What does a designer do when AI can produce a working interface in minutes? In this Design Review conversation, Shopify product design studio lead Marvin Schwaibold gives a grounded answer: the designer still sets the bar, breaks the problem down, and edits the result. He shows Aaron Epstein how an interactive prototype can expose a bad assumption faster than a polished static mockup, then applies the same standard to six submitted websites.
Build something real early, but judge it as a user would. Make the next action obvious, let people interrupt motion, and expose the variables behind an interaction so you can see what is actually wrong. Faster output is valuable only when it gives the team more chances to test and refine.
Watch the Design Review Episode
Source note, 29 October 2026: The episode and supplied transcript are the source for Marvin's workflow, the 91-iteration account and the six live critiques. I checked the linked projects and Shopify's admin article separately. I did not reproduce the shader, run usability tests on the submitted sites, or audit Shopify's internal process.
The Process Is Still Iterative; the Prototype Arrives Earlier
Marvin does not say that Figma is obsolete. He uses it as a scratchpad when a diagram explains an interaction better than a paragraph. He then asks coding agents to build working demos that he can test on a phone or computer. When a drag-to-dismiss gallery would not respond to a second tap until the first animation had finished, he drew a three-box sequence to show the interruptible behavior he wanted. That small visual specification gave the agent something concrete to implement.
The practical shift is from approving a picture of a product to trying the behavior itself. Marvin describes the human role as editing and quality control: choose what to keep, reject the generic version, and decompose a hard visual problem into smaller parts. Code generation speeds up the loop; it does not decide whether the experience is clear or pleasant.
Shopify Needs Both a Better Interface and an AI Way Through It
Marvin describes two related efforts at Shopify: improving the admin interface and making Sidekick useful while a merchant works in it. Shopify's own admin redesign account documents a collapsible side navigation, a more consistent visual language, and Sidekick as a floating prompt that can expand into a panel for longer work. The design tension is not "chat versus UI". A merchant still needs to see orders, settings and context; an assistant should be available without covering the task.
That makes the interview's design principle operational: a new layer should shorten the path to action, not merely make the screen look more futuristic. Shopify's write-up explains its intent; it does not prove that every merchant prefers the new layout.
The Grass Shader Worked When the Hidden Structure Became Visible
For a weekend experiment, Marvin gave Codex video of real grass and asked for a browser shader. The first render missed the reference badly. Instead of repeating "make it more realistic," he compared screenshots, used another model to interrogate what the first model had misunderstood, and broke the image into lighting, shadows, blades, wind and terrain. The decisive clue was a topographic map: the grass sat on small mounds, not a flat sheet. Once the terrain became explicit, the output could be corrected.
He then used DialKit to surface blade length, wind and other parameters. That exposed a skewed terrain map from an earlier iteration. This is the strongest lesson in the episode: when repeated prompting stalls, expose the system's intermediate state and tune its real controls. A visual control panel can reveal a wrong assumption that a plausible final render hides.
Marvin published the Grass Shader repository with reference recordings, WebGL2 demos and an offline iPhone project. Its README documents a Node.js 22.13+ browser setup; the separate native iPhone build needs a Mac and Xcode. It is a runnable study, not a one-click mobile app for every device.
Why 91 Carousel Iterations Were Not 91 Wins
Marvin also shows a portfolio carousel inspired by Mike Matas. He says it took 91 prompts to reach the interaction he wanted. The number is not a productivity benchmark. Its value is that he could keep adjusting the drag, release and neighboring-card response without treating an early version as finished. DialKit turned abstract requests like "less force" into visible parameters such as pull boost.
There is an important boundary here: studying the behavior of an admired interaction can teach you a technique, but a client deliverable still needs its own identity, assets and validation. The episode's closing mention of Blossom Carousel offers a separate, public implementation built on native scrolling with pointer drag. Its documentation is useful for developers; it does not remove the need to test keyboard use, pause controls and mobile behavior. The W3C carousel guide explains those accessibility requirements.
Six Site Reviews Point to One Rule: Show the Value, Then Let People Explore
The six critiques are not a ranking of the companies. They are live observations from one viewing session. The recurring issue is a mismatch between what the product can do and what its first screen lets a visitor understand or try.
| Site in the episode | What Aaron and Marvin notice | Practical design move |
|---|---|---|
| Endstack | A playful desktop metaphor competes with its headline and several similar-looking actions. | Decide the first user task and make its entry point unmistakable. |
| SkillSync | Its purpose is fairly clear, but dense tabs and decorative motion hide the strongest capabilities. | Lead with a few concrete benefits; let readers open deeper details. |
| Instaplay | Visitors can browse games, but the make-your-own-game value is easy to miss and a sign-in interrupts the first try. | Let the visitor experience the core result before asking for commitment. |
| Mintlify | The restrained hierarchy and social proof feel credible; some small labels and repeated effects are less useful. | Keep the focus, but make important product claims legible. |
| Knowlify | The video work is strong, yet animated headline fragments and difficult-to-open examples delay comprehension. | State the video offer plainly and make examples directly explorable. |
| Instinct Space | The lunar focus lands immediately, while important real footage is buried and some hover sequences take too long. | Move proof of the technology closer to the first screen and shorten the wait. |
Marvin's objection to scroll jacking fits the same rule: an animation that takes control of the page can delay the user who already knows what they want. For a motion-heavy page, test whether people can pause, navigate by keyboard and use a reduced-motion version. These are our additional review checks, grounded in the W3C carousel tutorial and MDN's reduced-motion guidance, not a claim that the episode audited accessibility.
Try a One-Page Design Review
Use one real visitor task, not a general request to "make the site better." This original prompt is an editorial exercise based on the interview; it has not been tested on the six sites.
One page, one user task
Find the first real point of friction.
Projects and Sources by Role
| Resource | Why it matters here | Before you use it |
|---|---|---|
| Grass Shader | Marvin's source, reference footage and live WebGL2 study. | Follow its README; browser GPU support and the separate iPhone build matter. |
| DialKit | Exposes visual and motion parameters for live tuning. | Choose deliberate ranges; the panel is a development aid, not a substitute for user testing. |
| Blossom Carousel | Public, native-scroll-based carousel mentioned near the end. | Test accessibility and interaction on your actual content and devices. |
| Shopify admin redesign | First-party account of the admin and Sidekick placement. | Product intent is not independent proof of user satisfaction. |
Video Chapters
Design process and Shopify
| Time | Topic |
|---|---|
| 00:00 | How the design process has changed |
| 03:25 | Inside Shopify's product design studio |
| 05:58 | Molly, the Brooklyn studio |
| 08:58 | The Collins bookshelf |
| 11:51 | Why the future needs design |
| 13:24 | How the studio works inside Shopify |
| 15:32 | Sidekick and the redesigned admin |
| 18:27 | Designers who build |
Prototyping motion
| Time | Topic |
|---|---|
| 21:58 | Rebuilding real grass as a shader |
| 23:48 | The topographic-map insight |
| 26:16 | DialKit and exposed variables |
| 28:45 | Advice for founders and designers |
Reviewing six websites
| Time | Topic |
|---|---|
| 31:04 | Endstack: a single primary goal |
| 33:12 | One primary action per page |
| 35:02 | SkillSync: show the important features |
| 36:43 | What tabs can hide |
| 38:25 | Illustrations that add no information |
| 40:49 | Instaplay: show the playable product |
| 43:04 | Show the games, not just thumbnails |
| 45:34 | Mintlify: trust through restraint |
| 47:25 | Social proof and less-is-more |
| 49:35 | Homepage sections as chapters |
| 51:42 | Knowlify: let people explore the work |
| 54:02 | Make examples directly explorable |
| 56:06 | Why scroll jacking frustrates |
| 58:34 | Instinct Space: show the real technology |
| 1:00:42 | Launch, cruise, land |
| 1:02:42 | The technology page |
| 1:04:16 | Blossom Carousel |
Sources and Limits
Checked 29 October 2026. The Y Combinator interview and the user-supplied transcript support the demonstrations, quotes and site-review observations. Project details come from Marvin's Grass Shader README, Josh Puckett's DialKit README, and Jesper Vos's Blossom README. Shopify's admin article describes its own design choices. The accessibility additions draw on W3C WAI and MDN. We did not independently reproduce the builds or test the featured websites.