Create a copy-paste prompt pack for a customer support agent handling [PRODUCT OR SERVICE] under these policies: [POLICY AND APPROVAL RULES]. Our voice is [BRAND VOICE], and the support channels are [CHANNELS]. Use only facts supplied in the prompt; never invent an order status, outage cause, refund eligibility, account action, legal commitment, or promised timeline. Provide six numbered prompts, each 90–140 words, with placeholders an agent can fill in: 1. First response to a frustrated customer. 2. Troubleshooting response that gathers the minimum useful diagnostic details. 3. Delay, bug, or outage update that distinguishes known facts from investigation.
4. Refund, replacement, or exception request that follows stated policy and identifies approval needs. 5. Escalation brief for engineering, billing, or a manager. 6. Final resolution and follow-up response. Each prompt must instruct the AI to produce a customer-facing draft plus a short internal note listing facts, actions taken, owner, and next step. Require empathy without apologizing for facts not established. Require direct answers before explanations, plain language, and a clear next action or time-bound checkpoint. For account security, payment, privacy, safety, legal, or harassment issues, require the AI to pause normal handling and route the case through the stated process. Before answering, check that every prompt separates customer statements from verified facts and contains no unsupported promise. Ask up to 3 clarifying questions only if a required input is missing.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [PRODUCT OR SERVICE] | Describe what you sell and the common support context. | Cloud-based team scheduling software for small retail businesses |
| [POLICY AND APPROVAL RULES] | Paste the relevant refund, replacement, privacy, escalation, and approval policies. | Refunds within 30 days for annual plans need billing approval; never disclose account details until ownership is verified; P1 outages go to incident commander |
| [BRAND VOICE] | Describe the approved tone and any words or style rules to avoid. | Warm, direct, professional; no slang or blame |
| [CHANNELS] | List the channels where agents respond, such as email, chat, or social messages. | Email, in-app chat, and Help Scout tickets |
How to use
- Paste your live policies first; do not rely on the model to know what your company permits.
- Choose the prompt matching the case and replace its placeholders with verified ticket facts.
- Check any refund, credit, security, privacy, or timeline statement against your systems and policy.
- Follow up with: “Rewrite this reply in 90 words, keeping only claims supported by the ticket notes.”
Variations
QA scorecard
Use this to audit a sample of support conversations.
Audit these [NUMBER] support conversations against [SUPPORT POLICIES] and [QUALITY STANDARDS]: [CONVERSATIONS]. Score each on issue understanding, factual accuracy, policy compliance, empathy, ownership, next step, and writing clarity. Return a table with evidence from the conversation, a 1–5 score per criterion, and one coaching action. Separate agent mistakes from missing documentation or product defects. Do not infer facts absent from the records. Ask up to 3 questions if a scoring rule is unclear.
Knowledge article
Use this to turn repeated tickets into a self-service article.
Write a help-center article for customers who need to solve [PROBLEM]. Use these verified product details and steps: [SOURCE MATERIAL]. Audience: [AUDIENCE]. Return a title, short symptom description, prerequisites, numbered steps, expected result, troubleshooting branch for each likely failure, and when to contact support. Mark anything that needs product-team verification rather than filling gaps. Do not include internal tools, security controls, or unsupported workarounds. Ask up to 3 questions only if required.
Escalation summary
Use this when handing a complex case to another team.
Create an escalation brief for [DESTINATION TEAM] about this customer issue: [TICKET HISTORY]. Include customer impact, affected account or order identifiers, reproducible steps, expected versus actual behavior, timestamps and timezone, diagnostics already collected, actions already attempted, severity rationale, requested decision, and customer communication status. Separate verified evidence from the customer’s report. Exclude unnecessary personal data. Flag missing evidence that would change priority. Ask up to 3 questions only if necessary.
Tips
- Write policy constraints into the prompt, especially exceptions and who can approve them; tone guidance cannot replace policy.
- Ask for the customer-facing response and internal note separately so private operational detail does not leak into the reply.
- For outages, state the next update checkpoint only if your incident process supports it; do not convert an estimate into a promise.
- Use ticket IDs and exact timestamps in escalations so the receiving team can reconstruct the case.
FAQ
Can AI send customer support replies automatically?
Only after your team has defined approval, data-access, and escalation controls. High-risk cases such as payment, privacy, safety, and legal issues should have human review.
How can I stop it from promising refunds?
Paste the refund policy and tell it to label approval requirements instead of offering outcomes beyond the agent’s authority.
Should I include customer data in the prompt?
Use the minimum necessary data and follow your company’s privacy and vendor rules. Remove sensitive details when they are not needed to draft the response.