别做出僵尸界面：Katie Dill 谈如何在 AI 造物潮里放大意图和手艺

Dill opens with architecture. After World War II the US saw the largest building boom in history, and builders reached for the modernism of the 1920s: simple geometry, bare surfaces, monochrome palettes and no adornment. The founders of modernism had real intent, with a point of view behind the Villa Savoye and principles behind the Bauhaus.

In the rush to build, the style was copied again and again without that intentionality. The thinking got thinner until only generic patterns were left, ill-suited to their context. Banks that once signaled trust, reliability and security gave way to buildings that differentiate nothing and show no care for the user, the surroundings or the brand.

Dill says we are in another building boom now, in which "everybody in their mother can build" and teams of three can do what once took thirty. She treats the parallels with the post-war era as watch points to navigate, and she names three of them in the next stretch of the talk.

The first watch point is that LLMs are best at the most probable answer. That means they tell you what has been popular or worked in the past, and are less good at what is original to you. Dill shows a website and asks the audience what the brand sells; the answer is Korean barbecue, though the site has no character and no sense of the user or the context of use.

The second is the temptation of done. She admits to eating microwave burritos for lunch, overlooking serious flaws because she is enamored by the 90-second speed. AI works the same way: type a sentence and you get an interface with nice corners and drop shadows, but an apparently polished state can mislead about whether it solves the problem or differentiates.

The third is that work this quick feels disposable and gets treated that way, so people let responsibility for their decisions slide and ignore the long term, such as who will maintain it. Combined, these produce "zombie UI" that is monotonous, vacant or uncared for, which matters because we spend half our waking lives looking at screens.

Against that, Dill points to small signs of care: the animation and life in the Grokbot, the current date shown in a calendar tab, and a link agent wallet that anticipates problems an agent could hit while buying online and builds troubleshooting in. Such details show the builder thought about the user and give a brand its soul.

Dill's first recommendation is to have a point of view, because if you don't, AI will give you one, likely generic and backward-looking. Teams need to define what the brand is, what it is for users, who it wants to be to them and what they care about. This forms the basis of standards, and matters more now that building is distributed and ownership is diffuse.

At Stripe, the value of optimism is put into every detail: the colors, the writing, the subjects and the products meant to support entrepreneurs. The cost of missing alignment showed up in a design crit of an ad combining the Stripe parallelogram with users' visuals. A cross-functional partner who wanted to ship fast asked what the quality standard is for something made with AI.

Dill's answer was that how it was made doesn't matter to users; whether it is good does. The discussion produced alignment and 17 improvements, such as a little frost here, softer edges and slightly smaller bubbles. Stripe now uses "Pepsi bubbling" as a verb for meticulous craft, aiming not for perfection but to go one level deeper than the customer could see.

To build a point of view, she says, get really good at noticing: what users need and want rather than just what they say, and what signals good, great or meh in products and in analogous fields like art and science. That sense of taste is what gives you standards to scale.

The second recommendation is to encode standards into the machine. Humans will not be part of all decisions going forward, with interfaces built without a designer in the room, agents fixing problems overnight and generative UI built in real time. In the past a designer filled in the blanks so nothing had to be written down; now decisions must be easy to define, which is why design systems are having a moment again.

Dill says the object of design is no longer the screen but the system itself. She cites Gutenberg, who built not just 26 lowercase and 26 uppercase letters but 290 characters with varied widths, abbreviations and ligatures, so justified text avoided rivers and lakes of white space. The system was extensible and opinionated enough that machine-made pages could feel handmade.

Stripe first built an MCP that understood its design documentation, but results weren't specific enough, and three people with the same prompt got three different results. It then built a CLI on top of its design system, acting as a harness that makes the AI far more obedient, consuming documentation at the right time to avoid context rot.

The difference from older design systems is that it contains full templates and flows rather than only components, so the system knows how the product is supposed to behave. Dill calls it far more opinionated than before, but only a baseline, quoting Christopher Alexander that a system can satisfy every rule and still be dead.

The third recommendation is to refuse to confuse done with good. Quality filtering used to be built into every stage of product development: with 20 ideas and staff for one, teams poked, prodded and pruned along the way. Now 20 ideas can be built in a week, which lets teams react to the real thing instead of sitting in meetings about hypothetical products.

The catch is that filtering now has to happen after the build, when it is much harder to say no. Dill says this makes the role of an editor ultra important and asks who in the audience's organization looks end to end at whether the team is building a whole. People must unlearn that built means done and done means good.

Editing means experiencing the work as a user would: does it solve the problem, match how the user thinks and hold together as a journey. Many things look good in isolation but feel disconnected when assembled. Quoting Nabeel Qureshi, Dill says great art lies in unexpected details, deeper meaning and continuing themes, areas where AI is weak, and an editor takes accountability for every decision, in effect every pixel.

Her example is the opening animation for the event itself. Designer Stefan and team started from a 3D model and used AI to create animations. The first version was cool but off pixel by pixel, a little jolted, so he iterated 56 times to reach something with realism and life. Without AI he might not have built such a complex scene or tried so many viewpoints, though it took scrutiny and was never one-and-done.

The fourth recommendation is to unleash creativity and artistry. AI can manufacture monotony, but Dill calls it the greatest creative catalyst we've ever had, and differentiation will matter more as everyone builds. She argues chats, charts and CLIs are not the epitome of interaction, and that we can now talk to computers, so it is time to invent new interfaces and aesthetics, as multi-touch and the synthesizer once opened new possibilities.

She sees evidence online of more interesting data visualizations, restaurant websites with personality and people painting with AI. At Stripe, the recent cover for its Built to Grow publication started with human iterations by an artist, then AI fine-tuned the colors and lines, taking human judgment and helping scale it.

To make AI a stronger creative partner, she suggests improving inputs with specific prompts that state what you believe and care about, and adding source material that builds your standards. Then stress the outputs, pushing a step past the burrito dilemma, and use adversarial agents to critique, while still pushing for better yourself.

The harder part is cultural, since cookie-cutter patterns are safe and cozy. Leaders should not merely tell teams to use AI but give room to explore and "protect the strange", spending some of the saved cost on something special. She closes with John Ruskin's praise of Gothic architecture, where no two columns were alike and each detail showed the maker's hand, and says users want products with that kind of care.