Logan GelzerClaude Code for Shopify: My AI Development Workflow (with Cursor, Gemini & ChatGPT)
By Logan Gelzer, Senior Shopify Expert
TL;DR
Claude Code works well for Shopify development once it's connected to Shopify's AI Toolkit or Dev MCP, so it reads current docs and validates Liquid and GraphQL against Shopify's schemas. I use it for multi-file theme, app and Functions work, Cursor for precise edits, Gemini for browser QA, and ChatGPT for research and review. A human still owns the plan, the tests and the publish button.
I've been building on Shopify since 2014, for more than 200 brands. In that time the tools have changed a lot, from FTP and the old theme editor to Theme Kit, Shopify CLI, Online Store 2.0 and Hydrogen. AI coding agents are the biggest change yet. They don't change what good Shopify work looks like. They change how fast you can get there, and how much time is left over for testing.
Here's the workflow I actually use on client projects, what each tool is good at, and where I don't trust them.
Why generic AI struggles with Shopify (and what fixed it)
Ask a general chatbot for a Shopify section and you'll often get something that looks right: valid-looking Liquid, a plausible {% schema %} block, a GraphQL query with sensible field names. The trouble is that Shopify moves fast. API versions ship quarterly, checkout moved from checkout.liquid to Checkout Extensibility, and Shopify Scripts gave way to Functions. A model working from memory will confidently use fields, filters and patterns that are out of date.
Shopify's answer is a proper context layer. The Shopify Dev MCP server (released March 2025) and the Shopify AI Toolkit (April 2026) connect coding agents to shopify.dev documentation and API schemas, and they validate GraphQL, Liquid and UI-extension code against Shopify's schemas before you ever run it. The toolkit supports Claude Code, Cursor, Codex, Antigravity CLI and VS Code.
Hydrogen goes a step further. npx @shopify/hydrogen setup writes agent skills that match your installed Hydrogen version into .claude/skills and .agents/skills, so the agent follows the version you actually run.
Rule #1 of my workflow
No agent touches a Shopify codebase until it has this context layer.
My stack: which AI tool for which Shopify job
| Job | Tool | Why |
|---|---|---|
| New sections/blocks, metafield-driven templates, multi-file refactors | Claude Code | Reads the whole theme or app, plans before editing, and runs hooks such as Theme Check after edits |
| Line-level Liquid, CSS and JS fixes, and reviewing diffs | Cursor | A fast editor loop with rules scoped to sections/ or snippets/ |
| Clicking through preview themes, visual QA, a second opinion | Gemini (Antigravity) | Its agents can work in the browser as well as the editor and terminal |
| Research, migration mapping, content drafts, PR review | ChatGPT / Codex | A different model catches different mistakes, and Codex can review PRs |
| Docs, schemas, validation, local preview, deploys | Shopify AI Toolkit, Dev MCP, Shopify CLI | Shopify's own source of truth |
I'm not loyal to any one vendor. Models leapfrog each other every few months. The workflow around them (context, plan, validate, review) is what stays the same.
The workflow, step by step
Set up a safe sandbox
Every project starts on a duplicate or unpublished theme, or on a dev store for apps and Functions, and everything goes into Git. I pull the theme with `shopify theme pull`, preview it locally with `shopify theme dev`, and only push with `shopify theme push --unpublished`. The agent never edits the live theme.
Give the agent project memory
In Claude Code, a `CLAUDE.md` file holds the rules for the project: theme architecture, naming conventions, which apps inject blocks, performance budgets, "never use jQuery", "always use `image_url` with `srcset`", and so on. Cursor gets the same rules in `.cursor/rules`, and Codex gets them in `AGENTS.md`. Ten minutes spent here saves hours of corrections later.
Map the codebase before changing it
On an inherited store, I first ask Claude Code to explain: how the product template is put together, which snippets the cart drawer depends on, where app scripts load. This is where AI saves the most time. Reading a messy 1.0-era theme used to take me half a day, and now I get a first-pass map in minutes. I still verify it, because a confident summary can be wrong.
Plan, then edit
For anything beyond a one-liner, I use plan mode: the agent proposes which files change and why, and I edit the plan before any code is written. Most AI mistakes are planning mistakes, and they're cheap to catch at this stage.
Build with validation in the loop
Claude Code drafts the section, Function or GraphQL operation and validates it through the AI Toolkit. A hook runs `shopify theme check` after edits. For Shopify Functions, I scaffold with `shopify app generate extension`, have the agent write the input query and logic, then run it locally against fixture carts with `shopify app function run` before it goes anywhere near a checkout.
QA like a senior developer, not like a demo
This is where the saved time goes. I test the theme editor settings merchants will actually use, translations, mobile layouts, accessibility, app blocks, and edge cases like sold-out variants, stacked discounts and subscriptions. I run Lighthouse before and after. Gemini in Antigravity can click through flows in a browser, and that adds coverage on top of my own testing. It doesn't replace it.
Second-model review, then human review
Before merging, a different model (Codex or ChatGPT) reviews the diff. Then I review it line by line, because I'm the one accountable for it.
Ship with a human on the button
The client approves the preview link, I publish, and I watch the launch. With Git, rolling back takes seconds.
Where I don't trust AI on Shopify
Shopify draws a similar line in its own product. Sidekick can now generate simple admin apps for merchants, but those apps can't touch Functions, themes or checkout. The storefront and checkout work still needs a developer.
What this means if you're hiring a Shopify developer
You shouldn't pay agency rates for a developer to hand-type boilerplate. You also shouldn't hire someone who pastes chatbot output into your live theme. The right question to ask a Shopify developer in 2026 is: "How do you use AI, and how do you check its work?" If the answer includes a sandbox, schema validation, a test plan and human review, you're in good hands.
That's how I work. My rates start at $100/hr with a $1,000 minimum project, and the time AI saves goes into scope and QA, not into cutting corners.
FAQ
Can Claude Code build a Shopify theme?+
Yes. With Shopify CLI and the Shopify AI Toolkit, Claude Code can scaffold and edit Liquid sections, blocks and templates and validate them against Shopify's schemas. It works best with a senior developer planning the changes, testing in the theme editor and reviewing before publishing.
Do I need the Shopify Dev MCP or AI Toolkit?+
For serious work, yes. They give the agent current shopify.dev docs and API schemas, and they validate GraphQL, Liquid and UI-extension code. Without them, AI tools often use outdated APIs.
Claude Code or Cursor for Shopify?+
Both. Claude Code is stronger for multi-file features and refactors. Cursor is faster for precise edits and diff review. Many Shopify developers, me included, use both.
Can Gemini or ChatGPT write Shopify code?+
Yes. Gemini via Google Antigravity and ChatGPT via Codex can both use the Shopify AI Toolkit. I mainly use them for browser QA, research and second-opinion code review.