Shopify AI Generator

The problem

Every new Shopify store starts from Dawn, the default theme, and looks like every other Dawn store until someone spends a day moving settings around by hand. What a client actually hands you is not a spec, it is a link to a site whose look they like.

What it does

Scrapes a reference website’s structure and branding, uses an LLM to map that visual language onto the Shopify Dawn theme, then drives the Shopify CLI to actually build it (colors, typography, homepage sections) plus a sample product dataset.

Originally specced against a paid SDK. Rebuilt on the headless CLI instead so it costs nothing per run, matching the same rule every other tool here follows.

Why it is built this way

Five linear steps: read a config, fetch the reference site and reduce it to a compact summary rather than raw HTML, have a model map that summary onto Dawn’s own setting names, shallow-clone Dawn itself, then merge the result into the theme. Generation runs through the local claude CLI in headless mode rather than the paid SDK, which stays in package.json as a documented fallback and is imported by nothing. Images come from a free keyless endpoint, so a complete run still costs nothing per store.

What was hard

Two things. Getting usable JSON back on Windows: the prompt goes in over stdin rather than as an argument, because the CLI is a .cmd shim and Windows quoting mangles anything long, fenced code blocks are stripped before parsing, and a parse failure retries once with a "return only valid JSON" reminder before giving up instead of writing a broken theme. And writing into Dawn without breaking it: settings_data.json is merged read-modify-write so unknown keys survive, and every rebuilt homepage section needs a unique key, a real Dawn section type and an entry in the order array, or the theme renders nothing at all and says nothing about why.

What it ships with

• 5-step pipeline
• Theme + product CSV output
• Paid SDK designed out

Stack