The Claude Setup Playbook for E-Commerce Marketing

A founder told me on a call last week that AI is stupid, especially for marketing and branding. He was right about his AI. He had been asking it to write product descriptions for a brand it knew nothing about, and it had been handing him copy that would fit any product on the internet. Swap the product, the copy still works. That is the tell, every time.

Your AI is not stupid. It is untrained. This page is the fix, written out completely: the whole setup, every step, and every prompt, in the order I would do them.

This is what I run for my own business every day, and it is the first thing I build for a brand. It is six steps across four build days, about ninety minutes of real work, and at the end of it you have an AI that opens already knowing your products, your customer, your voice, and the things your brand would never say.

If you found this from the video series, the map is: seven episodes, six steps, four build days. Episodes 2 to 5 are the build. Everything in all seven is written out below, so you never need to watch a video to follow it.

What do you need before you start?

Five files you already own, one folder, and the Claude desktop app. Nothing you have to make from scratch, and nothing you have to buy beyond an account.

What you need:

1. Product copy for your ten bestsellers. Whatever is live on the site right now, however rough. Copy and paste it into one document if it is easier than exporting.
2. Your last ten marketing emails. Campaigns, not transactional receipts. These carry more of your voice than your product pages do, because somebody wrote them in a hurry and sounded like themselves.
3. A review export. Every review platform exports to CSV. This one file is worth more than the other four combined, and almost nobody uses it.
4. Brand guidelines, if they exist. If they do not, skip it. This setup is how most brands end up writing their first real ones.
5. Your about page, plus the names of the two brands you lose customers to. The competitors are not for copying. They are so the AI knows what you are being compared against when someone is deciding.

What you do not need: to be technical. To write code. To learn prompt engineering. A new subscription to anything. A perfect version of any of the five files above. Written badly is fine. It has to be true and specific, not beautiful.

One thing to check first. This whole setup rests on the AI being able to read a folder on your computer. Before you start, confirm your Claude plan and version support local file access, because everything after Step 1 depends on it.

Step 1. How do you install Claude and point it at your business?

Download the desktop app, make one folder named after your brand, drop your five files in, and point the app at the folder. That is the entire step, and it takes about fifteen minutes.

The desktop app, not the browser tab. This is the difference people skip, and it is the one that matters.

Build the folder

One folder on your desktop, named after your brand. Inside it, four folders and nothing clever:

1. products/ your product copy
2. email/ your last ten campaigns
3. customers/ your review export
4. brand/ guidelines, your about page, the two competitor names

Then open the desktop app and point it at that folder. The exact click path moves as the app updates, so rather than give you a menu path that will be wrong in three months: look for the option to add a folder or a project directory, and select the brand folder you just made.

Then run the test that proves it worked

Do not skip this. It takes thirty seconds and it is the difference between building on something that works and building on something that does not.

What a pass looks like: it names real products, describes a recognizable customer, and quotes lines you can find in your own files.

What a fail looks like: a fluent, confident, entirely generic answer with no quotes in it. That means it never read the folder. Go back and check the folder is actually connected before you write a single other file, because every step after this one is wasted otherwise.

Nothing was created in this step. It was pointed at. That distinction is most of what people get wrong about AI at work.

Step 2. What goes in an AI brand context file?

Six things, on one page, in plain English. One page about your brand beats a hundred clever prompts, and I have watched both for seventeen years.

Make a file called brand/context.md and answer these six. Twenty minutes. Do not write it beautifully, write it true.

1. What you actually sell. Real products in real language. "Linen bedding in eleven colorways, plus a small towel range," not "premium lifestyle goods for the modern home."
2. Who buys it. A person, not a demographic. The most useful sentence you can write here is what they were doing five minutes before they landed on your site.
3. How you sound. Three adjectives, one sentence you would publish tomorrow, and one sentence you would never publish. The second one teaches more than the first.
4. The two brands you lose customers to. Name them, and say in one line why someone picks them instead.
5. Your goal this quarter. One number. Not three.
6. What is off limits. Claims you cannot make, topics you stay out of, anything legal or regulated.

What a filled-in one looks like

Below is a worked example. Harrow Home is invented for this playbook. It is not a real company and not a client. It is here because an empty template teaches nothing.

1. What we sell. Stonewashed linen bedding in eleven colorways, sold as sets and separates, plus a six-piece towel range. Average order around three items. Our bestseller is the sand-colored duvet set in queen.

2. Who buys it. A woman in her mid-thirties who has just moved into a place she plans to stay in. Five minutes before she lands on our site she was looking at someone else's bedroom on Pinterest and thinking her own looks unfinished. She is not price shopping. She is trying to find the thing she saw.

3. How we sound. Calm, specific, unhurried. We would publish: "It softens every wash for about a year, then stops. That is the point." We would never publish: "Elevate your sleep sanctuary."

4. Who we lose to. Brand A on price, because they run sales constantly and we do not. Brand B on delivery speed, because they hold stock and we make to order.

5. This quarter. Repeat purchase rate from 18 percent to 24 percent.

6. Off limits. No health or sleep-quality claims of any kind. No sustainability language beyond what our supplier certification actually covers. We never name a competitor in customer-facing copy.

Notice that none of that is marketing writing. It is a brief. If you would be comfortable handing it to a new freelancer on their first morning, it is right.

If the blank page is the problem, have it drafted from the files you already gathered, then correct it. Correcting is faster than composing.

Step 3. What should your brand never say?

The most useful file in my whole setup is a list of things I am not allowed to write. It is four lists, it takes fifteen minutes, and it does more for how the output sounds than any other single thing in this playbook.

Here is why it works. A language model has read almost everything, so with no instruction it writes toward the average. The average is precisely where every brand in your category already sounds the same. Telling it what you are is useful. Telling it what you are not is what actually separates you, because it cuts away the shared middle.

Make a file called brand/never.md with four lists.

1. Words you never use. Aim for thirty, not five. Mine include elevate, unlock, journey, and no em dashes, ever.
2. Claims you cannot make. If you cannot prove it, it belongs here. This list is also your fastest protection against a compliance problem.
3. Tones that are not you. Mine: fake enthusiastic, corporate, salesy, self-deprecating.
4. Formats you do not do. Listicles, emoji bullets, all-caps subject lines, whatever it is for you.

Then rerun something it wrote before you built the list, unchanged. Watch what disappears. That list is the most concentrated piece of positioning you own, and almost nobody writes it down.

Step 4. How do you make Claude remember your brand?

One file that loads automatically at the start of every session, plus a second file it updates itself. Twenty minutes, and it is the step that turns four days of work into something permanent.

Right now, most teams retype who the brand is every time they open a chat, and the tool forgets it the moment the window closes. That is a tax, and you are paying it several times a day.

The file that loads itself

Make a file called CLAUDE.md in the top level of your brand folder. This is the one that gets read automatically at the start of every session. Here is the actual text to paste in and fill:

The file that updates itself

Make a second, nearly empty file called brand/lessons.md. Every time you correct the output, the correction gets written in here as a numbered rule, and it is applied from then on. That last line of the memory file above is what makes it happen automatically.

This is the difference between a tool your team keeps re-briefing and a system that runs on your business.

Now test it cold

Open a completely new session. Say nothing about the brand. Ask for something specific and see whether it already knows.

Source: Anthropic, Effective context engineering for AI agents.

Step 5. How do you write a product description in your brand voice?

You ask for one, and you give it nothing else. That is the whole point of the four days. No context pasted, no prompt engineering, one instruction.

How to judge what comes back. Do not ask whether you like it. Ask these three, in order:

1. The swap test. Could a competitor publish this by changing only the product name? If yes, it is still generic and the context file is too vague. Go back to Step 2 and get more specific about the customer.
2. The never test. Did anything on the never list get through? If so, the list is too short. Add what slipped through and rerun.
3. The evidence test. Did it name a real spec and a real objection, and could it show you where it got them? If it invented either, say so explicitly and add "never invent a product detail" to the memory file.

Every description after this one starts from here. That is the compounding part, and it is the reason the setup is worth ninety minutes.

Step 6. What do you do when the output still sounds wrong?

You say what is wrong as a rule, not as a feeling, and then you write the rule down. This is the step almost everyone skips, and it is the one that makes the system improve instead of staying flat.

If you have been typing "make it better," that is why it is not getting better. "Better" is not information. Compare:

The rule version fixes this description. Writing it into brand/lessons.md fixes every description after it.

Do that ten times and it stops being wrong in the same ways. That is the whole mechanism. It is not clever, and it is why it works.

The prompt appendix

Every prompt in this playbook, plus the five I use most once the setup is running. Copy them as they are, then change the bracketed parts.

Prompts 01 to 06 are above, in the steps they belong to. These are the ones that come after.

Prompt 10 is worth a note, because it is the one that feels wrong. Pasting your whole catalog in to write copy for one product makes the output worse, not better, because attention gets spread across the hundreds of products that do not matter for this job. The skill is not giving it everything. It is giving it the smallest set of things that matter, which is also how briefing a person works.

Auditing a product page is a separate job with its own prompts, so it lives in its own kit rather than here. Comment AUDIT on any of my videos and I will send it over.

What comes next?

You now have the foundation: a folder it can read, a page that describes the brand, a list of what the brand never says, a file that loads all of it automatically, and a loop that makes it better every time you correct it.

The next layer is the brand brain. That is where the one-page context file grows into the full e-commerce stack: brand voice, ideal customer profile, product catalog, competitors, and the customer review library that feeds all of them. Same architecture, much deeper payload, and it is what makes the difference between output that is on-brand and output that is unmistakably yours.

I teach it in public, one piece at a time, and it is free.