The Shopify AI Toolkit: Claude, Cursor and VS Code

Every week someone asks which AI tool they should use for their Shopify store, and the honest answer is that the question is slightly wrong. These three tools are not competing for the same job. They are good at different things, and the reason people bounce off them is almost always that they picked the one built for a different task.

Here is what each one is actually for.

Table of Contents

• The one-line version
• VS Code — the free entry point
• Cursor — the theme editing seat
• Claude Code — the operator seat
• Which one do you actually need
• A worked example
• What none of them fix

Key Takeaways

The one-line version

VS Code edits files. Cursor edits files with an AI that can see the whole project. Claude Code does work — including work that has nothing to do with files.

That last distinction is the one that matters for most store owners, and it is the one nobody explains.

VS Code — the free entry point

Microsoft's editor, free, with a Shopify Liquid extension that gives you syntax highlighting and basic validation. Paired with the Shopify CLI, it is a perfectly good place to edit a theme.

What it is genuinely good at: learning what a Liquid template actually is. If you have never opened a theme outside Shopify's own code editor, doing it here once — with syntax highlighting and a local preview — teaches you more about how your store renders than a month of tutorials.

Where it stops: the AI assistance is bolted on rather than native. It completes lines; it does not hold your whole theme in its head. For a one-line fix that is fine. For "why is my collection page rendering the wrong template," it is not.

Reach for it when: you want to make a small, known change and you already know which file it lives in.

Cursor — the theme editing seat

An editor built around the AI rather than beside it. It indexes your whole project, so you can ask about the theme as a system instead of file by file.

What it is genuinely good at: sustained theme work. Shopify themes spread one visual change across a section, a snippet, a stylesheet and sometimes a schema block. An assistant that has read all of them at once can tell you which four files a change touches. One that reads a single open file cannot.

Where it stops: it is an editor. It is excellent at changing code and has no opinion about your store's data, your ad account or your email list. If your question is "which products lost conversion rate last month," you are in the wrong tool.

Reach for it when: you are doing real theme development over days rather than minutes.

A caveat worth stating plainly: on Shopify themes, a confident AI edit can silently break something that only shows up on a template you did not open. Preview before you push, every time. That advice is not tool-specific and it is not optional.

Claude Code — the operator seat

A terminal agent rather than an editor. It reads and writes files, but it also runs commands, calls APIs, and connects to external systems through MCP — including Shopify itself.

That last part is the difference in kind. Through the Shopify connection, Claude Code can query your actual store: orders, products, inventory, customers. So the work is no longer "edit this Liquid file," it is "find every product with no compare-at price in the sale collection and tell me the revenue impact." We covered the setup separately in connecting Shopify to Claude Code and what the Shopify connector can do, so this post will not repeat it.

What it is genuinely good at: multi-step work that crosses systems. Pull the data, analyse it, make the change, verify the change, write it up.

Where it stops: it lives in a terminal. There is no visual canvas, no design surface, and if you want to see your theme change as you type it, this is the wrong seat.

Reach for it when: the job involves your store's data, your marketing channels, or more than one step.

Which one do you actually need

Most Shopify operators need one, not three.

You are a store owner who edits the theme occasionally. VS Code with the Liquid extension. That is genuinely it. Adding an AI-native editor to a job you do four times a year is paying a subscription to make an infrequent task marginally faster.

You are building or heavily customising a theme. Cursor. The whole-project awareness pays for itself within a week of sustained work.

You are running the store rather than building it — marketing, merchandising, analysis. Claude Code, because most of that work is not file editing. This is where the majority of Shopify operators actually live, and it is why they bounce off the editors: they were sold a code tool for a business job.

You do both. Cursor for the theme, Claude Code for everything else. That is a real and common split, and the two do not conflict — they are looking at different things.

A worked example

This site runs on the terminal-agent pattern, and one piece of it is a useful illustration because it is the kind of thing an editor cannot do.

Every blog post here is a markdown file. A build script reads them, generates the page data, writes each post body out as JSON, and then generates a static HTML file per route with the right title, description and canonical tag. That is already multi-step: read files, transform, write, generate.

The part that makes it an operator job rather than an editing job is the gate on top. Since August 2026, any post dated after the cutoff must declare an evidence: field naming the one thing in it that exists nowhere else. If the field is missing or too short, the build fails. Not a warning — a non-zero exit that stops the deploy.

Writing that gate meant: changing a build script, defining the policy, wiring it into the pipeline, then deliberately breaking a post to confirm the gate actually fires and restoring it. That is a sequence of actions across files and commands, with a verification step. An editor helps you write line 47. An agent does the sequence.

The same shape covers most Shopify operator work: pull the data, decide, change something, check it worked.

What none of them fix

Worth saying, because the tooling conversation tends to skip it.

None of these tools knows your brand. They know what you tell them. The difference between AI output that sounds like your store and output that sounds like every other store is entirely in the context you supply — which is a separate problem, covered in making Claude sound like your brand.

None of them makes a bad offer work. If cold traffic does not convert, faster theme edits will not change that. The constraint is upstream.

None of them removes the need to check. Every one of these tools will produce confident, plausible, wrong output. On a live store that costs real money. Preview, verify, then ship.

If what you actually want is the marketing work rather than the tooling, here is the full map of what we hand to Claude across client accounts — and what we deliberately don't. Or book a strategy call.