Context
Marvin Schwaibold leads the product design studio at Shopify, where his team is rethinking major product surfaces for the AI era, especially around Sidekick and the Shopify admin. Before Shopify, he co-founded the design studio Molly, where highly crafted motion, hierarchy, and interaction design were part of the product story. In this episode, he explains how AI changed his workflow from Figma-first design into rapid coded prototyping, then applies the same thinking in live reviews of startup websites. It matters to PMs because the same shift changes product work too: if prototypes are now cheap, the value moves from producing artifacts to making better product judgments faster.
The Big Idea
AI is turning design from a craft bottleneck into a judgment bottleneck. When almost anyone can generate interfaces, motion, and code quickly, the real advantage shifts to deciding what deserves to exist, what should stay hidden, and what actually helps the user.
At Shopify, Marvin's team uses AI to move from ideas to working demos quickly, then uses human taste to edit, refine, and reject. For PMs, the implication is clear: static specs and broad idea decks matter less when a team can show the product, test the behavior, and align around something real.
Key Insights
Figma became a scratchpad
Marvin's process still starts with iteration, but the center of gravity moved.
- Figma is now mostly a scratchpad for rough layouts, visual thinking, and explaining a concept to AI.
- The real work quickly moves into coded demos built with tools like Codex and other AI coding workflows.
- He returns to Figma only when he needs to lay something out clearly or show AI a visual structure that is hard to describe in words.
For PMs, this is a shift in what an early artifact should do. A mockup is no longer the end product of discovery. It is often just the fastest bridge to a testable prototype.
Taste got more valuable
AI can do more of the execution, but that does not reduce the importance of design judgment. It increases it.
Marvin's point is that the output itself is no longer the hard part. Designers are now more like editors who curate what AI produces, decide what meets the bar, and keep pushing until the result feels right. He is explicit that AI can write code and produce complexity, but it does not know what looks good, what feels coherent, or what a user will trust. That is still a human call.
For PMs, the equivalent is product taste. If AI can generate ten flows, ten landing pages, or ten positioning options, the scarce skill is not creation. It is choosing the right one.
Decompose complexity for AI
AI is powerful, but Marvin does not describe it as magic. He describes it as a collaborator that needs structure.
- Good results depend on understanding the problem well enough to break it into smaller parts.
- When AI failed to recreate a realistic grass shader, Marvin did not just ask again. He decomposed the problem into light, shadow, terrain, blade variation, and motion.
- He even used ChatGPT to help him diagnose why Codex was failing, then generated a topographical map so the model could understand the terrain under the grass.
This matters to PMs because AI work quality often reflects problem framing quality. If a team cannot decompose the job, the model will usually produce vague or brittle output.
Prototypes beat documents internally
At Shopify, Marvin says prototypes are far more useful than PRDs for getting attention and alignment.
- His team can redesign a surface in Figma, port it into a tangible prototype quickly with AI, and circulate that demo across the company.
- Those prototypes spread "like wildfire" because people can react to something concrete.
- If Shopify CEO Tobi Lutke sees a prototype he likes, the team can move much faster than if the same idea arrived as a written concept.
For PMs, this is a strong argument for showing, not just describing, especially when the product change is experiential or workflow-heavy.
Small teams can show more
One of the biggest leverage changes Marvin sees is team size versus output.
He says he is now seeing one-person efforts or teams of two do work that previously required many more counterparts across design and engineering. The reason is not that collaboration disappears. The reason is that more of the early proving work can happen before a large team needs to get involved.
For PMs, this changes how buy-in happens. Instead of gathering a big group to imagine a future state, a small group can build it first and invite others to react to the evidence.
Restraint clarifies product value
The website reviews all pointed to the same product lesson: most teams add too much.
- On weaker sites, too many buttons, tabs, screenshots, and moving parts competed for attention.
- On stronger sites like Mintlify, the page felt trustworthy because it was reductive. It showed only what was necessary, gave each section room, and made the primary action obvious.
- Marvin repeatedly pushed toward one clear goal, fewer competing elements, and progressive disclosure instead of dumping all information on the first screen.
That is not just web design advice. It is core product design advice for onboarding, dashboards, settings, and AI copilots.
Mental Models & Frameworks
Designer as editor
This is Marvin's clearest model for the AI era.
- AI does the heavy lifting on generation and implementation.
- The human sets the bar, selects the right direction, and rejects low-quality output.
- The work becomes less about drawing every pixel and more about curating a system of options.
Use this model when a team is overwhelmed by AI-generated choices. The right question is not "can the model make this?" It is "who is responsible for taste and final judgment?"
Progressive disclosure beats clutter
Marvin described Molly's website approach as making content feel explorable instead of demanding attention.
- Keep important information one layer away instead of exposing everything at once.
- Let each section lead into deeper exploration.
- Avoid overwhelming users with too much content on the first view.
Use this when designing homepages, product onboarding, or complex admin surfaces. It is especially useful when a product has many capabilities but a first-time user only needs one entry point.
One surface, one job
A recurring critique in the website reviews was that many pages did not know what they wanted the user to do first.
- Every homepage or product surface should have one primary goal.
- Every other element should support that goal or get out of the way.
- If a user lands and sees five equally loud actions, the design has not made a decision.
Use this in roadmap reviews too. If a launch tries to explain every feature, it often fails to communicate the one reason the user should care.
Interruptible interfaces feel better
Marvin cares a lot about interfaces that respond immediately and let the user change course mid-action.
- In one example, an AI-built interaction was technically correct but wrong in feel because the user had to wait for an animation to finish before tapping the next item.
- He drew a simple Figma diagram with colored boxes to show the desired behavior: while one item is dismissing, another should be able to take over immediately.
- He makes the same argument against scroll jacking. Users should never feel trapped by an interface's timing.
Use this model for motion, mobile gestures, conversational UI, and any workflow with multiple fast user decisions.
Expose the underlying variables
Marvin repeatedly values prototypes that let the team inspect and tune the system.
- With Dialkit, he exposed variables like grass blade length, wind, and interaction force instead of tweaking them blindly through prompts.
- Exposing controls also surfaced hidden mistakes, like a terrain map that had been implemented incorrectly several prompts earlier.
This is a strong PM pattern for AI features too. If a system depends on thresholds, ranking weights, or prompt choices, create ways to inspect and tune them instead of treating the model as a black box.
Decision Principles
Principle: Prototype before alignment
When: a product idea depends on feel, sequence, workflow, or interaction quality. Why: people align faster around something they can use than around a description of something they must imagine.
Principle: Reduce before decorating
When: a page or surface feels busy, unclear, or over-explained. Why: trust usually comes from clear hierarchy and restraint, not from adding more motion, more buttons, or more visual effects.
Principle: Keep hidden content small
When: a team wants to place core messaging behind tabs, carousels, or hover states. Why: most users will not inspect every hidden area, so the product should assume unseen content is effectively absent.
Principle: Default to native interactions
When: deciding whether to add custom scrolling, unusual gestures, or novel navigation patterns. Why: users already know how websites and apps should respond. Fighting those learned behaviors creates friction that distracts from the product itself.
Trade-offs & Nuance
Full-stack needs guardrails
Marvin argues that designers are becoming more full-stack, but not in a reckless way.
- Shopify designers build in an admin playground, not directly in the main production repository.
- AI makes it easier for designers to visualize and test ideas in code.
- That does not mean every prototype should ship or every designer should own production implementation.
The trade-off is speed versus system safety. Teams should widen who can prototype without pretending that prototype code and production code are the same thing.
Chat is not always best
Shopify is simultaneously redesigning traditional admin UI and making more of the product controllable through Sidekick.
That is an important nuance. AI does not eliminate interfaces. In many cases, users still need visible controls, context, and confidence. The opportunity is not "replace all UI with chat." It is choosing where conversation reduces complexity and where direct manipulation remains better.
Expressiveness can hurt clarity
Marvin loves motion and crafted interactions, and he explains why details can matter. On a design studio site, a polished interaction can signal depth of craft and become part of the sales story.
But he also shows the failure mode. If motion slows navigation, if hover states hide meaning, or if animation competes with the headline, the same craft becomes noise. The trade-off is brand expression versus comprehension.
Common Mistakes
Mistake: Letting every feature shout
Several reviewed sites tried to show every capability at once.
- The result was too many equally loud buttons, panels, and visual elements.
- Users could not tell what to do first.
- The product felt broader, but also less understandable.
The better approach is to pick one primary action and build the page around that action.
Mistake: Hiding value behind tabs
Marvin pointed out that users often will not click through every tab or explore every hidden section.
That makes tabs risky when they contain core value propositions or critical product explanations. If the most important feature is hidden by default, many users will never see it. Better options are a shorter list of top capabilities, longer scrolling sections, or tappable cards that progressively open more detail.
Mistake: Forcing custom scrolling
Marvin is unusually direct on this point: smooth scrolling and scroll jacking often make products worse.
- They slow down interaction.
- They make users fight the interface to stop where they want.
- They break learned expectations from the rest of the web.
The fix is simple. Use native scrolling unless there is a very strong reason not to.
Mistake: Confusing motion with meaning
AI makes it easy to generate slick-looking animations, but motion is not automatically helpful.
In the reviewed sites, some animations added atmosphere without improving understanding. Marvin's standard is that motion should either clarify the product, show a real behavior, or strengthen a deliberate brand feeling. If it does none of those, it is decoration competing for attention.
Practical Application
Prototype the next workflow
Before the next roadmap or design review, turn one important user flow into a working prototype instead of presenting it only as slides or a PRD.
- Pick a workflow where sequence or interaction quality matters.
- Use Figma only to rough out states if needed.
- Move quickly into a coded demo, even if it runs in a sandbox or prototype environment.
This works because teams react faster and more honestly to something they can click.
Sketch states before prompting
When an AI tool is failing to build the interaction you want, stop rewriting the same prompt.
- Draw the states, transitions, or gesture logic in Figma or on paper.
- Keep the sketch crude if needed. Marvin used simple colored boxes to show interruptible behavior.
- Feed that visual back into the model so it can parse the structure.
This is especially useful for gestures, animations, and any state change that is hard to explain in prose.
Audit first-screen focus
Run a quick focus audit on any homepage, onboarding step, or dashboard.
- Ask what the single primary user action should be.
- Count how many other elements compete with that action on the first screen.
- Demote, remove, or defer anything that pulls attention away without increasing understanding.
If a first-time user cannot tell what to do in a few seconds, the page is trying to do too much.
Add controls to prototypes
For any AI-assisted or interaction-heavy prototype, expose the key variables early.
- Create dials, toggles, or visible settings for the behaviors you are tuning.
- Use those controls to test edge cases and catch hidden implementation mistakes.
- Keep the controls in the prototype long enough for product, design, and engineering to learn from them.
This speeds iteration and makes model behavior easier to reason about.
Keep onboarding interactions native
Review the first-use experience of any new product or launch page and remove avoidable friction.
- Make cards tappable if users will expect them to open.
- Avoid tiny controls when the whole object could be interactive.
- Remove custom scroll behavior that changes the speed or feel of navigation.
The goal is not novelty. The goal is to make first contact with the product feel obvious.
Questions to Consider
- Where is your team still using a document to explain an experience that would be clearer as a working prototype?
- What part of your product currently suffers from "everything is shouting" because AI made it easy to add more copy, features, and visual treatments?
- Which important capabilities in your product are hidden behind tabs, carousels, or secondary screens that most first-time users are unlikely to open?
- If AI can generate ten product solutions cheaply, what specific judgment criteria does your team use to decide which one is good enough to ship?
Bottom Line
AI makes output cheap, which means judgment gets expensive. For PMs, the practical shift is to rely less on static artifacts and more on working prototypes, stronger decomposition, and ruthless clarity about what the user should notice first.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Figma | Interface design tool | Marvin now uses it more as a scratchpad for layouts, states, and visual explanations to AI than as the final place where design work lives. |
| Codex | AI coding tool | Marvin uses it to turn rough concepts into working demos and prototypes quickly. |
| ChatGPT | General AI assistant | Marvin used it as a second agent to diagnose why another AI tool was failing and to help break a visual problem into parts. |
| Dialkit | Prototype control toolkit | It let Marvin expose and tune variables inside prototypes, such as grass length, wind, and motion behavior. |
| Shopify Sidekick | Shopify's AI assistant layer | It is central to Shopify's redesign work, especially around conversational control of store management tasks. |
Notable Quotes
"We're more of the editors now than ever before." (Marvin Schwaibold)
"AI has definitely raised the ceiling for all of us, but it's not a crafter." (Marvin Schwaibold)
"Prototypes are a million times more useful than any PRD document you can write." (Marvin Schwaibold)
"Use the native scroll on websites and don't hinder the speed at which a user wants to interact with the content of your website." (Marvin Schwaibold)
