Design standards

How Stripe puts design intent into AI workflows

Katie Dill describes Stripe's move from design documentation to a CLI with templates and flows, while keeping editing and human judgment part of the finished product.

Begin with a point of view

A design system can supply components without explaining how a product should behave. In her talk on Lenny's Podcast, Katie Dill describes a different task for Stripe's system: carrying the decisions that designers once supplied in person into workflows where they may never enter the room. The old system scaled consistency, she says, but "the new system needs to scale intent." That means specifying how the parts fit together, then examining the experience they produce.

Dill introduces the problem through architecture, describing how modernism's principles became repeated patterns during the building boom after World War II, while attention to context and purpose grew thinner. She sees echoes in the AI building boom, where an interface can take shape without conveying much about the business it serves. A Korean barbecue website supplies her example: its appearance, she says, gives little indication of what the restaurant sells or the setting in which people would use it.

An apparently finished interface also invites what Dill calls "the temptation of done." She compares it with overlooking a microwave burrito's flaws because it becomes lunch so quickly; rounded corners and shadows can create a similar feeling of completion before anyone establishes whether the interface solves the problem or differentiates the product. That speed also raises a question of responsibility, as work that feels disposable still needs someone to maintain it.

"You have to have a point of view," Dill says, beginning with what a brand means to its users and what they care about. At Stripe, she describes optimism appearing in colors, writing and products intended to support entrepreneurs. The question of how to apply those standards surfaced during an advertisement review, when a colleague asked what quality should mean for something made with AI. Dill's answer was to judge the output that users experience, so the discussion ended with specific improvements to the advertisement.

Put standards where building happens

Dill describes developing judgment by noticing what users need and examining products, while drawing on interests beyond the immediate work. Distributing that judgment across builders and agents requires making it explicit: interfaces can now be built without a designer present, so the system needs to carry decisions that once emerged in conversation. As she puts it, "the object of design is no longer the screen. It is the system itself."

Stripe initially connected an MCP to its design documentation, but Dill says the results were neither specific nor consistent enough. The team then built a CLI on its design system, putting documentation where builders work and supplying it when the task needs it. Full templates and flows sit alongside components, giving the AI guidance about intended behavior and how the product comes together. The written standard becomes part of the means of producing the interface.

Related posts approach the surrounding system through engineering ownership. itamar_mar on X describes agents across the development lifecycle as a step toward a software factory, while arguing that the harder work lies in building and owning the system around them. fjzeit on X connects accountability with being able to describe, reason about, defend and change a software system. Their concern with ownership sits alongside Dill's concern with design standards: each places responsibility beyond obtaining an agent's immediate output.

Examine the whole experience

The encoded system remains a baseline in Dill's account. "Refuse to confuse done with good," she says, explaining that product development once filtered ideas at many stages, beginning with deciding which projects a team could staff. Faster building moves more of that filtering until after something exists, so editing becomes a distinct responsibility. Someone must examine the complete experience as a user would, checking whether its parts form a coherent journey and whether the work solves the problem.

The event's opening animation shows what that scrutiny involved. Dill's team began with a 3D model and used AI to help animate the scene, but she found the initial movement slightly wrong and wanted to see more of the setting. Stefan and the team continued through 56 iterations, with Dill pointing to changes in realism and movement that conveyed greater care. In her telling, AI expanded what the team could attempt, while reaching the accepted version still required sustained examination.

Keep judgment in the iteration

Dill's final recommendation extends that exploration to creativity and artistry. She questions whether familiar chats, charts and CLIs exhaust the possibilities for interaction, pointing to online work in data visualization, restaurant websites and art. At Stripe, the cover for Built to Grow began with a human marbler's work, which AI then helped refine by adjusting colors and lines. Human judgment guided the work, while AI extended the team's treatment of its details.

Her practical advice connects those standards to the material a builder gives the model and the scrutiny applied afterward. A restaurant website prompt can explain what the business believes and cares about, supported by the source material behind the team's standards. Dill then recommends pressing beyond the initial output and using adversarial agents to critique it, while retaining the builder's responsibility to push further. "Always push a step further," she says.

The harder work, Dill says, is cultural: teams need room to explore alongside the instruction to use AI. "Protect the strange," she says, before returning to architecture through Gothic buildings whose individual details show the hand of their makers. She connects that care with products that anticipate users' needs. From the restaurant site to Stripe's design CLI, the revised animation and the marbled cover, her examples trace how a builder's intent reaches the finished experience.