Every new client store was its own build, and the design was only as good as the person rebuilding it.
- Each restaurant brand gets its own store, designed to look like the brand.
- The store manager has to be able to change any word or image without a developer.
- Products and images come from the company's catalog systems, joined by SKU.
- The client's live store stays read-only. Everything is built on a copy and a draft theme.
A design goes in, a store comes out
A designer finishes every page in Claude Design or Figma. The agent builds it as one client inside a shared Shopify theme and fills it with real products. On the left is the design. On the right is the store it built.
Ask people only what people must decide
Most answers are already in the files. The agent decides what it can, labels anything it wrote itself as editable, and asks in plain conversation only for things like a price promise, a legal line, or a brand conflict. It hands back a store to click through and a short account of every decision it made.
Designers label intent, not code
I wrote one page of rules a designer and their AI design tool can follow. Every word and image is one of four things: live store data, editable content, an action, or fixed. A labeled design builds without back-and-forth, and an unlabeled one still builds, with the agent asking when it can't tell.
Every word stays editable
Headlines, banners, and promo lines become settings the store team changes in Shopify. Prices, sizes, colors, and stock come live from the catalog. Nothing the designer drew is lost, and nothing needs a developer to change.
Pages the design never drew
A design rarely covers every page a store needs. When one is missing, the agent builds it in the design's own type, color, and spacing, labels it as ours, and leaves every word editable.
The same store on a phone
Every page is built for both sizes and checked against the design at both. On a phone it keeps the design's type, spacing, and order, not a squeezed desktop.
A team of agents, run like a team
Product manager, architect, engineers, UX, and QA, each an AI agent, with every assignment, review, and handoff in GitHub. No agent merges its own work, QA is assigned by the PM and never by the author, and every decision is written down with how it's enforced. I set the scope and answered what only a person could.
- One shared theme, with each client's design built as itself.
- A one-page contract that designers and their AI tools can follow.
- Questions asked only when a person has to decide.
- Every word and image editable by the store team.
- "It renders" means someone looked at a picture of it, not that a test passed.
Five real store designs run as stores on one shared theme, and a sixth was added through the same intake.