Support teams do not lose their day to hard problems. They lose it to the twentieth version of the same reply, the angry message that needs careful wording, and the bug report that has to be translated into something engineering will act on.
These Claude prompts for customer service target exactly that work. Below are 25 prompts covering first replies, de escalation, refund and policy explanations, bug summaries, macros, escalation handoffs, and follow up after a resolution.
On the numbers you will not find here. This guide does not quote resolution time improvements, because the figures circulating for AI in support are mostly vendor marketing without a verifiable method. Measure your own first reply time and handle time before and after instead.
Hand your tier 1 tickets to an AI agent.
The tickets that eat a support team’s day
Ticket volume is not evenly distributed by difficulty. A small set of shapes consumes most of the hours.
Four categories dominate:
- The repeated question. Answered a thousand times, still needs a personalized reply.
- The angry message. Genuinely needs care, and drains the agent who handles it.
- The vague bug report. “It doesn’t work.” Someone has to extract reproducible steps.
- The policy explanation. Refunds and exceptions, where wording carries real risk.
Notice what these have in common. None of them require deep product knowledge, but all of them require careful writing. That is precisely the gap Claude prompts for customer service fill.
Why Claude suits support work
Three properties make Claude a good fit here, and it is worth being specific rather than generic.
It holds long threads. A ticket that has been running three weeks with six replies is a lot of context. Because Claude keeps long conversations coherent, you can paste the whole thread and ask for a summary that does not lose the thread’s history.
Its default tone is measured. Support writing fails when it is either robotic or overly effusive. Claude tends toward careful and plain, which is closer to the right register for an apology or a policy explanation.
It follows negative instructions well. “Do not apologize more than once” and “do not promise a timeline” are the kind of constraints support writing depends on, and Claude holds them reliably.
If your team already uses a different model, our ChatGPT prompts for customer service cover the same ground.
25 Claude prompts for customer service
Every one of these Claude prompts for customer service follows the same pattern. Replace the bracketed fields with your details. In addition, append your tone paragraph from the section below to every prompt.
First reply drafting (prompts 1 to 4)
1. “You are drafting a support reply for [PRODUCT]. Here is the customer’s message: [PASTE]. Write a first reply that acknowledges the specific issue, states what I will do next, and gives a realistic next update. Do not promise a resolution time.”
2. “Rewrite this reply so it leads with the answer instead of the explanation: [PASTE REPLY]”
3. “The customer asked three questions in one message. Draft a reply that answers each one clearly, using their own wording as short headers. Message: [PASTE]”
4. “Write a holding reply for a ticket I cannot resolve today. Be honest that it is unresolved, say what I have done so far, and commit to a specific next update. Under 100 words.”
Angry customer de escalation (prompts 5 to 9)
5. “This customer is angry and partly right. Draft a reply that acknowledges the specific failure without excessive apology, states what we are doing, and does not get defensive. Message: [PASTE]”
6. “This customer is angry and mistaken about what happened. Draft a reply that corrects the record gently, without making them feel stupid. Message: [PASTE]”
7. “Rewrite this reply to remove defensiveness while keeping every fact the same: [PASTE REPLY]”
8. “The customer is threatening to leave a public review. Draft a reply that takes the underlying issue seriously and does not reference the threat at all.”
9. “Draft a reply for a customer who has had three separate problems this month. Acknowledge the pattern honestly rather than treating this as an isolated incident.”
Refund and policy explanations (prompts 10 to 13)
10. “Explain this policy to a customer in plain language, without quoting it verbatim and without sounding like a legal document. Policy: [PASTE]”
11. “Draft a refund approval message that is warm and simple, confirms the amount and timing, and does not upsell.”
12. “Draft a refund denial that explains the actual reason, avoids hiding behind ‘policy’, and offers whatever alternative exists. Situation: [PASTE]”
13. “The customer is asking for an exception we can make but rarely do. Draft a reply that grants it, is clear it is an exception, and does not set an expectation for next time.”
Bug report summarizing (prompts 14 to 17)
14. “Turn this customer message into a bug report for engineering with: expected behaviour, actual behaviour, reproduction steps, environment, and severity. Flag anything the customer did not tell us. Message: [PASTE]”
15. “Read these five tickets and tell me whether they describe the same underlying issue. If so, write one consolidated report. Tickets: [PASTE]”
16. “Draft the follow up questions I need to ask this customer to make their report reproducible. Maximum four questions, in plain language. Report: [PASTE]”
17. “Translate this engineering update into something the customer will understand, keeping every commitment accurate. Update: [PASTE]”
Macro and help centre writing (prompts 18 to 21)
18. “I answer this question about twice a day. Write a macro that covers the common case and flags where an agent needs to personalize it. Question: [PASTE]”
19. “Review this macro for anything that reads as robotic or condescending, and rewrite those parts only: [PASTE MACRO]”
20. “Turn this resolved ticket thread into a help centre article with a clear title, the problem, the solution in steps, and a note on when it does not apply. Thread: [PASTE]”
21. “Write five help centre article titles for this topic, phrased the way a frustrated customer would search rather than how we describe it internally. Topic: [TOPIC]”
Escalation summaries (prompts 22 to 23)
22. “Summarize this ticket thread for a manager taking it over: what the customer wants, what we have done, where it stands, what the customer has been promised, and the sentiment. Thread: [PASTE]”
23. “Draft an internal note for engineering explaining why this ticket is being escalated, including business impact and what the customer has already been told.”
Follow up after resolution (prompts 24 to 25)
24. “Write a short follow up message three days after resolving a difficult ticket. Check the fix held, do not ask for a rating, do not upsell. Under 60 words.”
25. “Draft a message to a customer who gave us a low satisfaction score after a resolved ticket. Ask what we got wrong, genuinely, with no defensiveness and no attempt to change the score.”
Try these on your own ticket queue.
Giving Claude your tone of voice in one paragraph
All of these Claude prompts for customer service improve sharply with one addition, and it takes ten minutes to write once.
Draft a single paragraph describing how your team writes, then append it to every prompt. Be concrete rather than aspirational:
“Write like a support agent at [COMPANY]. Plain language, short sentences, no corporate softening. Apologize once at most and only when we are actually at fault. Never say ‘we understand your frustration’. State what happens next with a specific time. Sign off simply. Match this example: [PASTE ONE OF YOUR REAL REPLIES].”
The example at the end does the heaviest lifting. Because a real reply carries register, rhythm, and vocabulary that no description captures, one pasted example beats three paragraphs of instruction.
Then store it where agents will actually use it: a pinned note, a snippet, or the top of your prompt library.
Turning your best replies into reusable prompt templates
The highest value use of these Claude prompts for customer service is not one off drafting. Instead, it is turning what your best agent already does into something everyone can use.
Work through it in four steps:
- Find the top 20 repeated tickets. Your help desk reporting already knows these.
- Pull your best agent’s actual replies for each one, not the macro.
- Run prompt 18 on each, feeding in that reply as the example.
- Have the agent edit the output. They will catch the two sentences that are subtly wrong.
The result is a library that sounds like your best person rather than like a template. As a result, quality stops depending on who picked up the ticket.
Review it quarterly. Because product changes silently invalidate macros, an unreviewed library slowly starts lying to customers.
Handing tier 1 tickets to an AI agent entirely
Here is where prompts stop. Every prompt above still requires an agent to open the ticket, paste it, and send the reply. That is faster than writing from scratch, but a person is still in every loop.
For the genuinely repetitive tier 1 volume, that person does not need to be. Dapta builds AI agents that handle those conversations end to end across voice and messaging.
What an AI agent should and should not own
- Should own: order status, password resets, hours and policy questions, appointment changes, plan and billing lookups
- Should own: the whole conversation in English or Spanish, with natural regional accents
- Should own: writing the interaction back to your help desk so the record is complete
- Should escalate: anything where the customer is upset, anything involving money owed, anything ambiguous
- Should never own: refund exceptions, complaints about a serious failure, anything legal
- Should always identify itself as an AI agent and offer a human on request
The division is not about capability, it is about judgment. Because a customer who is already upset gets more upset when an automated system handles them badly, the escalation rule matters more than the automation itself.
For the full picture, explore our related guides:
- ChatGPT prompts for customer service, the same tasks with a different model
- Claude prompts for sales, the parent guide this list builds on
- AI customer service scripts, full call structures for voice support
- Best AI chatbot for customer service, a platform comparison
- AI customer support system, the operational overview
- WhatsApp AI agent for support, for teams supporting on WhatsApp
- More AI prompt libraries, the full collection
Frequently asked questions
Will customers notice replies were AI drafted?
Not if an agent edits before sending, and especially not if you supply a real tone example. What customers do notice is generic phrasing like “we understand your frustration”, which is why the tone paragraph explicitly bans it.
Is it safe to paste customer messages into Claude?
Redact first. Remove names, order numbers, addresses, and payment details, since none of them improve the draft. In addition, use a business tier account where conversations are excluded from training and retention is controlled.
Should agents send AI drafts without reading them?
No. These Claude prompts for customer service produce a strong first draft, not a final reply. Because the model cannot know what your product actually did last Tuesday, a human check catches confident errors.
Can Claude write our whole help centre?
It can draft it from resolved tickets, which is a much better source than writing from scratch. However, someone with product knowledge has to verify each article, since an inaccurate help article generates more tickets than it deflects.
Does this replace support agents?
No. It removes drafting time so agents spend more of the day on the tickets that genuinely need a person. The volume that gets fully automated is the repetitive tier 1 tier, and even that needs a working escalation path.
Where should we start?
Start with your five highest volume repeated tickets and prompt 18. Because those are predictable and low risk, you get the clearest return before touching anything sensitive.
Stop rewriting the same reply twenty times a day. , no card required.
