Original Summary

I've spent the last few days building Shopify Theme Builder, an open-source Claude Code plugin (also a plain Agent Skill for Codex, Cursor and OpenCode). You describe a shop, and it produces a real Shopify theme you own. What it does You write one message: what the shop sells, who buys, how it should feel, what it should never look like, where the photos are. Claude asks only what it can't work out (languages, currency, menus, shipping), creates a free Shopify development store, and writes three different designs. A small local app shows them live on the real store. You pick one, and Claude builds the rest: product, collection, cart, search and 404 pages, the copy in the shop's languages, the photos in place. Before handing over it runs Theme Check, an axe accessibility audit with keyboard tests and Lighthouse, then uploads the theme unpublished. It never publishes anything. The screenshot is a demo shop it built from the example prompt in the README. The brand is made up and the photos are AI-generated; the theme, the store setup and the copy are what the plugin produced, unedited. How I built it with Claude Claude Code wrote nearly all of it. I'd describe what I wanted, Claude turned it into small GitHub issues with acceptance criteria, and a workflow ran a fresh agent per issue: tests first, then typecheck, Theme Check, commit. A final agent checked CI and fixed anything red. The most useful part was the testing loop. Every few hours I'd start from an empty folder and have Claude build a real store from one prompt, while a second Claude session watched the transcript and opened an issue for every problem. Six of those runs found things no unit test would: a French shop whose design names with accents were refused, Shopify serving its password page to every new browser, the EU cookie banner covering every screenshot, a coffee shop previewed with Shopify's demo snowboards. Each one went into the next queue. What I'd reuse on any Claude Code project One fresh agent per issue, run in sequence by a workflow. Each agent starts clean, reads only its issue, and has to pass the same checks before committing. Quality stayed steady over 200+ issues instead of drifting as one long session filled up. Push without waiting for CI, then one sweep agent at the end. Waiting on CI took about 40% of the queue's time; dropping it took a queue from about 15 minutes per issue to about 5. A hook that blocks editing a file type until the matching skill is loaded (Liquid, React, Tailwind). It sounds strict, but it cut the "forgot the conventions" kind of mistake almost to zero. End-to-end runs watched by a second session, as above. Unit tests told me the code did what I asked; these runs told me I'd asked for the wrong thing. The prompt I used for the demo ``` Use the shopify-theme-builder skill to build a Shopify theme for my shop. Shop: Olmo Ceramica, handmade stoneware tableware from a two-person studio in Florence. Refe

  • 情报分类：技术学习与提效
  • 分类依据：内容涉及技术、AI、软件工具或工程实践
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/26 03:18:42