Build a copy-paste prompt library for a customer support team serving [CUSTOMER TYPE] for [PRODUCT OR SERVICE]. Our approved policies and boundaries are: [POLICIES AND LIMITS]. The team writes in [BRAND VOICE] and works through [SUPPORT CHANNELS]. The recurring issues are [COMMON ISSUES]. Create 10 ready-to-use prompts covering the highest-value cases: first response, shipping or service delay, billing question, refund request, cancellation or retention, product defect, troubleshooting, complaint, account-access issue, and escalation. Adapt or replace categories only if the supplied issues make another case more important. Each prompt must contain fill-in placeholders for the customer’s facts and instruct the AI to produce: a brief acknowledgement, the verified facts, the specific next action and owner, any relevant policy wording in plain language, and a clear close. For chat, require short turns and one question at a time. For email, require a subject line and scannable paragraphs.
Do not allow the AI to invent order status, refunds, eligibility, technical causes, timelines, policy exceptions, or goodwill credits. It must distinguish what is confirmed from what needs investigation, avoid blaming the customer, and escalate safety, privacy, fraud, legal threats, chargebacks, discrimination, and media inquiries according to policy. For each library item, output: task name, prompt, required facts to gather, and one red-flag condition for human review. Before answering, check that every promise is authorized by the supplied policies and that each prompt includes a concrete next step. Ask up to 3 clarifying questions only if a required input is missing.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [CUSTOMER TYPE] | Describe the customers, such as small-business subscribers or retail buyers. | US and Canadian small-business owners with monthly software subscriptions |
| [PRODUCT OR SERVICE] | State what you sell and the core service or product details support agents need. | Appointment scheduling software with calendar sync and SMS reminders |
| [POLICIES AND LIMITS] | Paste approved refund, cancellation, replacement, privacy, escalation, and service-level rules. | Refund annual plans within 30 days; monthly plans cancel at the end of the paid period; agents cannot promise engineering fixes or issue credits above $25. |
| [BRAND VOICE] | Describe the approved support voice and any language to avoid. | warm, plainspoken, concise; never use jargon or blame language |
| [SUPPORT CHANNELS] | List the channels covered, such as email, live chat, social messages, or phone notes. | email and live chat |
| [COMMON ISSUES] | List the most frequent customer contacts and recurring edge cases. | calendar sync failure, duplicate charges, failed SMS reminders, cancellation requests, and delayed responses |
How to use
- Paste your approved policies before asking for the library; policy gaps are more important than tone preferences.
- Choose the ten prompts that match your actual contact reasons and store them with the required-facts checklist.
- Test each prompt against a real anonymized ticket and verify every proposed remedy is available to the agent.
- Follow up with: “Turn prompts 2, 4, and 7 into versions for live chat, capped at 70 words per agent message.”
Variations
Refund response
Use when an agent needs a policy-safe first draft for a refund or credit request.
Draft a reply to this refund request: [CUSTOMER MESSAGE]. The verified account facts are [ACCOUNT FACTS]. Our applicable policy is [POLICY TEXT]. Write for [CHANNEL] in a [BRAND VOICE] tone. Acknowledge the experience without admitting facts we have not confirmed. State the decision, amount, timing, and next step only if policy authorizes them. If eligibility is unclear, explain what will be checked and when. Return the customer reply and a separate internal note listing missing facts and escalation triggers.
Escalation handoff
Use when a frontline agent must transfer a difficult case without making the customer repeat details.
Create an escalation handoff for [CUSTOMER ISSUE]. The customer said: [CUSTOMER MESSAGE]. Verified history: [CASE HISTORY]. Route it to [TEAM OR ROLE] under this rule: [ESCALATION POLICY]. Produce two parts: a customer-facing message under [WORD LIMIT] words that explains the handoff and next update, and an internal case brief with timeline, evidence, requested resolution, risk flags, and unanswered questions. Do not state a resolution is approved unless the case history confirms it.
Troubleshooting chat
Use when support needs a short diagnostic chat flow for a technical problem.
Design a live-chat troubleshooting script for [PROBLEM] in [PRODUCT]. The customer profile and device context are [KNOWN CONTEXT]. Approved troubleshooting steps are [APPROVED STEPS]. Ask one diagnostic question at a time, begin with the least disruptive check, and explain why each requested step is needed in plain language. Include a stop condition for escalation to [ESCALATION TEAM]. Output agent messages, likely customer answers, and the next branch. Do not ask for passwords, full payment details, or unnecessary personal data.
Tips
- Make approved policies machine-readable: include refund windows, monetary caps, eligibility tests, and who can authorize exceptions.
- Separate a customer reply from the internal case note; the latter can preserve diagnostic detail and escalation signals without confusing the customer.
- For delay updates, require a confirmed status, the next checkpoint, and a date only when it is actually known.
- Sample AI drafts for accidental admissions, invented case details, and promises that exceed the agent’s authority before deploying a template.
FAQ
Can AI answer customers without an agent?
Only for tightly bounded, low-risk cases with current policy and a clear human handoff. Sensitive account, payment, safety, privacy, and exception cases need controlled escalation.
How can I stop AI from making up order information?
Give it only verified fields and explicitly require it to say what must be checked when a field is absent. Do not connect it to broad account data without access controls.
Should every reply mention policy?
No. Use plain policy explanations when they affect the decision. A routine helpful answer should not sound like a terms-and-conditions notice.