Aaron Epstein

1. Key Themes

The Designer Becomes the Editor and Curator

Marvin argues the bottleneck in design has flipped from producing output to selecting among abundant output. The craft moves from hand-building every pixel to taste, judgment, and steering. He said: "We're more the editors, now than ever before, um, which is, I think, fun, that's great. I mean, that was always the role, but getting to the point of editing just took a lot of, like, handcrafted design" [00:02:34]. He added that "the output isn't the issue anymore. It's us being extremely selective about what we go forward with" [00:01:40].

Figma Is Demoted from Canvas to Scratchpad

The tooling has shifted from Figma-as-source-of-truth to Figma-as-sketchpad feeding coding agents (Codex, Claude Code, terminal). "I'll use Figma as a scratchpad for some of the ideas I have, uh, but then, very quickly, that'll transcend into- into- into, like, real demos that I can use" [00:00:33]. He notes the toolchain is now "a world now that, uh, that in- includes tools like like like Codex, and and and Cloud Code, and and being in terminal a lot" [00:01:09].

Prototypes Replace PRDs as the Currency of Persuasion

At Shopify, a working prototype is the fastest way to win organizational buy-in, even with the CEO. "Prototypes are a million times more useful than any PRD document you can write. Those spread around the company like wildfire. And if Tobi sees a prototype that he likes, then we can go off and build that" [00:14:00]. He continued: "I can show up with a prototype, and get people's attention faster than I can if I have an idea and I have to, um, get a whole team rallied around an idea" [00:15:13].

Full-Stack Roles Are Collapsing Across Design, Engineering, and PM

Marvin states directly that role boundaries are dissolving, and that teams of one or two can do the work of many. "Designers are becoming more full-stack. As are engineers and PMs, it goes it goes it goes in every which way" [00:26:06]. And: "Right now, I'm seeing one-off armies or or teams of two going off and and and doing the work of many, um, because the tooling is just allowing us to to do that" [00:14:48]. Aaron echoed it: "there's no limitations now, right?" [00:26:13].

Build Tiny Tuning Tools: Expose the Underlying Parameters (DialKit)

A repeated pattern: when taste can't be communicated in chat, build a control panel that exposes the variables. "The back-and-forth with chat just doesn't work. Um, you need to be able to to to always expose those kind of those kind of values" [00:28:11]. Exposing structure also reveals AI mistakes: "AI had misinterpreted something five prompts ago, that I didn't know about anyway. DialKit kind of unpacked that for me" [00:25:11]. Aaron framed why this is new: "Previously, it wouldn't have been worth anybody's time to do that... Now, it's so easy to create new little software tools to be able to control these variables" [00:26:30].

Decompose Complexity, Then Feed the Agent Sub-Tasks

Output quality depends on the human's ability to break a problem apart. "As long as you understand the complexity of what it is that you're working on, and you're able to break that down into in- into a multitude of of almost like sub-tasks that then you can feed to the to the agent bit by bit" [00:02:13]. In the grass-shader project: "You have to be able to take a a concept... and break it down into all its little components. You have to be able to understand those components, and then you can put those components back together" [00:21:26].

Multimodal Communication Beats Text for Taste

Visual and structural inputs communicate design intent better than prose. Marvin drew colored boxes in Figma to explain interruptible motion: "I couldn't articulate it via via text or voice, but I could draw it out" [00:21:18]. For the shader he had ChatGPT render "a topographical map of the underlying structure" [00:23:20]. He also recommends feeding the AI references: "Send it a website where you really like the carousel animation, or give it a screenshot from from Savvy, or Cosmos" [00:29:29].

Craft Still Differentiates, and Small Details Signal Quality

Even in an AI era, intentional micro-detail signals the standard a team will hold. On the Collins bookshelf bounce animation: "the clients coming to the website are looking for those kind of Easter eggs. They want to know that if I hire this agency... what kind of Easter eggs are they going to put into our projects" [00:10:52]. The ambient-lighting gradient, inspired by wall-mounted TVs, made people "inclined to scroll into it" [00:07:25].

Design as a Strategic Mandate at a Founder-Led Company

Shopify's CEO is explicitly using design to paint the future. "Tobi... basically said, um, 'We're entering a new era, and that future needs to be designed'... 'I need to hire good designers who can help me paint a picture'" [00:11:48]. Marvin says design is "very much helping shape the narrative of where the company is going" [00:12:15].

AI as an Infinitely Patient Collaborator and Teacher

Marvin frames AI as a patient teammate and a learning tool, reducing the social cost of iteration. "Codex won't get mad at me if I need 91 approaches at getting the animation correct. If I had done that with my engineer, they would probably flip out" [00:29:25]. He also uses it to learn by copying: "I'm trying to imitate this this carousel animation... because now I have an infinitely patient teacher" [00:29:57].

2. Contrarian Perspectives

The Process Hasn't Changed. Only the Tooling Has

Most observers claim AI is rewriting design from scratch. Marvin says the cycle is the same, and the real shift is where iteration happens. "I feel like the process itself hasn't really changed. It's still extremely iterative, uh, and- and sequenced out" [00:00:33]. The leverage is in speed to a testable artifact, not in a new methodology.

AI Raises the Ceiling but Cannot Crafts Taste

Against the narrative that AI makes designers obsolete, Marvin argues taste is the scarce asset. "AI has definitely raised raised the ceiling for all of us, but it's it's it's it's not a crafter... It can write code, but it doesn't know what looks good and what looks bad, and that's for me to decide" [00:03:02]. Implication: the value of judgment rises as generation cost falls.

Hand-Craft Still Wins Brand-Critical Moments, and Figma Remains Right for "Too Expensive to Test in Code" Work

Counter to "code-first for everything," Marvin explains that for the Collins project, highly detailed Figma was the better tool when code prototypes would burn engineer time: "These are too heavy and too expensive to test in code. We don't want to test this, um, in code right now, because it would require an engineer to take a look at it, and build it for us. So, we test it all in Figma" [00:19:06]. The nuance is that tool choice depends on cost of the test, which AI is shifting but not eliminating.

Designers Should Prototype in a Sandbox, Not Ship to Production

Despite the "everyone is full-stack" rhetoric, Marvin keeps a hard line: prototype code stays out of the main repo. "That's not production code, and that wouldn't be allowed back into the main GitHub repo of Shopify" [00:18:24]. The empowerment is about visualizing and persuading, not bypassing engineering rigor.

Agent Failure Often Originates in the Human's Missing Mental Model

When Codex produced bad grass, the fix wasn't a better prompt but a better decomposition: the AI didn't know the terrain was hills, not a flat layer. "It doesn't understand that there's it's not a flat layer underneath, these are little hills" [00:23:38]. Marvin's view is that you must "know what you're doing" because "you can use these tools to make anything that you want" [00:21:26].

3. Companies Identified

Shopify

Commerce platform (about 20 years old) with millions of merchants. Marvin leads its product design studio. Mentioned as an organization where design is being repositioned as strategic. "It's a founder-led, engineering-led organization. Um, and design hasn't always played the role that it's playing now. Design is very much helping shape the narrative of where the company is going" [00:12:15]. Notable initiatives: the admin redesign, the "green app" mobile app, and Sidekick.

Sidekick (Shopify)

Shopify's AI infrastructure layer for managing a store through chat. Mentioned as the key surface of the new era. "Sidekick, which is our, uh, AI infrastructure layers, that basically manage our entire admin" [00:03:30]. The vision: "a world where a a merchant can spend a lot more time on the home surface, and use Shopify as as a dashboard to delegate and interact with with Sidekick, and have Sidekick directly manipulate the store" [00:16:56].

Molly

Brooklyn design studio founded by Marvin and Jeytul, run about two years before being integrated into Shopify. Known for high-craft web and product design with a small client load. "We never took on more than two or three clients at a time to stay really focused" [00:06:36].

Collins

Branding agency and long-time Molly client. Marvin worked with them over a year on their website, including the "bookshelf" case-study section with bouncing, organic animations. "We worked with them for, I think, over a year, um, on this on this website" [00:09:00].

Apple and Google

Named as large clients of Molly. "We worked a lot with, uh, with with Collins, and Apple, and Google, uh, and also a lot of smaller clients, too" [00:06:26].

Spencer (illustrator) website

A Molly client project referenced as a contrast with Shopify's scale. "We were having fun designing websites for other studios, like Collins, um, or Spencer, the illustrator" [00:12:15].

Codex (OpenAI)

Marvin's primary coding agent for turning design ideas into prototypes. "I was working in Codex, and Codex built this out for me" [00:20:15].

Claude Code (Anthropic)

Named as a core tool in the new workflow. "Tools like like like Codex, and and and Cloud Code, and and being in terminal a lot" [00:01:09].

ChatGPT

Used as a second agent for meta-reasoning when Codex failed. "I would open up ChatGPT, and I gave it to another agent there. I explained that agent exactly what I was doing as well" [00:22:57], and for generating the topographical map of structure.

Figma

Now used as a scratchpad rather than the primary design environment. "Figma is the playground and the and the world we we were building everything in" [00:19:06] describes the previous era, while currently it is "more a scratchpad to, like, map out ideas" [00:20:56].

DialKit

Tool by Josh (see People) that exposes tunable parameters in a UI panel, installable with NPM into a project. "You can just run NPM run via via his website, uh, that he has a little prompt for it. You can just plug it into to whatever you're working on" [00:27:01]. Aaron observed: "which I guess is where the name DialKit comes from" [00:26:49].

LoveFrom

Design collective referenced via Mike Matas, whose portfolio carousel inspired a Marvin experiment. "Mike Matas, he's a design lead at, um, at at LoveFrom, and and his portfolio carousel works in this way" [00:27:26].

Savvy and Cosmos

Visual inspiration sources Marvin suggests screenshotting as references for AI. "Give it a screenshot from from Savvy, or Cosmos, of something that you that that you think looks really beautiful" [00:29:29].

Nstack

YC-style startup reviewed on the show: "the only AI workspace you'll need. One powerful agent, one intuitive desktop workspace." Marvin praised the desktop emulation ("you're basically just in their product already") but critiqued the lack of a focal point [00:31:18].

GitHub

Referenced as Shopify's main repository host, where prototype code is not permitted. "That wouldn't be allowed back into the main GitHub repo" [00:18:24].

4. People Identified

Marvin Schwaibold

Leads the product design studio at Shopify; formerly co-ran Molly in Brooklyn. A model of the "full-stack designer." Aaron: "I've been a big fan of your work for a long time" [00:03:22]. His own summary of the shift: "AI has definitely raised raised the ceiling for all of us" [00:03:02].

Host (Y Combinator). Contributed the synthesis that the pattern is "creating prototypes, and then creating, uh, little widgets that allow you to control different variables" [00:26:13] and that "a lot of all the projects... is just having setting an unreasonably high bar, and now we have the tools to actually be able to meet it much faster" [00:28:27].

Tobi (Tobias Lütke)

Shopify's founder and CEO. Mentioned as the one who set the design mandate and responds to prototypes. "Tobi... basically said, um, 'We're entering a new era, and that future needs to be designed'" [00:11:48]; "if Tobi sees a prototype that he likes, then we can go off and build that" [00:14:24].

Jeytul

Marvin's studio partner at Molly; credited with the ambient-lighting video-blur idea. "I think Jeytul, and this was pre-AI, um, basically duplicated the video, and had it, like, blurred out in the background" [00:07:31].

Jesper

Engineer who built the bookshelf bounce animation on the Collins site. "I would then hand it off to someone like Jesper, um, who would then work on like the bounce animation" [00:09:48]. Aaron highlighted him as "supremely talented, whose entire job on this project was to make the tab interaction" [00:10:21].

Michael Speicher

Collaborator with Marvin and Jeytul on the Collins project. "I think this is a mixture of me, um, Michael Speicher and Jeytul, working on this project together" [00:19:06].

Josh

Creator of DialKit; suggested exposing grass blade length at a lunch. "He's like, 'This looks pretty good, but you should look at the length of the grass blades.' I was like, 'Woah, okay, that's true.'" [00:24:37].

Mike Matas

Design lead at LoveFrom; his portfolio carousel is the reference for Marvin's collision-ripple experiment. "Inspired by, uh, Mike Matas... his portfolio carousel works in this way" [00:27:26].

Marvin's German graphic design professor (unnamed)

Source of the "diktatisch" critique: a poster should dictate where to look. "This poster doesn't dictate, it's not diktatisch, it means there's nothing in this poster right now that is drawing my attention" [00:32:16].

5. Operating Insights

Branch-and-Merge Prototyping in a Shared Demo

Shopify's design team aligns on a base demo, then everyone forks their own idea and merges back into the same link. "Once a demo is at a point where we kind of all align, everybody branches off, and builds a version of something that they think they that brings value. And then we all come back together, and it actually lives all inside the same the the same demo and the same link" [00:17:29]. A concrete tactic for parallel exploration without losing a shared artifact.

Keep a Dev-Panel Playground Inside the Real Product Shell

Rather than isolated mockups, the team builds in an admin playground with a dev UI panel, so ideas are tested in context with live-feeling data. "There's a dev UI panel inside here, and basically, we're we're building all of these the the design team is building all of these, um, changes and and demos and and and and ideas right into the interface here, so we can just test and build on top of it" [00:16:56]. This also keeps production code protected.

When the Agent Stalls, Escalate to a Second Model for Prompt Diagnosis

Marvin's debugging loop: screenshot the output plus the reference, hand both to a different model, and ask why the first agent isn't understanding. "I said, 'Why is the other agent not understanding what I'm trying to get to?' And that ChatGPT then helped me kind of like understand my own process" [00:22:57]. The suggestion: break the target into about six parts, specify light direction and shadows, and rebuild piece by piece.

Specify Interaction Requirements (Interruptibility) Explicitly

Animations built by agents often fail on responsiveness. Marvin caught that the interaction "wasn't interruptible... That's just not how you want to use this pattern. This has to be very responsive very quickly, otherwise, you'll get very annoyed with the interface very fast" [00:20:31], and fixed it with a simple colored-box Figma sketch. Specify behavior states, not just visuals.

Keep Client Load Tiny to Keep Quality High

From the Molly era: "We never took on more than two or three clients at a time to stay really focused," with weekly meetings and high-focus stints [00:06:36]. A transferable principle for studios and internal teams: limit concurrent commitments to protect craft.

6. Overlooked Insights

AI Hides Its Own Mistakes Several Prompts Deep

In passing, Marvin revealed that the terrain map Codex implemented was "completely wrong and s- and and skewed" and that "AI had misinterpreted something five prompts ago, that I didn't know about anyway" [00:25:11]. This is significant: iterative prompting accumulates silent compounding error, and the fix is instrumentation. Expose the intermediate state, which is exactly what DialKit-style panels do. It suggests a whole category of debugging and observability tooling for AI-built creative work.

Design and Automation Are Converging Within a Single Product

Marvin noted that Shopify's two design efforts are "two almost controversial, paradoxical uh, sectors of the company, but they work really well together... you're building really beautiful, new UI, but you're also using new UI to basically generate um, AI infra, and AI workflows" [00:05:16]. The deeper implication is the "bring UI to the user, and into the context" concept, where conversational agents surface just-in-time interface rather than users navigating a dashboard [00:16:30]. That shifts the unit of design from the screen to the suggestion-plus-diff, a subtle but major change in what product design means for SaaS.