Claude Prompts for Customer Support Replies

The short version: Claude writes better support replies than most tools because you can hand it your tone, your policy and the full email thread in one go, but the prompt has to include constraints, not just a request. Give it a bad customer example, your refund rules and a word limit, and it stops sounding like a corporate robot. Skip any of those three and you'll be rewriting half of what it gives you.

Why Claude specifically, and not just "AI" in general

I've written before about how Claude AI can handle customer service enquiries for a small business, and the short reason it's worth bothering with is context length. A lot of support tickets aren't one message, they're a back-and-forth thread with a customer who's already annoyed, already repeated themselves twice, and already threatening to leave a review. Claude handles long pasted threads well and tends not to lose track of what was already said three messages ago. That matters more than people think, because a reply that ignores something the customer already told you is the fastest way to make them angrier.

That said, Claude is not magic. It will write a plausible, grammatically perfect, completely wrong answer about your refund policy if you don't tell it the policy. It doesn't know your business. It knows how support emails are generally structured, and it will fill gaps with generic nonsense unless you stop it.

The structure every good support prompt needs

I use five parts, every time, no exceptions:

• Role - tell it who it's writing as (a support agent for a specific type of business, not "a helpful assistant")
• The actual customer message - pasted in full, including the angry bits, not summarised
• Your policy or facts - refund windows, shipping times, what you can and can't offer
• Tone instruction - specific, not "friendly" (more on this below)
• A length limit - without this Claude pads replies with filler sentences nobody reads

Miss the length limit and you'll get a reply that opens with "Thank you so much for reaching out" and closes with three lines about valuing the customer relationship. Nobody wants to read that. Customers want the answer and then the sign-off.

A worked example

Say you run an online homeware shop and a customer emails: "I ordered a lamp 11 days ago, it still hasn't arrived, and your tracking page says 'label created' with no update. This is the second time this has happened with your company and I want a refund, not another excuse."

A weak prompt would be: "Write a reply to this angry customer." Claude will produce something apologetic but vague, and it won't offer a refund because it doesn't know if you allow one.

A working prompt looks like this:

• "You are a support agent for a homeware shop. Here is the customer's message: [paste in full]. Our policy: orders not delivered within 10 working days qualify for a full refund or a free reshipment, customer's choice. We do not currently offer express shipping so avoid promising faster delivery next time. Write a reply that acknowledges this is their second bad experience, offers the refund or reshipment choice clearly in the first two sentences, and does not use the phrase 'we apologise for any inconvenience.' Keep it under 120 words."

That prompt gives Claude a fact (the 10-day rule), a constraint (no empty apology phrases), a structural instruction (choice up front, not buried), and a hard limit. The output is a reply you could send with minor edits, not a draft you have to rebuild.

Prompts for the situations that come up

These are the five I'd keep saved somewhere, because they cover most of what a small support inbox deals with.

1. The repeat complainer

"Write a reply to a customer who has contacted us three times about the same unresolved issue [describe it]. Acknowledge the repetition directly, do not use the word 'understand' more than once, and end with one specific next step and a timeframe, not a vague promise to look into it."

2. The policy no

"The customer is asking for [specific thing] which is outside our policy because [reason]. Write a reply that says no clearly in the first sentence, explains the reason in one sentence, and offers one alternative if one exists. Do not soften the no with multiple hedging phrases."

3. The five-star-turned-one-star

"This customer left a five-star review last month and is now furious about [issue]. Write a reply that references their past loyalty without sounding manipulative, and keep it under 80 words."

4. The confused first-timer

"This customer is new and confused about [process]. Write a reply that explains it in three short steps, uses no jargon, and assumes zero prior knowledge of how our system works."

5. The template for everything else

"Draft three alternative openings for a reply to this message, one direct and brief, one warm and brief, one that leads with the solution before the apology. I'll pick the one that fits."

That last one is useful because it stops you accepting the first draft Claude gives you. Three openings takes ten seconds longer and usually one of them is better than your first instinct.

On tone words, and why "friendly" is useless

Every guide on this topic tells you to specify tone. Almost none of them tell you that "friendly," "professional" and "empathetic" are nearly meaningless to a language model because they've been trained on millions of emails that all claim to be those three things while sounding identical and hollow. Instead, give Claude a reference. Paste in a past reply you liked and say "match this tone." Or describe tone by what to avoid: "no exclamation marks, no corporate filler phrases, write like a person replying to a friend who's annoyed at them." That produces something distinct. "Be friendly and professional" produces the same beige email every company already sends.

Here's the uncomfortable part most posts on this topic skip over: AI-drafted replies, even good ones, read slightly off to customers who deal with your business regularly, and the fix isn't a better prompt, it's a human skim before sending. Not a full rewrite, just thirty seconds checking the specifics are right and the tone matches how you'd talk to that person. Treat Claude as the first draft writer, not the final sender, and the whole approach works a lot better than if you're hoping to fully automate it from day one.

A simple workflow if you're doing this daily

If you're handling more than a handful of tickets a day, here's a rough process that keeps quality up without eating your morning:

• Keep a document with your core policies (refunds, shipping times, cancellation windows) and paste the relevant bit into every prompt rather than retyping from memory
• Save your five best prompts somewhere you can copy from fast, rather than rewriting instructions each time
• Batch similar tickets together, write the prompt once with the pattern, adjust the specific detail for each customer
• Read every draft before sending, even when you're confident, because the one time you don't is the time it invents a policy you don't have

If your team is bigger than just you, this is also where a shared setup helps, and I've written separately about the Claude Team plan and a seven-day rollout if you want multiple people working from the same saved prompts and policy documents rather than everyone improvising their own version.

What this doesn't replace

None of this replaces having an actual policy to begin with. If your refund rules live in someone's head rather than written down, Claude can't help you, because it'll either guess or make something up that sounds confident and is wrong. Get the policy written first, even three bullet points, then build prompts around it. The same logic applies to broader email work, not just support tickets, and if you want the wider version of this I've covered it in how to use AI to answer customer emails without it reading like a robot wrote it, plus there's crossover with how email marketing supports lead nurturing if your support replies are also doing a bit of relationship-building, which most of them should be.

If you want a faster, cheaper model for high-volume simple replies and save Claude's fuller version for the complicated ones, it's worth reading up on when to pick Claude Haiku 5.5, since not every ticket needs the heaviest model to get a decent answer, and running everything through the most expensive option is a quiet waste of money most small teams don't notice until the bill arrives.

Frequently asked questions

Can Claude write my entire customer support inbox for me?

It can draft most of it, but someone should skim every reply before it's sent, because Claude will occasionally invent a policy detail or miss a nuance in a thread, and that one mistake can cost you more than the time you saved.

What's the single most important thing to include in a Claude support prompt?

Your actual policy or facts, pasted in rather than assumed. Without them Claude fills the gap with a generic, plausible-sounding answer that may be entirely wrong for your business.

How long should a Claude-drafted support reply be?

Set a hard word limit in the prompt, somewhere around 80 to 150 words for most tickets. Without a limit, Claude pads replies with filler apologies and closing lines nobody wants to read.

Is it obvious to customers when a reply was written by AI?

It can be, if the tone instruction was vague. Replies built on words like "friendly" or "professional" tend to sound generic. Giving Claude a real example of tone to match produces something far less obviously templated.