When AI makes design output cheap, selection becomes the job

1. Designers as editors — ▶ 1:40

   The design process is unchanged, still iterative and sequenced. The tooling has moved from the Figma canvas to Codex and Claude Code in the terminal, with output that can be tested on a phone or computer almost immediately. Because output is no longer the constraint, Schwaibold says the discipline is being extremely selective about what moves forward. AI raised the ceiling, but it is not a crafter. It can write code he cannot handwrite, yet it does not know what looks good and what looks bad. It works best when complex work is broken into subtasks fed to the agent bit by bit.

2. Admin redesign meets Sidekick — ▶ 4:44

   Shopify's design initiative has two tracks. One redesigns the admin, starting in Figma and ending in polished web interfaces. The other, the Sidekick team, builds the AI layer that lets merchants manage and automate their store through chat. Schwaibold calls them almost paradoxical but complementary, since new UI is used to generate AI workflows. In one prototype, Sidekick suggestions on the home feed, such as rewriting a product description, start a conversation and directly change the admin. The idea is to bring UI to the user in context. He frames a merchant delegating from the home surface as a possible future, not current reality.

3. Craft as a sales signal — ▶ 10:19

   At his former studio Molly, Schwaibold paired designers with engineers on details like the bounce of a bookshelf of case studies on the Collins website. Each book had to land organically or the idea of a living bookshelf would be diluted. His explanation for the effort is commercial. Clients of a branding studio look for these Easter eggs as evidence of the care they would receive. An agency that sweats an interaction might also build a system behind a color palette rather than just picking five colors. Detail works as a demonstration of capability.

4. Prototypes beat documents — ▶ 11:47

   Tobi's mandate to Schwaibold's team, as he recounts it, was: "We're entering a new era and that future needs to be designed." Design now sits in the foreground of a founder-led, engineering-led company. AI lets designers turn Figma work into tangible prototypes quickly. Schwaibold says these are a million times more useful than any PRD and spread through the company like wildfire. If Tobi likes one, the team builds it. He describes one-person or two-person teams doing the work of many, and says arriving with a prototype gets attention faster than rallying a team around an idea.

5. Sketching for the agent — ▶ 19:43

   Codex built Schwaibold's 2x3 video grid, where tapping opens a video and dragging down closes it. But the interaction was not interruptible, so he had to wait for one asset to settle before tapping another. He could not articulate the fix in text or voice, so he drew red, blue and green boxes in Figma showing a swipe-away interrupted by a new tap. Codex understood the sketch. Figma is now a scratch pad for mapping ideas to hand to the agent, because it is easier for him to decompile and, he argues, for the model to read visually.

6. Decompose before prompting — ▶ 22:16

   Turning a video of real grass into a web shader failed at first. Codex produced something that looked nothing like the target. Schwaibold used ChatGPT as a second agent to diagnose why. It advised him to explain where the light comes from, which shadows apply, and to split the image into six parts. He then had it render a topographical map of the underlying structure. The model had treated the ground as a flat layer, when it was little hills. His conclusion is that the tools can make anything, but only if you can break a concept into components, understand the commands and reassemble them.

7. Expose the controls — ▶ 24:40

   After lunch advice from Josh about blade length, Schwaibold asked Codex how many grass-blade variations existed. It reported about 217,000, differing only by roughly 0.7-0.8. They added a DialKit toggle to see it directly. Exposing the structure also revealed that his terrain map was skewed from a misreading five prompts earlier. His carousel, inspired by Mike Matas, took 91 prompts and relied on exposed parameters like pull boost. He says he knows when the feel is off, but back-and-forth chat cannot tune it. The host adds that such control widgets were previously not worth building.

8. The infinitely patient teacher — ▶ 29:17

   Schwaibold treats AI as an "infinitely patient" collaborator. Codex will not get angry at 91 attempts at an animation, whereas an engineer might. His advice to non-designers is to tinker, not give up when results disappoint, and feed it references such as sites with carousels he likes, screenshots and color palettes. For designers his trick is to copy: he imitates a long-standing carousel animation on weekends, without bothering engineers, to look under the hood. He says this teaches him to be a better manager and collaborator, a claim based on his own experience.

9. One job per page — ▶ 31:34

   Reviewing Nstack, Schwaibold says nothing on the page draws the eye first, citing a typography professor's critique of posters that lack a focal point. The host adds that there should be one primary goal or KPI, and here multiple identical buttons compete. Schwaibold calls the result "the opposite of progressively disclosing information." At Skill Sync he would reduce features to a three-up of top capabilities, each leading to more depth, and let the page run longer. Instaplay shows a related failure. Visitors went for the game screenshots and missed the prompt box that was the product's core, while sign-up came before they saw a created game.

10. Cheap animation, no meaning — ▶ 38:26

    The host warns that AI makes it easy to produce animations and images no human would bother to program, such as a dot traveling between corners, but they do not necessarily add understanding. Schwaibold agrees that curation needs the right skill sets. AI can help assemble a whole website, but abstracting features into illustrations may justify an illustrator who specializes in information graphics. His example is a Gemini animation of braided strands forming a helix, built by talented people over a long time and then translated into a scrollable web animation. It stayed in his memory and explained the model's capabilities.

11. Restraint is the expensive part — ▶ 45:33

    Mintlify earns praise as reductive in the right ways. It has a clean headline, a prominent get-started button, a subtle background graphic and a credible logo wall. Schwaibold still faults small card headlines and a reused braid animation that loses its magic. He structures homepages so each section is the first chapter of a child page, ordered by importance because scroll depth drops off. The host notes that spacious pages take more work, since choosing what to cut is harder than adding. Schwaibold agrees. Every team wants its features included, and deciding what to reduce is the designer's job.

Marvin Schwaibold's central claim is that AI has moved the bottleneck in design from production to judgment. The process itself, he says, is unchanged. It is still iterative and sequenced. What changed is the tooling. He now works in Codex and Claude Code in the terminal, using Figma as a scratch pad, and gets to something testable on his phone far faster than before. In his telling, designers were always editors. The difference is that hand-crafted Figma work used to stand between the idea and the editing. With output no longer the issue, the work is being selective about what moves forward.

He is blunt about the limits of the arrangement. AI can write code he cannot handwrite, but it does not know what looks good and what looks bad. So the human supplies the standard and the structure. His grass-shader project shows how. Codex's first attempt at recreating a video of real grass looked nothing like it. He used ChatGPT as a second agent to work out why, and it advised him to explain the light and shadows and to split the image into six parts. He also had it render a topographical map. The model had assumed a flat layer, when the ground was little hills. His lesson is that the tools can make anything, but only if you can break a concept into components and put them back together.

A second mechanism is visibility. Schwaibold says chat is a poor medium for feel. His carousel took 91 prompts, and he knows when it feels off, but he cannot fix it by talking. So he exposes the values, such as pull boost, as controls. The grass project shows a further benefit. When a colleague named Josh prompted a look at blade length, Codex reported about 217,000 variations differing by only roughly 0.7-0.8, and a DialKit toggle made that visible. The same exposure revealed that his terrain map had been wrong since a misreading five prompts earlier. The editorial inference is that judgment needs an inspectable system. A slider turns vague dissatisfaction into something that can be adjusted.

The channel matters as well. Codex built a video grid whose interaction was not interruptible, and he could not describe the fix in text or voice. He drew three colored boxes in Figma, and the model understood. Interruptibility recurs in his website critiques. He calls scroll-jacking a disaster because people's thought patterns are interruptible, and says a site should not force a user through a scroll pace the designer imagines. He treats Apple as proof of the principle. In both cases the interface should yield to the person using it.

Inside Shopify, he says the consequence of cheap output is persuasion. Tobi's mandate was that the future needs to be designed, and Schwaibold's team answers with prototypes, which he says are a million times more useful than a PRD. They spread through the company, and if Tobi likes one, the team builds it. He describes teams of two doing the work of many, where at Molly there were a thousand counterparts to coordinate. The Sidekick prototypes illustrate the ambition. Suggestions on the home feed start a conversation and directly change the admin, so the merchant never clicks through the interface.

The evidence for all this deserves scrutiny. It comes almost entirely from Schwaibold's own workflow. The Shopify prototypes live in a playground that he says is not production code and would not be allowed into the main repository. So the speed gains he describes apply to exploration and persuasion. Nothing in the discussion says how much of a prototype survives the trip to production, or what that handoff costs. His claim that designers, engineers and PMs are all becoming more full stack is asserted, not measured.

The website reviews show what happens when generation outruns judgment. Nstack puts a sidebar, chat history, competing buttons and its main message on one screen, which Schwaibold calls the opposite of progressive disclosure. Skill Sync spreads features across tabs and animated graphics, and the host notes that users rarely open every tab and that AI makes it cheap to produce animations that explain nothing. Instaplay's visitors missed the prompt box that was the whole product. The common prescription is one primary goal per page, a three-up of top features, and a longer page instead of hidden content.

Mintlify is the positive case, and the discussion makes a point about cost. The page is clean because someone cut things, and the host observes that restraint takes more time than piling everything in. Schwaibold says the difficulty is organizational: every team wants its feature on the page, and reducing is the designer's job. Even Mintlify errs by reusing its braid animation until it loses its magic. His own homepage rule gives the discipline a mechanical basis. Sections follow the order of importance because users drop off as they scroll.

There is a tension in his account of craft. At Molly he explained the hand-built bounce of a bookshelf as proof of care, the kind of Easter egg a branding client reads as evidence of systematic thinking. If AI makes polish cheap, as he argues, visible polish may signal less. His own answer is implicit. He suggests an LLM could help render a 3D moon as a show-stopper on a space company's page, but the human decides where the one or two such moments belong. What remains scarce is the decision about where attention should go.

That leaves the question of how the skill is acquired. Schwaibold's advice is to tinker, copy references, and use an infinitely patient AI to decompile work you admire. Yet the capacity he treats as essential, knowing what looks good and breaking a hard problem into parts, is exactly what the tool lacks. Whether imitation with AI builds that capacity for newcomers, or only amplifies people who already have it, is the part his own examples cannot settle.