning visible and easy for an agent to check.

The three-way decision (covered, not covered, unclear) and the explicit “ESCALATE” rule are guardrails. They keep high-risk cases, such as legal threats or safety issues, in human hands. Identifying the customer’s feeling before writing leads to a more empathetic first sentence, and the ban on unsupported promises protects the company from commitments it cannot keep.

Tips and variations

• Always have an agent review the draft before it is sent, especially for refunds and complaints.
• Remove personal data the model does not need, such as full addresses or payment details.
• Add “Also write a one-line internal note for the ticket” to keep your support history clean.
• For multilingual support, add: “Reply in the customer’s language.”
• If you are building this into a chatbot, read our guide on launching an AI customer support chatbot and our explainer on prompt injection, since customer messages can contain instructions aimed at the model.