travcjohnson Claude for ecommerce and retail businesses
This page is four things you can set up in claude.ai this week: a batch of product descriptions written in your voice, one approved reply for each question your inbox keeps asking, a full campaign set built from one promotion, and your reviews and tickets sorted into a fix list. Every prompt is here to copy, and every one of them tells Claude to keep its hands off facts only your systems know. No code, no plugins, nothing to install.
Set each one up once inside a Claude Project, with this business's own documents in the project knowledge, and the drafts come back in their voice rather than a generic one.
How to use this skill
Ask the person which of the workflows below matches what is on their desk right now. Then walk them through that one only. Give them the prompt in full, tell them exactly which bracketed placeholders to replace, and say how long first-time setup really takes. Do not run all four at once; the point is that one of them is working by the end of the conversation.
Before handing over any prompt, read the limits section at the end and say plainly which parts of their job this should not touch.
1. Write a batch of product descriptions in your voice
What happens today. Forty new SKUs arrive as a supplier spreadsheet of attribute codes and a folder of photos. You write six good descriptions on Monday, and by the twelfth you are copying the last one and swapping the color.
What happens instead. You paste the spec rows once and get a first draft for every SKU in your format, with the title, the bullets, the long copy, and a flag on any claim nobody has verified yet.
First-time setup: about 30 minutes.
You are writing product descriptions for my store. Match the voice in the examples below. Use only the attributes I give you.
Our voice, taken from five descriptions we are happy with: [paste five of your best existing descriptions here, or write SEE PROJECT KNOWLEDGE if you uploaded them]
Words and phrasings we never use: [list them, for example luxurious, must-have, elevate, perfect for everyone]
Who buys these and what they compare them to: [one or two sentences]
The products, pasted from our supplier sheet or PIM export: [paste the rows here, one product per line, with every attribute you have]
Rules that override everything else:
1. Never state a measurement, material, weight, care instruction, certification, country of origin, warranty term, or compatibility fact that is not in the data I pasted.
2. Never state stock levels, shipping times, delivery dates, or prices, even if the source sheet contains them. Write [STOCK], [SHIPS], or [PRICE] and I will fill them in from our system.
3. If an attribute is missing and the copy needs it, write [MISSING: what you need] instead of guessing.
For each product give me:
1. A product title under 70 characters, front loaded with what the thing actually is.
2. Four bullets, each one a concrete attribute and why a buyer would care about it, in that order.
3. A description of 90 to 130 words that opens with the problem this product solves, not with the brand.
4. A meta description under 155 characters.
5. Alt text for the main photo, describing what is visible rather than repeating the title.
6. A line called CHECK BEFORE PUBLISH listing every claim in your draft that a human should verify against supplier documentation.
Then, at the end, give me one list: every attribute you needed and did not have, grouped by product, so I can go get it once instead of forty times.
2. Answer the same twenty questions without retyping them
What happens today. The shared inbox asks the same things every week. Sizing, where is my order, can I change the address, is this the right fit for my model, how do returns work. Two people retype those answers all day and every version sounds slightly different.
What happens instead. You have one approved reply per question, in your voice, with blanks where real order details go. A new hire on their second day sends the answer you would have sent.
First-time setup: about 25 minutes.
You are helping me build a reply library for my store's customer service inbox. The point is one approved answer per recurring question, in our voice, that anyone on the team can use.
Our actual replies, pasted from our inbox so you can hear how we talk: [paste ten to fifteen real replies you were happy with, with customer names, emails, and addresses removed]
Our policies, in our own words: [paste your returns, exchange, warranty, and shipping policy text, or write SEE PROJECT KNOWLEDGE]
The questions we answer over and over: [list them, for example sizing, order status, address change, return request, damaged on arrival, discount code will not apply, is this compatible with X]
Rules that override everything else:
1. You cannot see our store, our inventory, our carriers, or any order. Never write a stock level, a tracking number, a ship date, a delivery estimate, or an order status. Use a placeholder like [ORDER NUMBER], [SHIP DATE FROM SYSTEM], or [TRACKING LINK] wherever a real detail belongs.
2. Never invent a policy. If a question cannot be answered from the policy text I pasted, say so and write the question in a list called POLICY GAPS instead of drafting an answer.
3. Do not apologize twice in one reply and do not promise anything the policy does not already promise.
For each question give me:
1. A short reply, under 80 words, for the version of this question that is simple.
2. A longer reply for the version where the customer is already annoyed. Acknowledge the specific problem in the first sentence, state what happens next, and give one clear action.
3. The subject line, if it is an email rather than a chat reply.
4. A one line note: SEND AS IS, or NEEDS A HUMAN, and if it needs a human, what they have to check first.
End with POLICY GAPS: every question our written policy does not actually answer, in priority order.
3. Turn one promotion into a full campaign set
What happens today. The sale starts Friday. You still owe the announcement email, two reminders, the banner line, the callouts for the featured products, and the social captions, and the offer terms are still buried in a Slack thread.
What happens instead. You paste the offer, the dates, and the featured SKUs once, and every piece comes back as a first draft with identical terms and dates, in the same voice, ready for you to cut and approve.
First-time setup: about 25 minutes.
You are drafting a full campaign set for one promotion at my store. Everything you write has to use the same offer terms and the same dates, word for word.
The offer, exactly as it will run: [paste the discount, the code, what is included, what is excluded, and any minimum]
Start and end dates and time zone: [paste them]
Featured products with the prices I want shown: [paste the SKUs, names, and prices, or write NO PRICES IN COPY]
Who this goes to: [for example past buyers in the last twelve months, list subscribers who have never bought, VIP segment]
One past campaign that did well, pasted so you can match the voice: [paste the email that worked, or write SEE PROJECT KNOWLEDGE]
Words and phrasings we never use: [list them]
Rules that override everything else:
1. Use only the prices, dates, discount amounts, and exclusions I pasted. Never calculate a sale price unless I gave you both numbers, and never round.
2. Never write scarcity you cannot see. No stock counts, no "almost gone", no shipping cutoff date unless I supplied it. If a line needs one, write [STOCK NUMBER] or [SHIPPING CUTOFF].
3. State the exclusions anywhere the discount is mentioned. If the terms are ambiguous, list the ambiguity instead of writing around it.
Give me, in this order:
1. The announcement email: five subject line options with the one you would send marked and why, preview text under 90 characters, and a body under 200 words with one primary call to action.
2. A reminder email for people who opened and did not buy.
3. A last call email, sent the final day, that does not repeat the announcement word for word.
4. Three site banner lines, each under 60 characters.
5. For each featured product, two sentences that could sit on the product page during the sale.
6. Three social captions, each a different angle, no hashtag padding.
7. Three text message length lines under 160 characters including the required opt out wording placeholder.
End with two lists: every placeholder a human still has to fill, and every place the offer terms need to appear for the promotion to be honest.
4. Turn reviews and tickets into a product fix list
What happens today. You know the three star reviews are telling you something useful. Reading four hundred of them is a job nobody owns, so the same complaint gets rediscovered in a meeting a quarter later.
What happens instead. You paste the export and get complaints grouped by cause, split into product problems, page problems, and expectation problems, with the customer's own words under each one so you can check the grouping yourself.
First-time setup: about 20 minutes.
You are analyzing customer reviews and support tickets for my store to find what we should actually fix. Work only from the text I paste. This is not a summary exercise, it is a diagnosis.
The product or category: [name it, and paste the current product description so you can compare what we promise to what buyers say]
The reviews and tickets: [paste the export here, with customer names, emails, addresses, and order numbers removed]
What we already know is a problem: [list it, or write NOTHING YET]
Rules that override everything else:
1. Count only what I pasted. Never estimate what percentage of all customers feel something, never extrapolate to our whole customer base, and never compare us to an industry norm.
2. Every theme must be supported by two or three verbatim quotes from the text I pasted. If you cannot quote it, do not name it.
3. If a theme rests on only one or two mentions, label it THIN and say so plainly.
Give me:
1. The themes, ordered by how many of the pasted items mention them, with the count of pasted items for each and the verbatim quotes underneath.
2. For each theme, one cause from this list: product defect, sizing or fit, description or photo mismatch, packaging or transit damage, instructions unclear, expectation set by our marketing, or price to value.
3. A split into three buckets: fixable by editing the product page today, fixable by changing packaging or process, and needs a product or supplier change.
4. The exact sentences in our current description that are setting a wrong expectation, quoted, with a replacement line for each.
5. The five things you would fix first and the reason each one is ahead of the others.
6. Anything in this text that reads like a safety issue, a returns pattern worth investigating, or a claim we make that a buyer is disputing. Flag it, quote it, and do not soften it.
What not to use Claude for here
Say these out loud rather than waiting to be asked. Being trusted is worth more than being impressive.
- Anything that depends on what your systems say right now. In a plain browser chat with no connectors, Claude cannot see your inventory, order status, tracking, carrier delays, live prices, or your storefront, so it must never state them. A description that says in stock and ships tomorrow, or a chargeback response that asserts a delivery date Claude cannot see, is worse than no draft at all. Keep those as bracketed placeholders that a person fills from the real system before anything goes out.
- Product claims a human has not verified against supplier documentation. Claude will write a confident sentence about materials, certifications, country of origin, warranty terms, or compatibility because that is how product copy usually reads, and it has no way to know whether yours is true. Health, safety, supplement, cosmetic, and children's products are the sharp end of this: phrases like clinically proven, FDA approved, non-toxic, hypoallergenic, or BPA free carry real regulatory weight in the US. Draft with Claude, then verify every claim against the spec sheet or certificate before it is published.
- Raw customer data. Strip names, email addresses, shipping addresses, phone numbers, and order numbers out of any export before you paste it, and never paste payment card details anywhere. You lose nothing analytically, because the useful signal is in what people wrote. Know your plan too: on Free, Pro, and Max accounts, whether your chats are used to improve Claude is a setting you control in Privacy Settings, while Team and Enterprise accounts sit under Anthropic's commercial terms instead.
Questions this business usually asks
Will Google penalize product descriptions written with Claude?
Google's published guidance on generative AI content is about the content, not the tool: the problem is material produced mainly to manipulate rankings rather than to help a buyer. Three hundred near identical descriptions were already a ranking problem when a human wrote them, and Claude makes it faster to produce that mistake at volume. The fix is the same either way. Give each product the specifics only you know, from fit notes to what buyers actually ask before they order.
Do I need a paid plan for this?
You can start free. Projects, which hold your voice examples, policy text, and spec sheets so you stop re-explaining your brand in every chat, are available on all accounts, and free accounts can create up to five projects. Project files are capped at 30MB each with no limit on how many, as long as the total fits Claude's context window. Paid plans add expanded project knowledge capacity and higher usage limits, which starts to matter once two or three people are running these workflows daily.
How do I stop three hundred descriptions from sounding identical?
Feed it your real work instead of adjectives. Put five descriptions you are genuinely proud of into the project knowledge, plus a list of words you never use, and require one concrete attribute per product that appears nowhere else in the batch. Tell Claude to write [MISSING] rather than fill a gap with atmosphere. Then rewrite the first line yourself on anything that matters, because the opening is where generic copy gives itself away.
Can Claude connect to Shopify and see my orders?
Not in the workflows on this page, and that is on purpose. You paste in, you paste out, and nothing touches your store. With web search turned on, Claude can read a public web page you give it a link to, which is useful for checking your own live product page or a competitor's, but it still cannot see your admin, your inventory, or a single order. Anyone telling you otherwise is describing an API or connector setup that needs a developer and a security conversation first.
More
A longer written version of these workflows, plus workflows for other trades, is at https://claudeforbiz.com/claude-for-ecommerce . Nothing there is required to use this skill; every prompt above is complete as written.