I thought I'd cracked eBay customer service when I set up a "polite professional template" in November. One prompt, every situation: late delivery, return request, VAT question, whatever. Copy the buyer's message, paste it into Claude, get a reply, send it. I was clearing fifty messages a day in under two hours, and my seller rating sat at 4.9.
Then Boxing Day hit. Returns flooded in—jumpers that didn't fit, toys the kids already owned, a coffee machine someone bought twice by accident. Royal Mail ground to a halt under the backlog. Buyers who'd ordered on 18 December were messaging on 3 January asking where their parcel was, and my one-size reply ("We apologize for the delay and appreciate your patience") earned me three neutral ratings and a warning from eBay that my response quality had dropped.
I sell vintage clothing and homeware—Pyrex bowls, 1980s cardigans, enamel bread bins, the sort of thing people impulse-buy at 11 p.m. then forget they ordered. I'm a sole trader registered with Companies House, I list about four hundred items a month, and I handle every message myself. I realized in early January that my generic prompt was costing me sales and ratings because it treated every situation the same: apologetic, vague, eager to please. A buyer threatening a PayPal dispute got the same tone as someone asking a simple VAT question.
I tested four different prompt formulas across two hundred real messages in January and February. Not four variations of one template—four structurally different approaches. I logged which formula I used for each message type, tracked the outcomes, and watched my rating climb back to 4.8 by mid-February. Here's what worked, what didn't, and the exact prompts I'm still using.
The four formulas I tested, and why I needed more than one
The problem-solution formula is the one I started with: acknowledge the issue, explain what happened, offer a fix. It works when there's a clear problem and you control the fix—a missing item, a damaged parcel, a wrong size. The structure is "I understand [problem]. Here's what happened: [reason]. Here's what I'll do: [action]."
The AIDA formula—Attention, Interest, Desire, Action—sounds like marketing copy, and it is. I used it for messages where the buyer wasn't upset but needed persuading: a question about whether an item would fit, a request for more photos, someone asking if I'd accept a lower offer. It front-loads the benefit, builds interest, then gives a clear next step.
The feature-benefit formula translates a product detail into a reason the buyer cares. I used it almost exclusively for pre-sale questions: "Is this bowl microwave-safe?" or "What's the fabric content?" It's a two-sentence structure: the fact, then the benefit.
The empathy-first formula starts with validation before it gives information. I used it for emotionally charged messages—late Christmas presents, a gift that arrived broken, a return request on an item the buyer clearly loved but couldn't afford to keep. It's "I completely understand [feeling]. [Relevant fact]. [What I can do]."
I didn't invent these; they're standard copywriting frameworks. But I'd never thought to match the framework to the message type, and that mismatch was killing my replies.
Late Royal Mail deliveries: problem-solution won, empathy-first backfired
Between 27 December and 10 January I had sixty-three messages asking where a parcel was. Royal Mail's tracking showed "in transit" for eight days on some orders. My original generic prompt produced replies like "We sincerely apologize for the delay and understand your frustration. Royal Mail is experiencing high volumes. We're confident your order will arrive soon."
Confident based on what? The buyer didn't want confidence; they wanted a plan.
I switched to problem-solution and gave Claude this prompt:
"You're a UK eBay seller. A buyer has messaged asking where their order is. Royal Mail tracking shows [status]. Write a reply that: 1) acknowledges the delay without over-apologizing, 2) explains what the tracking means in plain English, 3) offers one concrete action (refund, replacement, or wait with a deadline). Tone: helpful, not defensive. 150 words max."
The output: "Your order was dispatched on [date] and is currently in transit with Royal Mail. Tracking shows it's moving through their network, which is running 4–6 days slower than usual after the Christmas backlog. If it hasn't arrived by [date five days out], message me and I'll send a replacement or issue a full refund—your choice. I'll keep an eye on the tracking in the meantime."
Neutral ratings stopped. I had two buyers reply "Thanks, that's really helpful," which had never happened with the generic version.
I tried empathy-first on ten of these messages—"I completely understand how frustrating it is to wait for a parcel, especially after Christmas"—and three buyers escalated. They didn't want understanding; they wanted their item or their money back. Empathy without action reads as stalling.
VAT questions: feature-benefit was faster and clearer than problem-solution
I had nineteen messages asking about VAT. Most were from buyers in Northern Ireland or the Republic wondering if they'd pay extra fees, or businesses asking for a VAT invoice. My generic prompt produced replies that started with "Thank you for your question" and ended with "Please let me know if you need further clarification," but buried the actual answer in the middle.
Feature-benefit cut that to two sentences:
"You're a UK eBay seller. A buyer has asked [VAT question]. Write a reply that: 1) answers yes or no in the first sentence, 2) explains the benefit or implication in the second sentence. No filler. 50 words max."
Example output for "Do I pay VAT on this?": "No, this item is second-hand and sold by a sole trader, so there's no VAT to pay. The price you see is the full price—nothing extra at checkout."
For a business asking for a VAT invoice: "I'm not VAT-registered, so I can't issue a VAT invoice. I can send a standard receipt with my Companies House registration number if that helps for your records."
I tested problem-solution on five of these and the replies were twice as long for no benefit. VAT questions don't have a "problem" to solve; they're factual queries that need a factual answer.
Boxing Day returns: empathy-first kept the rating, problem-solution felt robotic
Forty-one return requests between 26 December and 15 January. About half were straightforward—wrong size, changed mind, duplicate gift. The other half had an emotional edge: "This was supposed to be a Christmas present and it arrived on 28 December," or "The bowl arrived chipped and now I look like an idiot in front of my mother-in-law."
Problem-solution handled the straightforward ones fine:
"You're a UK eBay seller. A buyer wants to return [item] because [reason]. Write a reply that: 1) confirms you accept the return, 2) explains the return process in three steps, 3) mentions the refund timeline. Tone: efficient, friendly. 100 words max."
Output: "No problem—I accept returns within 30 days. Here's how it works: 1) Pop the item back in the post to [address], 2) send me the tracking number when you've posted it, 3) I'll refund you within 24 hours of receiving it, usually 3–5 days total. Let me know if you need the return address again."
But the emotionally charged ones needed empathy-first or the buyer left a neutral rating even after I processed the return:
"You're a UK eBay seller. A buyer is upset because [situation]. Write a reply that: 1) validates their frustration in one sentence, 2) states the relevant fact (late delivery, damaged item, etc.), 3) offers a fix with no hoops to jump through. Tone: warm, no corporate jargon. 100 words max."
Output for the late Christmas present: "I'm really sorry this didn't arrive in time—I know how disappointing that is when you've planned a gift. Royal Mail's Christmas backlog hit harder than expected this year. I'd like to offer you a full refund, no need to return it. I'll process that today. If you'd rather keep it, let me know and I'll refund £10 instead as an apology."
Three buyers replied thanking me. One left a positive review specifically mentioning "excellent customer service on a late delivery." That never happened with problem-solution, which read as going through the motions.
Dispute threats: AIDA was a disaster, problem-solution with boundaries worked
I had eleven messages threatening to open a PayPal dispute or leave negative feedback. Most were impatient buyers whose parcels were in transit, two were return requests where the buyer wanted an instant refund before posting the item back, one was someone claiming an item was fake (it wasn't—I have receipts).
I tried AIDA on the first three because I thought I could "sell" them on waiting. Terrible idea. AIDA assumes the reader is open to persuasion. Someone threatening a dispute is past that point.
Problem-solution with boundaries worked, but I had to add a new instruction:
"You're a UK eBay seller. A buyer has threatened to [dispute/negative feedback] because [reason]. Write a reply that: 1) acknowledges their frustration, 2) states the factual situation (tracking status, return policy, etc.), 3) offers one fair solution, 4) calmly states what you can't do and why. Tone: firm but polite, no defensiveness. 150 words max."
Output for a buyer demanding an instant refund before returning a jumper: "I understand you want this resolved quickly. I'm happy to refund you in full once I've received the return—that's eBay's standard process and it protects both of us. I can't issue a refund before the item is posted back, but I'll process it within two hours of it arriving, usually 3–4 days total. If you post it today and send me the tracking number, I'll keep an eye out for it. Let me know if you need the return address again."
The buyer posted it the next day. No dispute, no negative feedback.
The one I got wrong: I used empathy-first on a buyer claiming an item was fake, thinking validation would de-escalate. It didn't—it made me sound guilty. Problem-solution with evidence (photos of the original label, purchase receipt) closed it in two messages.
What I'd do differently if I started over
I'd skip the "one template for everything" phase entirely and start with these four formulas from day one. I'd also log every message type for the first fifty messages so I could see the pattern—late deliveries cluster between 26 December and 10 January, VAT questions spike in January when businesses are sorting receipts, return requests are constant but the tone changes depending on the reason.
I wouldn't bother with AIDA for customer service. It works for pre-sale questions where you're genuinely trying to interest someone in buying, but it's too salesy for problem resolution.
I'd write the prompts in a Google Doc with the formula name at the top and the exact instruction below it, then keep that doc open in a second tab. Switching between four prompts sounds fiddly, but it's faster than redrafting a generic reply that doesn't land.
And I'd test the output on ten real messages before committing to a formula. I assumed empathy-first would work for late deliveries because it "sounded nicer," but buyers don't want nice—they want a plan. I only figured that out after twenty messages and three neutral ratings.
Questions people ask
Can I use the same prompt for eBay and Shopify customer service?
Mostly, yes. The formulas work across platforms, but you'll need to adjust the specifics—eBay buyers expect you to follow eBay's return policy, Shopify customers expect you to follow whatever's on your website. The structure (problem-solution, empathy-first, etc.) stays the same.
How do I know which formula to use if a message mixes two situations?
Go with the emotional temperature. If a buyer asks a factual question but sounds frustrated, use empathy-first. If they're calm but the situation is complicated, use problem-solution. The formula matches the tone you need, not the category of question.
Do I need to rewrite these prompts for every message?
No—I fill in the bracketed bits (tracking status, item name, return reason) and leave the instruction the same. The only time I rewrite is if the output misses the mark twice in a row, which happens maybe once a week.
Will this work if I'm not using Claude?
I've tested these prompts in Gemini Flash and Llama 3.3 70B and they work fine. The output's a bit more formal in Gemini, a bit more casual in Llama, but the structure holds. I'd avoid the very small models (GPT-OSS 20B, Amazon Nova Micro) for anything where tone matters—they're too hit-and-miss.
I still spend two hours a day on messages, but I'm clearing eighty instead of fifty and my rating's stable at 4.8. If you're handling eBay customer service at volume and your replies all sound the same, try splitting your prompts by situation. You can test this in Kryotta's compare mode if you want to see how different models handle the same formula—I did that for the first twenty messages and it saved me a week of trial and error.
