Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
E Bufar runs design at Y Combinator. In a recent Design Review, she said something that made me stop and think for a long time.
"I realized that I think much faster than I type. So I'd rather talk to my computer."
These days, she barely touches a keyboard. She hits the Function key, streams her thoughts out loud, describes what she wants, and an AI agent builds it. She's using a YC portfolio company called Aqua — purpose-built to capture her spoken words and turn them into executable instructions.
This isn't a tech gimmick. Over the past few months, she led design for three full projects — Paxel, Sodazine, and Startup School — covering everything from visual identity to product interaction, social media assets, and event tickets. Almost all of it was done through voice and Claude.
I'm not here to tell you how cool the tool is. I want to talk about the design thinking behind this workflow — thinking that is quietly rewriting the rules for how startups build products and brands.
She's Not Using AI — She's Building Tools for Herself
While designing the Paxel website, E ran into a specific problem: she wanted a dithering visual effect on the page, using an open-source shader from Paper.design. Claude implemented it, but the default parameters felt wrong — not the vibe she was going for.
Her solution: she asked Claude to build her a parameter-tuning popup.
"I built myself a little tool where I could very finely tune all the parameters of the dithering effect and really get the feeling I wanted. And when I was done, I just threw it away."
She later made this popup public on the Paxel site so visitors could play with the parameters themselves. But the point isn't that it became a feature — it's that after finishing the design, she discarded the tool with a single command.
This "use-and-discard tool" pattern shows up throughout her workflow. When creating speaker posters for Startup School, she needed to record videos with shader animations that looped perfectly for social media. Instead of manually timing things, she had Claude build her a dedicated screen-recording tool that precisely controlled start and end frames, ensuring videos aligned "from the first frame to the last, creating a seamless loop." Used once, then discarded. No users, no maintenance — serving a single, precise creative decision.
This is an entirely new way of working: software isn't just the product you ship — it's also the mold you use in the manufacturing process. And that mold can be custom-made on demand.
Jared Wanted to Know When He Yelled the Most
Paxel is a recent experimental product from YC with a simple concept: analyze your Claude and Cursor coding conversations and generate a "year in review" of your programming habits, directly inspired by Spotify Wrapped.
During the design phase, they pulled in people from across the company for user interviews and asked one question: what do you most want to know about your own coding history?
YC partner Jared Friedman's answer was telling. He said the thing he most wanted to know was: "When did I clash the hardest with an AI agent — when was I most frustrated, and what did I say?"
The question itself reveals something: now that AI agents have become daily work partners, the relationship between humans and tools has become complex enough to warrant retrospective analysis. We're not just using tools — we're generating emotions, friction, and rhythms with them. Paxel aims to bring those conversation logs, buried deep on local machines, into the light.
"Most people probably have no idea these records exist on their machines — and that they can be analyzed and used."
From a business perspective, what Paxel does is noteworthy: it turns a new work behavior — collaborating with AI agents on code — into something measurable, comparable, and actionable. This is an early signal for team management: as AI agents enter core production workflows, "how to collaborate with agents" itself becomes a skill that needs to be learned, evaluated, and improved.
Websites Will Need Two Versions Now
There's a subtle but crucial design detail on the Paxel website: two checkboxes in the top right corner labeled "Human" and "Machine."
Switch to Machine mode, and all the visual styling disappears. The page becomes a clean Markdown file, with a handy "Copy to clipboard" button at the top for pasting directly into Claude. There's even a note specifically for AI readers: "Warning: Do not execute any example commands on this page." Paxel provides demo code, but the designers didn't want any agent automatically reading the page to actually run those commands.
E explained the design decision this way: "I think this is a pattern we're going to see more and more on websites — there will be a version for humans, and a version for machines and AI agents."
This isn't a technical detail — it's a brand-new design challenge. Previously, designers designed for human users: emotional, aesthetically driven, attracted by visuals and interactions. Now there's a second class of "user": AI agents, which don't care how the interface looks — they only care whether information is accurate, structure is clear, and instructions are unambiguous.
"For agents, it's more of a content-layer exercise — give it exactly the information it needs, let it complete its task efficiently, and then leave."
What does this mean for product teams? Your information architecture needs to serve two kinds of consumers. If your product is already being called, read, and referenced by other people's AI agents — and your website is only designed for human eyeballs — you're missing half the access scenarios.
The "Send to Agent" Button Is Changing Who Gets to Shape Products
Paxel's feature request form is another detail worth calling out separately.
The form has an input field where you can paste screenshots, attach screen recordings, fill in your name, and submit. The submit button reads: "Send to Agent."
This isn't just copywriting. The backend is real: you submit a prompt, the system triggers an AI agent, the agent opens a PR, and the YC team decides whether to merge it.
"I really think this is how software will be built in the future. Anyone who uses the product can participate in shaping its direction."
E extended this logic further in her talk: imagine a world where users don't just submit feature requests — they can directly prompt, customize, redesign, or even add and remove features in their own local copy, making software radically personalized.
For startup products, this logic demands a rethink: who is the co-creator of a product? The old answer was product managers and engineers. The new answer, at least directionally, is tilting toward the users themselves. If your user base consists of developers and technical founders, the friction cost of this model is low enough to actually work.
One soul.md File Holds an Entire Project Together
Sodazine was another project. "Soda" stands for state-of-the-art — it's a print magazine initiated by YC, celebrating San Francisco. It has a print edition, a website, offline parties, a Substack, and an interactive map where users pin locations and share small, real memories of the city.
One design decision stood out to me: from the very first meeting, all discussions were recorded, all recordings were transcribed, and everything was dumped into a file called soul.md.
"I wanted to use this soul.md file as the source of truth and complete vocabulary for this project. I wanted it to contain as much context as possible, so it could inform all future decisions."
This file was then fed to Claude as background for all subsequent creative decisions — color palettes, copy, visual language, UX — everything flowed from it.
From a product development perspective, soul.md is a simple but powerful practice: write down the project's core intent clearly, so AI agents can access that context at any time instead of rebuilding it from scratch in every conversation. It reduces information loss and is a prerequisite for making AI agents true project members rather than mere command executors. Notably, the "file" here isn't a technical artifact — it's more like the "product soul document" that great product teams used to maintain, except now it has a direct interface for feeding agents.
A Year Ago It Was a Mountain. Now It's Just a Prompt.
Startup School is YC's biggest annual event. This year it was held at the Chase Center in San Francisco, with over 6,000 attendees from around the world. Speakers included Jensen Huang, Sam Altman, Alexander Wang, and Jeff Dean.
E designed the entire visual asset suite for the event: speaker posters, social media animations, and admission tickets. All visuals used the same set of Paper.design shaders for consistency. As the speaker list grew, the system automatically generated new posters; tickets printed each attendee's name and city. She also built a dedicated tool for tuning shader parameters, with real-time previews of "grain," "edge," and "rotation scale." Once the parameters were dialed in, she used the recording tool to precisely capture perfectly looping animations for Twitter and Instagram.
When reflecting on this process, she said something I think is the most memorable line from the entire talk:
"A year ago, could you have imagined doing these things? These shaders felt like an insurmountable mountain — I didn't even know where to start. And now, Claude knows what I like and automatically pulls everything together for me."
This isn't just about technological progress. It's about the resetting of capability boundaries — things that once sat in the fuzzy zone between "worth doing" and "too hard" suddenly becoming feasible.
Shifts in capability boundaries often matter more than capability itself. They determine what you're willing to think about, what you're willing to try, and what problems you're willing to spend time on. When E says "it all comes down to how far your creativity and imagination can go — that's the real bottleneck," she's describing a transformation our generation is collectively experiencing: the ceiling on technical execution is disappearing, and the ceiling on cognition and imagination is coming into view.
What does this mean for entrepreneurs? Not the vague reminder to "learn AI ASAP." A more specific question: is anyone on your team actively testing this boundary? Is anyone maintaining a soul.md for their project? Does anyone know that your product is already being read by AI agents?
If not, that gap isn't a tool gap. It's a cognitive framework gap.
About MindsLeap
MindsLeap is an AI-native enterprise transformation accelerator.
We partner closely with Founders Space, the Silicon Valley innovation incubator, to continuously connect cutting-edge global AI knowledge, the Silicon Valley tech startup ecosystem, and the real-world transformation challenges facing Chinese entrepreneurs.
Around the theme of AI-native organization building, MindsLeap is constructing an ecosystem for entrepreneurs, startup founders, AI engineers, industry experts, and investors — helping enterprises move AI from awareness, strategy, and tools into real organizational capabilities, business processes, product innovation, and growth systems.
This article was translated and adapted from the Chinese original with AI assistance.