Marvin Schwaibold on AI-Driven Design, Motion Craft, and Reviewing Five Viewer Websites

Aaron Epstein introduces Marvin Schwaibold as the leader of the product design studio at Shopify. The episode is framed as a behind-the-scenes look at his process, showing how he uses the latest AI tools to bring designs to life, followed by a review of viewer-submitted websites.

Schwaibold says the process itself has not really changed; it remains iterative and sequenced, with many iterations and variations. What has shifted is the tooling. Where he once worked with tools like Figma and aesthetic layouts, he now works in a more AI-driven workflow.

He uses Figma as a scratch pad for ideas, but those ideas quickly turn into working demos that he can use. He returns to Figma only occasionally, when he needs to lay something out or explain it to an AI. He also names tools like Codex and Claude Code and says he spends a lot of time in the terminal, reaching a testable output on his phone or computer much faster.

Asked how much of the work is driven by AI versus himself, he says he drives AI to do much of the heavy lifting, while quality control and decision-making remain with the designers. He says the output is no longer the bottleneck; the challenge is being selective about what to pursue and curating what AI creates.

He adds that AI is extremely capable of getting a lot done as long as the person understands the complexity of the work and can break it down into parts.

Before joining Shopify, Schwaibold ran a design studio in Brooklyn called Molly together with his studio partner J Tal. They ran the studio for about two years before joining Shopify and integrating the studio into Shopify.

The studio worked with clients including Collins, Apple and Google, as well as many smaller clients. The work covered a broad mix of web and product design, along with content strategy and hierarchy for clients' websites. Because the studio was small, it kept close relationships with clients, meeting them weekly for many weeks.

He says the studio never took on more than two or three clients at a time in order to stay focused. The episode then shows a reel of the studio's work, where the speakers note organic animations and a gradient and shadow that change to match the background color on each screen. Schwaibold says the idea came from the ambient lighting behind wall-mounted TVs.

Schwaibold describes a website made with the agency Collins, which his team worked on for over a year. The home page is organized as chapters, with each section leading into further exploration, so the site feels like exploring rather than being pushed to act. Content is progressively disclosed, so users are not overwhelmed.

The bookshelf section, which holds the Collins case studies, is one of his favorites. The speakers note that its hover bounce feels different from what is typically seen. Schwaibold says this was not made with AI tooling but was hand-crafted by a design and engineering partner. He laid out the spacing, size, colors and rows in Figma, then handed the bounce animation to an engineer named Jesper.

He explains that the books level up on top of each other and land once more as the last book settles, and that this has to feel organic. If it feels bouncy in that way, it reinforces the idea of a living bookshelf rather than a book, and he argues these seemingly unimportant moments are what people pick up on while scrolling.

Schwaibold describes the concept behind a prototype: bring UI to the user in context rather than having them click through the interface. In the envisioned version, a merchant could spend much more time on the home surface, using Shopify as a dashboard to delegate work to Sidekick, which would directly change the store.

He says the current version is still a prototype environment with a developer UI panel. The design team builds changes, demos and ideas directly into this interface so they can test and build on each other's work. Once a demo reaches a point where the team aligns, each person branches off to build a version they think adds value, and then the versions come back together in the same demo and shared language.

Asked how designers move from Figma to building live features, Schwaibold says the admin team has a playground where they test code that would not be allowed into Shopify's main repository.

Schwaibold says he uses Figma more as a scratch pad to map out ideas, which he then gives to Codex. He finds it easier to decompile something visually, and says Codex can also interpret such an image well. He says the tool understood what he was doing and how he was thinking about the motion. He could not articulate it in text or voice, but he could draw it out.

He identifies two points. First, current AI tools and models are extremely capable of creating almost anything, but the user has to know what they are doing and be able to break a concept, such as a shader, into its components and understand the commands for each part.

The speakers then discuss a grass shader. The speaker had a set of images and videos of real grass and wanted a shader that looked identical while still being controllable. He fed the video to Codex and asked it to be meticulous, pixel perfect, matching hex codes and understanding each part of the image before recreating it.

Schwaibold explains that a colleague, Josh, looked at the grass demo at a lunch and said it looked good but the length of the blades should be examined. Schwaibold asked Codex how many grass blade variations existed. Codex said there were about 217,000 different blades, but they differed by only about 0.7 to 0.8, so the team built a toggle to see them directly.

He wanted to avoid going back and forth through chat, so he asked for the underlying structure to be exposed. This also revealed his own errors: the original terrain map was wrong and skewed, because he had misinterpreted something several prompts earlier.

He says that adjusting values helps visualize what he wanted to see, and he tested wind movement, which can get quite wild. He says experiments like this would have been out of reach before AI, and that designers are becoming more full-stack, as engineers and product managers are too.

Asked how designers can raise the quality of their AI-assisted work, Schwaibold says AI is an infinitely patient collaborator and teammate. Codex will not get mad if he needs 91 approaches to get an animation correct, and he says a human engineer probably would.

His advice is to tinker with the tooling. Describe what you want, and when the result is off, do not give up; explain how it differs from your vision. He suggests sending a website with a carousel you like, a screenshot, or a color palette for the tool to copy, then going back and forth to shape the result together.

For designers, his method is to copy. He says he is currently imitating a carousel animation that has been around for a while, using the AI to look under the hood and decompile how it is built. He does this on his own time so he does not need to bother engineers, and he says this teaches him to be a better manager and collaborator.

The reviewers look at Skill Sync, a free local-first Mac app that carries memory and skills across agents and teammates. They say it is clear what the company does. The page is in dark mode, and the reviewers say the fonts are very small, with many percentage-based sizes in the typography that feel off.

They argue that when a page abstracts functionality while staying very detailed, it becomes hard to understand what the product does, requiring a lot of reading. The features sit in sidebar tabs, and the reviewers note that most users will not click through them, so the information stays hidden.

Schwaibold suggests condensing content into principles and main points, then rallying more information under each. He says founders also benefit from asking what the one thing they do is. For Skill Sync, if it is skills syncing, the page should focus on that, make it impressive, and get people using the product quickly.

The reviewers note the page is not very tall and that users are comfortable scrolling. Rather than compressing and hiding important information behind tabs, they suggest making the page longer.

They also warn that AI tools make it easy to generate many small animations that add no value. Schwaibold says a memorable abstract piece can communicate more than dense text, and that a specialist in illustration or information graphics may be better than AI at abstracting features. He says sometimes being more artistic and abstract goes a long way.

The reviewers look at Insoflight, a site for creating games from prompts. They say the labels create, my games, for you and explore are unclear, and that clicking my games leads to a sign-up page even for someone not signed in. For you returns to the game just viewed, and explore is the page already open.

They say the explore page removes the input field, so a user can no longer create a game there. The fast-forward icon used to return home is not obvious and can look like an error, so the speaker clicked back instead. The reviewers say the product is compelling, since people could play with games made from a prompt for hours.

The reviewers say that once someone has created a game and is proud of it, being able to share it with friends is probably a nice growth hack. They say there is a lot of value in this kind of work and praise the team.

The reviewers praise Mintlify, a company that provides knowledge infrastructure for agents and documentation, as clean and professional. They credit reductive composition that shows only what is necessary: a clear headline, a short paragraph, a prominent get-started button, and a dashboard that suggests where documentation lives. The background graphic adds to the mood rather than taking away from it.

The homepage sections act as first chapters to subpages. The reviewers say the order should reflect importance, with products, solutions, resources and pricing, because people drop off as they scroll, so important content must sit at the top. They also say the products section probably needs its own homepage because there are so many products.

They describe the page as more breathable, with everything given its own space, which helps users know what to focus on.

The reviewers look at Netlify, which turns documents and ideas into narrated animated videos. They say that cards in the carousel cannot be clicked; users must press a small icon, which is not good, and the card should be tappable. When a viewer opens one, it goes full screen, making it hard to close or move to other examples.

The reviewers would prefer the video to zoom up to about 90 percent of the screen rather than full width, so users can keep swiping through examples and see the company behind them. They also suggest that a video of the company's own studio could prove the point that users do not read long text.

The reviewers note that the moon imagery and a satellite convey aerospace work, but say the company's purpose should be made clearer. They describe the homepage copy as reductive, with room for one more line explaining what is built. The more specific section describes building the lunar lander for today's space industry, through launch, cruise and land.

On hover, the first launch clip did not show much to one reviewer, while the cruise clip started as soon as the cursor entered. The reviewers suggest the three clips could loop at the same time because each is a concise moment, which would make the point faster. They say the land clip does not clearly show the moon approaching until the shadow grows, and that this moment alone should be shown.

The speakers say that if technology, company, news and careers are important to a business, they should be the homepage chapters rather than making visitors use the top navigation. Simple sheets can then open any piece of that content, which would make the whole site more digestible.

They discuss a technology section whose video looks like real footage rather than renderings, and say its information could sit on the homepage as an explore-technology chapter instead of relying on a separate page. They also mention a carousel that does not drag or swipe with a trackpad and works only through its buttons.

On typography, the reviewers say a serif font should be used only for headlines as a branding element, with a sans-serif such as Inter or Instrument for body copy.