AI is turning designers into full-stack prototypers, and curation is the bottleneck

The workflow shift. Shopify's product design studio lead, Marvin Schwaibold, describes moving from a Figma-first process to one driven by AI tooling. Figma is now a scratchpad; he works in a terminal with Codex and Claude Code, turning rough sketches into functional prototypes he can test on his phone. The loop is still iterative, but the output arrives much faster, and designers at Shopify use this to prototype admin and Sidekick experiences instead of handing off static mockups.

Why it matters. Schwaibold argues the role is shifting from producing layouts to curating what AI creates. Designers are more the editors now than ever before; AI raises the ceiling but does not know what looks good. For engineering organizations, the implication is that the bottleneck in design work becomes taste, judgment, and the ability to break tasks into subtasks agents can execute. Teams that treat AI as an output machine miss the real skill: defining the problem and selecting among variations.

Concrete example. A side project shows the loop clearly. He fed a video of grass to Codex and got a weak shader, then used ChatGPT to break the scene into terrain, light, and depth. A topographical map helped the model understand the structure, and the final shader contains roughly 217,000 grass blades. He also built DialKit sliders to expose variables like wind and blade length, because chat alone could not tune how the motion felt.

The organizational tension. At Shopify, Schwaibold works inside Sidekick, where AI suggestions can rewrite product descriptions and directly manipulate admin dashboards. He calls it a paradox: beautifully crafted new UI plus AI that bypasses the UI through chat. For engineering leaders this is a governance question, not a taste question. Prototypes are built in separate playgrounds and not allowed into the main repo, yet they spread internally and shape what the CEO decides to build.

What is missing. The evidence is one experienced practitioner's account, heavy on demos and light on production metrics. There are no numbers for adoption, conversion, or engineering cost. Claims like teams of two doing the work of many sound plausible but are unmeasured. The website reviews are useful craft lessons, yet they are not the strategic signal. Engineering leaders should treat this as a direction signal, not a benchmark.

The takeaway. AI does not remove the need for taste; it increases the value of people who can judge and direct output. For organizations, the practical move is to give design teams safe prototyping environments, agent tooling, and a clear boundary between exploratory demos and production code. The full-stack designer is becoming real, but the new bottleneck is editorial judgment, not tooling.