Wonder is Swell’s AI agent. It builds a working online store from a short brief without the merchant leaving the dashboard: describe the brand, confirm a plan, and you get a live storefront with real products, cart and a checkout.
Swell
Visit site
↗
Swell is an API-first, headless ecommerce platform. Merchants run catalog, checkout, subscriptions and customer accounts on Swell and build whatever frontend they like on top. Seven years there, from early product designer to Chief Design Officer, is why Wonder could lean on everything the platform already does well.
YEAR
2019-2026
COMPANY
Swell
ROLE
Branding
User interface
Design engineering
Two stages, kept apart
Wonder works in two stages. The first is creative generation: from the brief it writes a plan, which pages, which sections on each page, and a set of design tokens for type, colour, density and shape, then generates the storefront from that plan on mock data so the model can concentrate on design without fighting APIs. The second is automated assembly: a build engine runs a sequence of Claude Code steps that import the catalog, wire products, cart, checkout and accounts to live Swell data, and push the result as a standard Swell store.
The split matters because the plan exists as data before any code is written. A full build takes about an hour and costs real money, and two kinds of error were only visible after that spend: a missing Contact page, or a brief that said minimal coming back loud. Rendering the plan as a summary needs no build, so a review step sits between planning and generation. It lists the pages, each page’s sections in plain language, how the style and colour were read, and a note on anything inferred from a weak signal. The merchant confirms or corrects before anything is spent.
The decisions
The storefront name is the only required field. The original flow opened with a brand form, name, brief, colour, pages and four visual style radios, and it failed at both ends: a first time merchant with no formed brand was asked to pick density before seeing anything, and an established brand with a brand book could not fit any of it into a textbox. Everything beyond the name is now offered as add what you have, and established brands can drop a brand book, URL or logo and let the agent extract voice, mood, colour and type. People start by asking for something, not by defining a style system.
Inference fills what the merchant did not touch, and an explicit choice always wins. The structural tokens that most decide whether a store reads as committed or as a template, density, shape, outline and motion, used to be radios defaulting to the safe middle, and inference was not allowed to set them. Leaving the radios alone produced bland structure with colour and type layered on top. The form now tracks touched against untouched state per token, so a radio sitting on its default is not a choice, and the agent can set those values when it reads minimal as spacious and outline free.
Sections are assembled from a reference library rather than generated from nothing. Each section role has a reference image, a written spec and a copy stripped component, and the agent can instantiate one with zero drift or author a new one with the library as the exemplar. Taste is easier to curate than to prompt, and the library lines up with the marketplace, where professionally designed sections and themes can double as starting points for generation.
Most of the editor was built rather than drawn. A local Claude Code instance ran against the real repos, each branch had one goal, each prompt stated what must not change, and every diff was reviewed. Design questions got answered with working software and engineering review happened on real merge requests.
In August 2026 Wonder generated and assembled a complete storefront end to end for an investor demo: eight sections, live catalog, checkout and accounts, with a library section inserted into the generated page without drift. The editor chrome, structure panel, inline text editing and per section regeneration, was feature complete and in review at the point of leaving Swell. The three ideas worth carrying anywhere are to ask less and infer more, to show the plan before spending the build, and to curate taste rather than prompt for it.