How I Built My Entire Shopify Store With Claude AI

That top block is what I wrote in February. Inline styles, an onclick handler, no accessibility, no structured data, nothing responsive. The bottom block is December. Same person, no bootcamp, nobody hired. What changed in between was not talent. It was the way I asked. Three hundred plus conversations across a year, and the output is a real Shopify store that takes real money. I am not going to tell you it was magic. The useful part is that the prompts got written down. Six of them do most of the work, and I am going to put all six on screen exactly as they are saved. Prompt one. It names the platform, the standard and the exact theme version in a single request. Most people write build me a hero section and then spend twenty minutes correcting assumptions the model had no way to get right. Every clause in that paragraph is one assumption removed before it is ever made. All six of these are on the class page, copy ready, at ddsboston dot com. Free, no signup, no email. Grab them now and follow along, because the next four are the ones that actually keep the code honest. Prompt two is the one I would keep if I could only keep one. Without it you get scaffolding, a promising shape with a to-do comment in the middle. With it you get a file you can deploy. Naming the specific things to verify matters more than the word test, because test on its own means nothing to a model. It is worth seeing the shape of the year, because it was not one leap. February was plain HTML with inline styles. By spring the work was structured data, JSON-LD and meta tags on every page. By early summer it was real Shopify sections, Liquid, the theme internals. And in September the important thing happened, which was that the accessibility and mobile rules stopped being something I asked for each time and became a standing instruction. Everything after that point is noticeably better than everything before it, and that is the line I would point at. In practice the loop is short. Paste the master prompt with the file you want changed and you get a complete section back. Then paste the audit demand as a separate message, and the model goes back over its own output and tells you what it got wrong. Keeping those two as separate steps matters more than it sounds. Asking a model to check work it has already produced gets you far more than asking it to be careful while producing it. Then theme check has to come back at zero errors before anything goes near the live theme. Prompt three exists because a model will happily invent an internal link to a page that has never existed on your store. It looks completely plausible and it is a dead end for a customer and a penalty from a search engine. The rule is simple. Ask which pages are real before linking to any of them. Four, five and six are not about code quality at all. One states the certifications the brand actually holds, and names the one it does not, because a model filling a gap with a plausible certification is a false claim on a live product page. One forbids inventing a human expert persona. One forces the native signup so the spam protection is real. These are the prompts that keep you out of trouble. Underneath the prompts is a fixed list of six standards. Accessibility, mobile first, performance, structured data, actually tested, and printable. The list does not change between projects, which is the entire point. A standard you renegotiate every time is a preference. A standard you never renegotiate is what makes the output consistent. Four habits do most of the cleanup work. Prefix every class so your CSS cannot collide with the theme. Declare your Liquid variables at the top so there is one place to look. Check that every schema setting is actually read by the markup, because a setting nobody reads is one nobody notices is broken. And honor reduced motion every single time. So here is the honest summary. The model got better over that year, but that is not what moved my code from the top block to the bottom block. What moved it was writing the constraints down and refusing to restate them from memory every time. The prompt became the standard, and the standard is what produced the output. The full case study is on the site with all six prompts copy ready, plus the before and after code and the standing rules. It is free, there is no signup, no email and no paywall. Take the audit prompt first. Paste it at the end of your next request and see what comes back different.