A customer service reply prompt is hardest not in the wording but in the balance: apologise too much and the customer feels brushed off; explain too much and it sounds like you are dodging responsibility. The prompt below hands over the customer's own words, the facts you have verified, the remedies you are actually authorised to offer and the reply deadline in one go. The draft comes back in a fixed order — acknowledge the problem, offer a remedy, end with one clear next step — so you can check it once and send it, instead of rewriting the tone line by line.
Ask an AI casually to "help me reply to this complaint" and you usually get one of two failures. Either the whole reply is "we are very sorry for the inconvenience" with no actual solution in sight, or it promises a full refund plus extra compensation that turns out to exceed your authority when it is time to deliver. The difference is whether the boundaries are written down first: how much you can refund, how far you can go, and when the customer hears back. Put that in the prompt and the model can only phrase things inside those limits.
The prompt (copy and use)
You are an experienced customer service lead. Draft a reply I can send to the customer as-is, following these six rules:
1. Acknowledge the problem first: restate the customer's concrete problem and loss in one short sentence so they know they were heard. Do not repeat their emotional wording, and do not open with explanations.
2. Apologise once: apologise for the specific problem only. No repeated empty phrases, and never shift the blame to the warehouse, the courier or the system.
3. Only offer what can be delivered: choose solely from the remedies I list below. For each one, state what the customer needs to do, what we will do, and how long it takes. Do not promise any compensation outside that list, not a word of it.
4. Give concrete times: wherever waiting is involved, name an exact point in time, such as "we will have a result for you by 6 p.m. tomorrow". Never write "as soon as possible" or "shortly".
5. End with one step: the last sentence leaves exactly one clear next step — the customer confirms the remedy, provides the order number, or waits for our follow-up call. Never offer three paths at once.
6. Length and tone: keep the whole reply under [length limit, default 200 words]. Calm, specific tone; no grovelling, no lecturing. If the customer writes angrily, respond only to the factual parts.
The customer's message, word for word:
[paste the complaint exactly as received]
Order and verified facts: [what you have confirmed, e.g. the parcel shows as delivered, the customer never received it, the courier depot has confirmed it is investigating]
Remedies I am authorised to offer: [e.g. a full refund, or a free replacement with shipping on us; vouchers up to 30 at most; anything beyond this is above my authority]How to swap the variables
| Variable | What to put in | Example |
|---|---|---|
| [customer's message] | The original text, not your summary of it | Their "waited ten days" aims the reply far better than your paraphrase "slow logistics" |
| [order and verified facts] | Only the status you have actually checked | Tracking records, call logs, warehouse dispatch times; leave guesses blank rather than inventing them |
| [remedies] | Options and ceilings you can actually deliver | The ceiling is the point: full refund or difference only, maximum voucher value — put numbers on it |
| [length limit] | Set it by channel | Private chat can run to 300 words; a public review reply should stay under 120, with the problem acknowledged and one contact path |
A 30-second check before you send
- Facts: compare every order status, amount and time in the reply against your records, one by one. If the model miscopies a single digit, the customer will screenshot it and ask.
- Promises: if lines like "extra compensation" or "guaranteed never to happen again" appear and you cannot deliver them, delete them or rewrite them into something inside your authority.
- Deadlines: count weekends and holidays into any time you name. If you wrote "by 6 p.m. tomorrow", something must arrive by then, even if it is only a progress update.
- Channel fit: a public review reply must not contain the customer's name, phone number or full order number. In private chat you may use their name, but never compare their case with another customer's outcome.
Complaints you should not send an AI draft for directly
Anything involving personal injury, a product causing harm, or spoiled food goes to a supervisor first; the reply states only that the case is accepted and when you will call back, with any position waiting for the investigation. If the customer has already mentioned a regulator, a lawyer or the press, the reply becomes evidence: state facts and actions already taken only, and have a supervisor read it before sending. Amounts above your authority, or requests to delete data or close an account, stay drafts until the internal process is done.
There is one more case: the same problem arriving in batches. Do not draft replies one by one yet. Gather the reviews and complaints, sort and count them first, and find whether it was one batch of goods or one leg of shipping that failed. The method in the user feedback prompt that sorts hundreds of reviews in one pass does that sorting; then come back to the prompt above for individual replies. Doing it in the reverse order only multiplies the replies.
Fill in the facts and your authority properly and this prompt carries across e-commerce after-sales, software support and replies to negative reviews for physical stores. Switching industries means editing only the remedies line; the rest of the structure stays as it is.
