# AI Tip #1: Design for the metric, not the dopamine hit > Source: > Published: 2026-10-27 14:10:35+00:00 **Pro tip:** Tell the model which business metric the design should move before you ask it for mocks. ## Context AI is moving so fast. Understanding the newest ideas, and how each one applies to your own work, is how you stay equipped. It lets you keep reinventing your design process from first principles, so the work gets more thorough and more creative. That’s why I’m starting this series. Each week I take one idea from something I read and turn it into a move you can use. This week there are two: Felix Rieseberg’s tutorial on being more ambitious with models, and Katie Dill’s talk on zombie UI. First, where I am today. My day-to-day is mostly exploration. I collect transcripts with Granola to understand what users are asking for. Then my own skills push the UI in many different directions, I choose a creative direction, and I take it back to users. The design process itself hasn’t changed much. What’s changed is that I can build a ton of mocks quickly, narrow them down, and still try the crazy ideas. That’s speed. In [his tutorial on being more ambitious with models](https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/), Felix Rieseberg, who works on Claude, names it as the first of four kinds of ambition, and the one that needs the least creativity. The next one up is altitude: you ask for the goal beyond the goal. At the start of last year he was still asking Claude to fix individual crashes. Eventually he asked it to keep his team’s apps crash-free. For designers, moving up a level isn’t just mocking the whole product experience. It’s thinking like a business owner: how do you create a product that moves the business outcome? Sometimes that’s the front end, sometimes the project management, sometimes consolidating the IA. What matters is that it moves the business metric. There’s a trap on the way up, which is zombie UI. You feel like you’ve finished the work, but you actually haven’t. Stripe’s Katie Dill uses the name in [her Lenny and Friends Summit talk](https://www.youtube.com/watch?v=GLvFTMtw4Jk). Behind it is what she calls the temptation of done. Her example is the microwave burrito she has for lunch all too often, 90 seconds from freezer to plate and nearly inedible, which she forgives because it’s so fast. ## Why does it work? Understanding the business metric tied to your design steers you away from making fast UI just to satisfy the dopamine hit. You make real UI decisions that come from understanding users, and those decisions tie back to the business outcome. A metric gives the model Felix’s altitude and turns Katie’s advice, experience the work the way a user would, into a check you can run on every screen. **Zombie UI check, the prompt I’d start with:** ``` Before we call this mock done, walk through it as [user type], from [entry point] to [the moment that moves the metric]. At every screen, say what the user is trying to do, what could stop them, and whether the screen helps or hurts [business metric]. List anything that looks finished on its own but breaks the journey. Don't call it done until that list is empty or I've signed off on each item. ``` ### How to do it Felix builds a three-rung ladder around a quarterly review deck. Here it is rebuilt for a design crit. **Rung 1: start from the metric.** ``` We want [product or feature] to move [business metric] from [current value] to [target] by [date]. Here's what I have: [analytics, research notes, call transcripts, current mocks]. Before you design anything, tell me what drives that metric today, which user problems in this material sit behind it, and which parts of the flow, front end and IA would move it most. ``` **Rung 2: decisions, same thread.** ``` Given everything above, pick the three design decisions in this flow that matter most right now. For each, sketch two or three directions with the case for and against, judged by what each would do to [business metric]. ``` **Rung 3: what crit missed.** ``` Here are the problems we found after our last [number] launches: [bugs, support tickets, metric drops, user complaints]. Make a table with one row per problem and three columns: the problem, whether our design crit caught it (yes, no, partly), and what check would have caught it earlier. Then turn that last column into a checklist I can run on next week's designs before they go to crit. ``` Then run that checklist beside your usual crit for four weeks and compare what each one catches. ### When should I use this? Felix’s tell is meetings. If people will meet to discuss the work, altitude fits. A crit counts. He also says it works less well for creative work, so try it on signup or checkout before a brand exploration. **The mistake to expect:** running Rung 1 blind. Before Claude could keep his apps crash-free, Felix had to let it fetch crash reports, map app versions to git commits, and reproduce fixes on Windows, macOS and Linux machines. Your version is the dashboard and the research. #### Depth and range are Felix’s other two kinds of ambition Depth is working the way a field-leading expert would. Range is drawing on fields the task doesn’t seem to need. Both apply to design work. ## How do you push a design task deeper? Start with instruments. Told to act like a professional, Felix says, a model often switches its vocabulary and keeps its methods, so you get markdown tables where a data scientist would reach for a database and a statistical model. “Act like a senior designer” fails the same way, with senior-sounding words on the same mock. Ask for the method instead, like a task-based usability script or a screen rendered at real breakpoints. Then pick goals that sound unreasonable and tell the model they’re reasonable. Need a page to load in 5 seconds instead of 10? Ask for 200 milliseconds, and when the model balks, tell it to keep going. For a flow, ask for the [number]-step signup in one step. Challenge what the team already believes. Felix’s example is a company convinced that churn is an onboarding problem, so he asks the model to reproduce that belief from the raw data. Designers say “users drop off at onboarding” just as easily, and session data can test that before anyone redesigns the first screen. Then push for the earliest churn signal nobody knows about yet, tested on data the model hasn’t seen. And don’t let the model grade its own work. Felix wants an outside judge, ideally a tool the model can run for instant feedback. For a screen, that could be an accessibility checker or a round with real users. **The prompt** (Felix’s “deeper, not bigger” prompt, rewritten for design): ``` I want a deeper version of this design task, not a bigger one. Before you start: (1) tell me what a top [product designer / UX researcher] would do here that I haven't asked for; (2) name the raw material they'd start from instead of my summary, like [session recordings, support tickets, analytics events]; (3) name the methods and tools they'd use; (4) say what, other than you or me, would show the design works. Then rewrite my request to ask for exactly that. My request: [request]. ``` ## How do you bring other fields into a design? In early research and exploration, it’s easy to get caught up in the timeline and not explore enough. Range is Felix’s answer. His first trick comes from [Tom Kelley and Jonathan Littman’s The Art of Innovation](https://www.amazon.com/dp/0385499841?lv=shuf&channelId=500&plpRedirect=mhFallback): replace the nouns with verbs. A quarterly review stops being a deck and becomes getting busy people to believe a story about cause and effect, and to decide under uncertainty, in 30 minutes. Onboarding stops being a set of screens and becomes getting a new user to trust the product enough to [first key action] in [time]. A verb like teaching points you to a teacher. Then Felix assembles the crew, which for his deck means an analyst, a journalist and an editor, plus a short seller and a forensic accountant as a red team. He warns that role-play makes the transcript weirder without improving the work. Don’t ask the model to write like a short seller. Ask it to do the research a short seller would do. **The prompt:** ``` Rewrite [the thing I'm designing] as verbs: what it has to get [user type] to do, believe or decide, and how fast. List the fields best at those verbs. For each field, describe the work that expert would do on this design, step by step, then do that work. Don't write in their voice. ``` #### Ten directions when three would do Felix ends on self-efficacy: we’ve always been held back by what we believe is possible. Locke and Latham went through 35 years of goal-setting research and found that people given specific hard goals do better than people told to try their best. In 1939 George Dantzig showed up late to a statistics class at Berkeley, copied two problems off the blackboard and handed in solutions, thinking they were homework. Both were unsolved problems in statistics. A brief that asks for ten directions on a screen instead of three is a specific hard goal, for the model and for you. #### Katie Dill’s warning starts with post-war buildings After World War II came the biggest building boom in history, and builders copied 1920s modernism again and again without its founders’ intent, until only generic patterns were left. Katie calls the results zombie buildings. She hears echoes of that boom in the AI building boom, where teams of three do what teams of 30 used to. She gives three watch points. LLMs are good at the most probable answer, so a Korean barbecue restaurant ends up with a website you’d guess sells software. The second is the temptation of done, the burrito. The iteration cycle is so much tighter now. A mock used to take a few days, sometimes two weeks. Now everything’s done in a day, so feedback comes back much quicker. That speed feeds her third watch point. Work this quick feels disposable and gets treated that way, and nobody stops to ask who’s maintaining it. ## Katie’s four recommendations #### **Have a point of view.** If you don’t, she says, AI will give you one, and it’ll be generic and drawn from the past. Stripe’s is optimism, which she says runs through its colors, its writing and the products it ships for entrepreneurs. In a recent crit on an ad, a cross-functional partner asked what the quality standard is for something made with AI. Her answer: users don’t care how it was made, only whether it’s good. The team walked out with 17 bullet points of fixes (a bit of frost here and there, softer edges, slightly smaller bubbles), and “Pepsi bubbling” is now the office verb for that kind of meticulous craft. The bar is one level deeper than the customer can see, not perfection. Write a few lines on who your product is to its users and what they care about, and paste them above every design prompt. #### **Encode your standards into the machine.** Stripe started with an MCP over its design docs, and three people could type the same prompt and get three different results. It now uses a CLI built on the design system. It feeds the model the docs at the right moment, to avoid context rot, and holds full templates and flows, not only components. In Katie’s framing, the old design system scaled consistency and the new one has to scale intent. Her model is Gutenberg, who didn’t stop at 26 lowercase and 26 uppercase letters. He made 290 characters, with different widths, abbreviations and ligatures, so fully justified lines wouldn’t open rivers of white space. She follows it with Christopher Alexander: “A system can satisfy every rule and still be dead.” Put whole flows in whatever your agents read, each with a sentence on why it’s shaped that way. #### **Refuse to confuse done with good.** The quality filter used to run through the whole process, when a team had 20 ideas and staff for one. Now it can build all 20 in a week, so the filtering happens after the build, when no is harder to say. Katie wants an editor there, someone who experiences the work end to end the way a user would. Her example is this summit’s opening animation. Stefan on her team took a 3D model of the scene into AI to animate it, and the first version moved a little off. Fifty-six iterations later it had more realism and a bit of life. Run the zombie UI check above, and name one person who’s allowed to say no. #### **Unleash creativity and artistry.** Her favorite. Improve the inputs with what you believe and what good means to you, stress the outputs instead of falling for the burrito, and let adversarial agents bang the work up a little. If you lead a team, she says, protect the strange. AI lowered the cost to create, so spend some of the savings on something special. It can raise the ceiling, not just the floor. **The prompt:** ``` Here's what [brand] believes and what good looks like to us: [beliefs, references, source material]. Give me [number] directions for [screen or flow] that only we could make. Then run a second pass that finds where each one looks like every other [product category], and rank what survives by its likely effect on [business metric]. ``` ### Too long, didn’t read - Rung 1 is the metric, Rung 2 the decisions, Rung 3 what crit missed. - To go deeper, ask for the methods experts use and let something other than the model judge the result. - Turn nouns into verbs to find the fields a design needs, then ask for each expert’s work, never their voice. - Zombie UI is work you feel you’ve finished but actually haven’t. - Katie’s fixes: a point of view, standards built into the tools, an editor who uses it like a user, and room for the strange. ### **Try it this week:** - **Ten minutes tonight:** run Rung 1 on whatever you’re designing now. - **One hour this week:** all three rungs, then the zombie UI check, on your next crit design. Reply with your metric and what Rung 3 found. If you want the longer read, start with [6 levels to becoming AI native](https://johnmaartifacts.substack.com/p/6-levels-to-becoming-ai-native-find). P.S. Forward this to the teammate who ships a prototype a day and never opens the dashboard. — John