Add preferred source
Marvin Schwaibold, who leads Shopify’s Product Design Studio and earlier ran the Brooklyn studio Molly, argues that AI coding tools have raised the ceiling for designers by letting them build working prototypes instead of static mockups — but the scarce skill is now judgment, not output. Speaking on the Y Combinator Startup Podcast, he describes a workflow where Claude Code and Codex do the heavy lifting, prototypes spread “like wildfire” inside Shopify until CEO Tobi Lütke blesses them, and teams of two now do the work of many. He also reviews five startup landing pages and finds the same recurring flaws: no visual hierarchy, tiny type, scroll-jacking carousels that aren’t tappable, and sign-up walls that appear before the user sees any value. The report connects his example to a broader market question: as AI tooling floods platforms with generated content — Etsy just banned AI-made patterns and recipes to protect buyer trust — curation and accountability become the scarce asset, and the design roles that command value will be the ones that can judge output rather than merely produce it.
Key Elements
Marvin Schwaibold runs the Product Design Studio at Shopify, a job that did not exist in this form two years ago. Before that he ran a small Brooklyn design studio called Molly, took on two or three clients at a time, and spent more than a year building a single website for the branding agency Collins. Then Shopify bought the studio, folded it in, and handed his team two of the most consequential surfaces in the company: the redesign of Shopify’s admin and the interface layer for Sidekick, its AI assistant.
The tool that made the leap possible, he says, is the same one making designers everywhere nervous. “AI has definitely raised the ceiling for all of us, but it’s not a crafter,” Schwaibold said on the Y Combinator Startup Podcast. “It can execute complexity and technology that I couldn’t hand write code… but it doesn’t know what looks good and what looks bad. And that’s for me to decide.”
That single sentence frames the shift that matters for anyone watching tech labor markets. The bottleneck in design used to be production — drawing screens, speccing motion, waiting on engineering. Now the bottleneck is judgment, and judgment was always the part no one could automate.
The prototype swallowed the document
The most concrete change inside Shopify is what happens to a design before anyone writes production code. Schwaibold’s team builds working prototypes, not static mockups, and the prototype travels further than a spec ever did.
“Prototypes are a million times more useful than any PRD document you can write,” he said. “Those spread around the company like wildfire, and if Tobi sees a prototype that he likes, then we can go off and build that.”
Tobi is Tobi Lütke, Shopify’s founder and CEO, who told the design team the company was “entering a new era and that future needs to be designed,” Schwaibold recalled. “I want to design for the future and I need to hire good designers who can help me paint a picture.”
The structure that came out of that mandate is unusual: two tracks that pull in opposite directions but resolve into one product.
| Track | What it does | Where it lives |
|---|---|---|
| Admin redesign | New home design and a facelift of the entire product, polished in Figma then brought to the web | Figma → web |
| Sidekick | Lets merchants run their store by conversation rather than by clicking | AI infrastructure layer |
The two converge in Shopify’s mobile app, which is getting more conversational — you talk to Sidekick about your business instead of manipulating menus. Schwaibold is explicit that the admin redesign is prototyped in a separate playground, not production code: it would never be allowed into Shopify’s main codebase. But the whole studio is deliberately embedded in the company’s actual roadmap, which he argues is the only way an internal studio earns real influence. “We also wanted to have a larger impact and see if our strategy would hold up on interface paradigms with millions of users versus the few users that would scroll around the Spencer website,” he said.
That shift in ambition — from a few thousand visitors on a studio site to millions of merchants — is what the Shopify acquisition bought. About two years at Molly, capped at a handful of clients, was the proving ground; Shopify was the stress test.
Teams of two, doing the work of many
The operating change Schwaibold describes is blunt, and it is the part worth watching for anyone modeling software-company headcount.
“Right now I’m seeing armies or teams of two going off and doing the work of many, and because the tooling is just allowing us to do that, it’s making it a lot easier to get by.”
At Molly, every animation was hand-built by a developer, and art direction meant coordinating “a thousand counterparts.” Now the designer starts in Figma, but treats it as a scratchpad that “transcends into real demos” — he tests on his phone, driven by Codex and Claude Code in the terminal, and returns to Figma mainly to lay out a layout or explain an intent the model isn’t getting from text.
The working method is decomposition: break the build into subtasks, then feed them to the agent one at a time. The agent’s real advantage isn’t brilliance, it’s stamina. Codex will not object to ninety-one attempts at an animation the way a human engineer would. He calls it “an infinitely patient collaborator,” and the human in that loop is less a maker than an editor-in-chief of generated output — a title he uses himself.
The grass shader: what agents can do, and where they break
The episode’s most detailed case study is grass. Schwaibold filmed real grass, fed the video to Codex, and asked for a pixel-perfect web shader using the same hex codes. The first output, in his words, “obviously doesn’t look at all” like the target.
Then came the interesting move: he screenshotted the failure, took it to a different AI agent, and asked why the first one wasn’t understanding him. The advice back was to specify where the light came from, define the shadows, and split the image into roughly six parts and render each separately before recombining.
The breakthrough came from structure. The model had been rendering grass as a flat layer. Once it understood there was terrain underneath — varying depths and heights — and re-rendered the grass on top of that terrain, the scene turned realistic. The discovery came with a warning: exposing the underlying terrain map revealed that it was “completely wrong and skewed,” a mistake introduced five prompts earlier that he hadn’t noticed.
A lunch with a colleague named Josh produced the tuning layer. They installed a tool called Dial Kit so Schwaibold could inspect the shader’s variables directly. When he asked Codex how many grass-blade variations existed, the answer was about 217,000 — differing only by roughly 0.6 to 0.8 in value. They built a toggle to see the variation, plus sliders for blade length and wind.
Something like that control panel would never have been worth anyone’s time before. Now it is trivial, and that is the quiet argument buried in the anecdote: AI doesn’t just generate the artifact, it generates the tooling around the artifact.
Why chat is the wrong interface for feel
Dial Kit exists because Schwaibold got tired of trying to describe motion in conversation. The tool — installable via npm run or a prompt from its creator’s site — lets you create and tweak variables in any project. He demonstrates it on a carousel inspired by the portfolio of design lead Mike Mattis, where dragging an artifact down and releasing it sends an energy ripple into neighboring artifacts. After “91 prompts,” a side panel finally exposes parameters like “pull boost,” the rate at which dragging accelerates.
His reasoning is a useful rule for anyone building with AI: he knows when the feel is off, but back-and-forth chat cannot get him there. The values have to be exposed as knobs, not argued about in prose.
This connects to a principle he returns to throughout a review of viewer-submitted startup sites — interfaces must be interruptible, because human thought is interruptible. He points to Apple’s hardware and software as the long-standing proof. His own weekend side project made the point: a grid of brass-study videos that open on tap and close on drag-down. Codex built it, but the interaction wasn’t interruptible — after pulling an asset down you had to wait for it to settle before tapping another. He drew crude colored boxes in Figma showing that a new tap mid-swipe should cancel the previous one, and the model read the drawing roughly as he intended. He couldn’t say it in words or voice, but he could draw it.
Five landing pages, one diagnosis: nothing tells the eye where to land
Schwaibold’s site critiques are where his rules get concrete, and the diagnosis repeats. He borrows it from a German typography professor: a poster that isn’t “dictatish” gives the eye nowhere to land.
| Site | What works | Primary flaw |
|---|---|---|
| Nstack | Drops you into a product-like desktop; whimsical retro feel | No single focus; sidebar, chat history and headline all compete; four same-colored CTAs |
| SkillSync | Clear what the company does; dark mode | Type too small and dense; features buried behind tabs users won’t click |
| InstaPlay | AI game generation is genuinely fun; strong sharing loop | No copy under the title, so users jump to thumbnails; sign-up wall appears before the value |
| Mintlify | Reductive in the right ways; credible logo wall; well-spaced sections | Card headlines too small; animation reused until it loses impact; key copy buried under a big screenshot |
| Nullify | Well-made videos | Two rotating headline placeholders disorient; five font sizes in one viewport; scroll-jacking; carousel cards aren’t tappable |
| Instinct | Instantly legible as a space company; serif headline is a bold choice; real launch video builds credibility | Launch animation too slow to read; logo wall competes with the message; serif used for body copy; carousel doesn’t drag |
The specifics matter because they’re the same mistakes AI makes easy to commit. On InstaPlay, the sign-up wall appears before the user has seen their generated game, which breaks the acquisition loop — value first, then sharing as the growth effect. On Nullify, he argues the company should use its own product: put a video at the top instead of paragraphs, since “we’re not going to read any of this anyway.” On Instinct, he wants the launch clip to play on hover immediately, because half a second is the attention span, and suggests all three clips loop simultaneously since each is a brief moment. On SkillSync, he’d compress the feature set to a three-up of the top three capabilities, each tappable into a sheet.
He leaves behind two reusable assets: a three-up component as a default for feature sections, and Blossom Carousel, built by colleague Jasper Voss, a free GitHub carousel that swipes, scrolls and stacks on mobile — his answer to the many carousels that only respond to arrow buttons.
Craft as a sales argument, not decoration
The standards Schwaibold applies trace back to Molly, and to a bet that craft is legible to buyers. The studio’s site for Collins used a bookshelf of case studies where books level up and land on one another; a dedicated engineer named Jesper worked on that bounce because an organic landing “eludes at the idea” that the shelf is alive. The homepage’s ambient gradient behind the video reel was copied from the lighting behind wall-mounted TVs — his partner JTel duplicated and blurred the video so the background matched the on-screen rhythm.
Why spend a person’s time on a tab interaction? Because clients hiring a branding studio read it as a signal. “Clients coming to the website are looking for those kinds of Easter eggs,” he said. “They want to know that if this agency takes this much detail to produce a website of that quality, what kind of Easter eggs are they going to put into our projects when we work with them?”
The same logic produced Molly’s content philosophy, which they called “primal.” As he described it: “It’s a way to just basically understand content on a website… everything is just one layer away and everything is progressively disclosed so you never overwhelmed the user with too much content.” That’s the lens he turns on every reviewed site.
The gap between what tools produce and what teams can judge
The tension Schwaibold names but doesn’t resolve is the one that should occupy investors and operators. AI makes it trivial to generate variations, which makes curation and taste the binding constraint. Yet the same tools let a solo founder ship a site with five font sizes and a scroll-jacking carousel. The tools raised the ceiling; they did not install the eye.
That gap shows up well beyond Shopify. Etsy this week banned AI-generated patterns and recipes from its marketplace and required sellers to show real photos of what they actually made, citing buyers who say they can no longer tell what is AI-made — a decision that treats curation as a trust asset with a revenue line attached. Siemens, working with Nvidia on factory automation, is promising to lift automation from roughly 30–40% to 70–80% and spending €500 million in Erlangen to get there, another case where the models are ready and the human judgment about where to deploy them is the constraint. Even the CFTC is convening a “Frontier Forum” on October 28 to work out how AI agents can trade inside existing derivatives rules.
For Schwaibold, the conclusion is that roles converge toward full-stack rather than disappear: design, engineering and product management moving toward the same person, or the same two-person team, with the agent doing the typing. The scarce input is no longer the ability to produce — it is the discipline to interrupt, reduce and decide what is good. That is a skill no model has yet been able to take, and it is the one worth paying for.
Full content available at:AI Raised the Ceiling for Designers
- Etsy Cracks Down on AI Chaos, Bans Sale of AI-Generated Recipes and Patterns
- US Stocks: All Three Major Indexes Rise as SpaceX and Super Micro Earnings Lift AI Shares
- Amazon Open-Sources TorchNeuron to Breach CUDA’s Moat, Heating Up AI Chip Software Ecosystem Battle
- Siemens Aims to Lift Factory Automation to 80% With AI, Backed by 500 Million Euro Erlangen Expansion
- US CFTC to Host ‘Frontier Forum’ on the 28th… AI Agent Trading Regulation Emerges as Key Issue
