E-Commerce and Customer Service Prompts: Product Pages and Support

A customer writes that their lamp arrived broken. You paste their message into a chatbot, ask for a polite reply, and the AI promises them a refund within 48 hours and a store credit. The text is flawless. The problem is that your shop has never offered either one.

In e-commerce, the risk of a poorly framed prompt has less to do with style than with the invented promise, the one your customer will quote back to you in writing. Everything below therefore starts from the same principle: the AI writes from your rules and your product data, and it's allowed to say it doesn't know.

The rules sheet, to prepare before anything else

Write down once and for all what your customer service can say. One page is enough: real shipping times, carriers, return conditions, who pays for the return, what you do in case of a damaged parcel, the point at which a human must take over.

Two points deserve serious checking, because they don't depend only on you. In France, a consumer who buys online generally has 14 days to withdraw, with exceptions (custom-made products, for example), and usually benefits from a two-year legal guarantee of conformity. Your sheet must reflect these rules as they apply to your products: rely on your terms of sale (CGV), on the DGCCRF website or on a lawyer if you have any doubt. The AI, for its part, can get the law wrong.

Once written, this sheet can live in your tool's custom instructions, so you no longer have to paste it in.

Replying to a complaint without promising just anything

Here's what the weak prompt gives you:

Reply politely to this unhappy customer: [MESSAGE]

And its corrected version:

You're writing a reply for the customer service of [STORE].

Store rules (only authorized source):
[PASTE THE RULES SHEET]

Customer's message:
[MESSAGE, WITHOUT NAME, ADDRESS OR PHONE NUMBER]

Order data:
[ORDER DATE, DELIVERY STATUS, PRODUCT]

Write a reply of 120 words maximum:
- start by acknowledging the customer's specific problem, without a stock phrase
- say what will happen next and what the customer needs to do
- keep a polite, formal register, stay simple, no repeated apologies

Prohibitions: don't promise any refund, goodwill gesture or timeframe that isn't in the rules. If the situation isn't covered, write "[FOR A HUMAN TO HANDLE]" and explain to me why in one line, below the message.

The AI now has a single source and an escape route when the case isn't covered. Also remove personal data before pasting anything, and look at what the tool does with conversations: our article on what not to paste into a chatbot covers the question. For the framework that applies to your store, the reference remains the CNIL.

A product page that answers purchase questions

Most generated pages line up "elegant," "superior quality" and "perfect for any occasion." A good page does something else: it removes the doubts of the person who's hesitating. Take a handmade soap maker or a hiking boot seller, the mechanism is the same.

You're writing a product page from this information only.

Features: [DIMENSIONS, MATERIALS, COMPOSITION, CARE, ORIGIN]
Questions and feedback received on this product: [REVIEWS, CUSTOMER QUESTIONS, REASONS FOR RETURNS]
Typical buyer: [WHO, IN WHAT SITUATION]
Price: [PRICE]

Structure:
1. One sentence saying what it is and who it's for.
2. Three benefits, each tied to a specific feature.
3. A "Before you buy" section that answers the most common doubts frankly, including who this product won't suit.
4. A table of features.

No superlatives, no claims the data doesn't prove. At the end, list any missing information that would help the buyer decide.

The "who this product won't suit" part may seem counterintuitive. It serves to limit avoidable returns: if your boots run small, you might as well say so. Be careful with claims about cosmetics, supplements or anything touching on health: a sentence embellished by the AI can be problematic, so read them closely.

Replying to negative reviews in public

A negative review mostly serves the future customers who read your reply. The prompt must therefore target those readers as much as the author of the review.

Here is a public review left on [PLATFORM]:
[REVIEW]

Facts verified on our side: [WHAT HAPPENED]
What we can offer: [REAL ACTION]

Write 2 public replies of 70 words maximum, one more restrained, the other warmer. Stay factual, don't contradict the customer about how they felt, don't reveal any information about their order, and offer a private contact. If the review contains a serious accusation, don't respond to it on the merits and flag it to me.

One limit not to cross: don't ask an AI to write fake positive reviews. The practice is deceptive and can be penalized, however they were written.

Making a hundred support tickets talk

Here's a less obvious use: export your latest tickets, anonymize them, and look for what keeps coming up. You get improvements to product pages and an FAQ based on real questions.

Here are [NUMBER] anonymized customer messages:
[PASTE THE MESSAGES]

1. Group them into 5 to 8 contact reasons, with the number of messages per reason and a short quote for each.
2. For each reason, say whether the problem can be solved by information to add (product page, shipping page, confirmation email) or whether it's an organizational or product defect.
3. Write the 5 corresponding FAQ questions, phrased in the customers' own words.

Present point 1 as a table. Base your work only on these messages. Check that the sum of messages per reason matches the total, and flag the ones you couldn't classify.

Check the counts on a sample, because models easily make mistakes when counting. If you want to plug the result into a spreadsheet or a ticketing tool, ask for a structured output, as explained in our guide to JSON.

Before automating anything, run your last ten complaints through the reply prompt and compare with the message you would have sent. If even one reply contains a promise that isn't in your sheet, complete the sheet and start again.