How to start a Shopify app in 2026 (and where AI actually helps)

Scaffolding a Shopify app has not been the hard part for years. The hard part is everything after npm run dev — auth, billing, webhooks, review. Here is how I start one now, and where an AI assistant earns its keep.

The scaffold

npm install -g @shopify/cli
shopify app init
shopify app dev

That gives you an embedded app with OAuth handled, a tunnel to your dev store, and App Bridge wired up. Pick the official template unless you have a specific reason not to — it tracks Shopify's own changes to session tokens and embedding, and you inherit those fixes for free.

Before writing a feature, get three things right:

• shopify.app.toml — scopes, app URL, webhook subscriptions. Declaring webhooks here instead of registering them at runtime means they survive a reinstall without extra code.
• Session storage — the template ships with something workable, but decide early whether it's Prisma, Mongo or Redis. Migrating sessions later is miserable.
• The mandatory compliance webhooks — customers/data_request, customers/redact, shop/redact. Shopify will not approve you without them, and bolting them on at review time always takes longer than doing it now.

Where AI actually helps

I use Claude on every app now, but not evenly. It is genuinely good at:

• GraphQL against the Admin API. The schema is enormous and the docs are spread thin. Describing what you want and getting a working query with the right connection fields saves real time.
• Polaris scaffolding. Component names, prop shapes, the layout primitives. This is memorisation work, and offloading it is pure profit.
• Webhook payload handling. Give it a sample payload, get a typed handler.
• Reading unfamiliar code. Dropping into a client's existing app and asking what a module does is faster than tracing it by hand.

It is unreliable at:

• Anything versioned. API versions change quarterly. Confident-sounding code against a deprecated field is the most common failure. Check the current version in the docs, not in the answer.
• Billing. The subscription APIs have edge cases around trials, upgrades and currency that are easy to get subtly wrong and expensive to discover in production.
• Rate limits and bulk operations. Knowing when to reach for a bulk query instead of paginating is a judgement call about your data volume.

The pattern that works: let it write the first draft of anything mechanical, and keep the architecture, the billing logic and the data model for yourself.

The part nobody scaffolds for you

App Store review is a real project. Budget for a demo store with realistic data, a screencast, listing copy, and at least one round of rejection. I have never had an app approved first time, and the feedback is usually about the listing, not the code.