A whole catalog, written from your own specs

What ecommerce and retail teams use it for

Retail prompts succeed or fail on one habit: the spec sheet and the policy go into the prompt, and the model adds nothing they do not say. Each job below lists the templates and tools that do it.

A worked example: 4,000 product pages before a migration

Take a hypothetical home goods store moving to a new platform with 4,000 product pages written over eight years by different people. The new theme needs a consistent title format, five bullets and a description of 80 to 120 words for every product, built from the spec sheets the store already has.

Twenty products are enough if they are the right twenty. Pick the ones a human copywriter would slow down on: a set sold as one item, a rug whose spec gives size in two units, a mug the supplier calls “lead-free,” a lamp with no wattage listed. In the Prompt Tester, a “JSON has keys” check confirms every output has title, bullets and description; a Length check holds the title to your theme’s limit; a “Does not contain” check lists claims you only make when the spec says so. An AI judge can then score whether every bullet traces back to the spec. The Prompt Tester and Model Compare need an API key; the template, Grader estimate and calculator do not.

What it costs

The chart prices a larger version of the same job: a hypothetical store rewriting 15,000 product pages in a month, 500 a day. Each call carries about 1,500 tokens in (a fixed block of style rules and examples, plus one spec sheet) and 450 out. Half the input is the fixed block and can be read from cache. Catalog work can wait for results, so the chart uses batch pricing where the provider offers it. That is about 22.5 million input tokens and 6.75 million output tokens.

Batch suits catalog jobs and does not suit a live support chat, where customers will not wait hours for an answer. Support is priced separately: a chat that resends its history on every turn costs more per conversation than its word count suggests. The support chatbot cost post works through that case, and the cost calculator takes your own volumes.

A sample prompt

A prompt that turns return reasons into a “fit and sizing” note for a product page, one of the most useful lines a clothing or footwear page can carry.

You write the "Fit and sizing" note for a product page at
{{store_name}}.

Product specs:
"""
{{specs}}
"""

Return reasons for this product in the last 90 days, one per
line, as customers wrote them:
"""
{{return_reasons}}
"""

Write a fit note of at most 60 words that tells a shopper what
the returns suggest, for example whether it runs small or the
sleeves are long. Use only patterns that appear in at least
{{min_mentions}} return reasons. Quote no customer. Make no claim
about materials, origin or care that is not in the specs.

If the return reasons show no clear pattern, write exactly:
NO CLEAR PATTERN

The fixed “NO CLEAR PATTERN” answer gives the Prompt Tester something exact to test: feed it a product with scattered return reasons and an “Equals expected” check fails any note the data did not support.

Guardrails

Practical guidance for US sellers, not legal advice. Other countries have their own consumer rules, and marketplaces add theirs.

Every product claim needs a basis

The FTC expects you to have evidence for what your copy says, and a model that writes “BPA-free” or “hypoallergenic” because similar products say so has given you a claim you cannot support. Origin claims are stricter: an unqualified “Made in USA” claim requires the product to be “all or virtually all” made in the US, as the FTC’s Made in USA guidance explains. Keep a list of claims your store only makes from the spec, and test for them.

Reviews: answer them, never write them

The FTC’s Consumer Reviews and Testimonials Rule bans fake reviews and testimonials, including AI-generated ones, along with incentives that depend on a review being positive and undisclosed reviews by staff. Asking all buyers for a review is fine. Generating one is not.

Shipping promises and delay notices

Under the FTC’s Mail Order Rule, you need a reasonable basis for any shipping time you state, and 30 days applies when you state none. When an order will be late, the notice must seek the customer’s consent to the delay and offer to cancel with a refund. If a prompt drafts your delay emails, put those elements in as fixed requirements and check for them.

Marketing email follows CAN-SPAM

Every marketing email needs a working unsubscribe, honored within 10 business days, a valid postal address and a subject line that is not misleading, per the FTC’s CAN-SPAM guide. A model rewriting your template can drop the footer. Test that it is still there.

Bots speak for the store

In 2024 a British Columbia tribunal held Air Canada responsible for wrong bereavement-fare information its website chatbot gave a customer (Moffatt v. Air Canada). Assume a customer can hold you to what your bot says. Keep refunds and exceptions with a person, test the bot against prompt injection, and never paste card numbers into any prompt.

Related templates, guides and posts

• Product listing from your specsTitles, bullets and descriptions with nothing invented.Template
• Explain a refund or return decisionWithin policy, in plain words, with the next step.Template
• AI prompts for customer support that follow policyThe support side of a store in detail.Guide
• What an LLM call really costsTokens, caching and the arithmetic behind the chart.Guide
• What a support chatbot costs at 10,000 chats a dayWhy conversation history drives the bill.Blog post
• The cheapest model that passesHow to choose a model for a catalog job.Blog post