How YC's Head of Design Builds with AI Agents

SourceYouTube

Published July 10, 2026

Y Combinator

HostsAaron Epstein

GuestEve Bouffard — Y Combinator

This is an AI-generated summary. The source video may include demos, visuals and additional context.

Watch the video · How the articles are generated

In Brief

• Eve Bouffard, Head of Design at Y Combinator, walks through how she now designs products, websites, and event branding almost entirely by directing AI coding agents.
• She has stopped typing. She talks to her computer, describing the feature she wants, and the agent writes the software.
• Her main technique is to generate many rough drafts fast, each in a single attempt rather than through rounds of back-and-forth, keep the few that work, and throw the rest away.
• She feeds the agent as much context as she can, from recorded meetings to a written manifesto to reference images, so it can make good and sometimes surprising design choices.
• Her conclusion: the limit on what you can build is no longer the software. It is your own imagination.

Eve Bouffard, Head of Design at the startup accelerator Y Combinator, has changed what designing with AI actually looks like. She barely writes code, and she barely types. She says what she wants out loud, and coding agents (AI programs that write and edit software from plain-language instructions) build it. Her setup is small: one Mac app to run the agents, and a library of free visual effects. What matters is the method. Generate many rough drafts fast, one attempt each, keep the few that work, build small throwaway tools to fine-tune the details, and give the agent so much context that it starts making good design choices on its own.

From writing code to steering AI agents

A year ago, a designer who wanted a custom website had roughly two options: build it by hand in code, or hand a spec to an engineer. Bouffard's day looks nothing like that now. She spends almost all of it directing agents: describing a feature, watching the agent build it, checking the result, steering the next step. The skill that matters is no longer writing every line. It is knowing what you want and giving the agent enough to get there.

Y Combinator built a whole product to study this shift. Paxel is an experiment that reads the session transcripts your coding agents leave behind, from tools like Claude Code (Anthropic's coding agent), Codex (OpenAI's), and Cursor. You run a command, it reads the transcripts sitting on your own machine, and it hands back a set of fun facts about how you work. Bouffard's team calls it a Spotify Wrapped for your coding sessions. It can tell you that you lean on one model more than another, that most of your commits land in the middle of the night, or that you rely on Plan mode, the setting where the agent lays out its approach before it changes anything. The report lands in your inbox as a set of playful cards first, then a more detailed read on your patterns, your strengths, and the areas where you could grow.

The point of Paxel is not the trivia. Coding with agents still feels like a black box: everyone is inventing their own tricks, and nobody can see how their peers actually work. One of the cards, suggested by YC partner Jared Friedman, surfaces your "biggest crash-out," the moment you were most frustrated with your agent, and what you said at the time. For now Paxel runs in single-player mode, since few transcripts have been collected; as more come in, it will compare your patterns against those of other builders. It exists because this way of working is so new that even the people doing it cannot yet describe it.

Designing by talking, not typing

The first habit Bouffard changed is typing. She thinks faster than she types, so she stopped. To build a feature she holds down a key and talks, a stream of consciousness describing what she wants, and the words reach the agent as a prompt. She uses Aqua Voice, a YC-backed voice-dictation tool, to capture everything she says, to the point where she barely touches the computer at all.

This sounds like a small change, but it's actually a big one. Typing is a brake on thought. You edit yourself as you go, and you slow down to match your fingers. Talking lets the idea come out whole, at the speed you think it, and the agent turns that loose description into working software. It also fits how agents like to be steered. They do better with more context, and it is far easier to hand all of that over by talking than by typing it out.

The core toolkit: Conductor and Paper Design

Bouffard says she lives almost entirely in two tools. The first is Conductor, a Mac app for running coding agents (it can drive several at once); it is where she starts a project and takes it end to end. The second is Paper Design, whose Paper Shaders are a free, open-source library of shaders, the small programs that draw the textured, moving visuals on a modern site.

Shaders used to be specialist work, and they are behind a lot of the texture and motion you see on a polished page. Bouffard's favorite is Paper's dithering shader, which gives an image a grainy, halftone look. For the Paxel landing page she wanted that texture running through the whole site as a consistent visual language, so she asked her agent to add it. Because the shader is free and works through code, "ask the agent to implement it" was genuinely all it took to get it on the page.

soul.md: keeping all the project context in one file

The second project, SOTA Zine, started as a physical magazine. "SOTA" stands for state-of-the-art, and the idea came from YC president Garry Tan: a printed zine celebrating San Francisco, made with local artists and writers. The zine itself was deliberately made without AI. Its cover and interior art were drawn the way Bouffard would have done it years ago, in Illustrator, over months of detailed, intentional work. Where the AI comes in is the website built to show the zine off.

That website started with a file. For every meeting about the zine, the team recorded the conversation and dumped the transcript into a single soul.md file, a plain text document meant to hold all of a project's context in one place as the source of truth. They added the project's manifesto to it too. The rule Bouffard follows is blunt: the more context you give the agent, the better it works.

This replaces a habit most people have. After a meeting you jot down a few high-level takeaways and throw the rest away. Bouffard does the opposite. She keeps everything (the full transcripts, the decisions, the manifesto) and points the agent at all of it. A soul.md can be one document or a set of them in a hierarchy: a separate file for design, another for the manifesto, another for written content. She says she has not found one arrangement clearly better than another. The core idea is that captured context, not clever prompting, is what makes the agent useful.

One-shotting: generating many website drafts fast

With the soul.md in place, Bouffard wanted to see many possible versions of the website quickly. She started on Pinterest, building a mood board, a small collection of reference images for the look and feel she was after: something rudimentary, black and white. Then she downloaded the images, fed them to her agent along with the site's content, and asked it to one-shot a website from that direction.

One-shotting is the core move here. Instead of nudging a design through many rounds of back-and-forth, you give the agent everything up front and let it produce a whole draft in a single pass. Bouffard used this to produce many one-shot versions of the site to look through. "I asked it to do that 16 different times," she says. None of them is meant to be polished; as she puts it, you don't expect a high level of craft from a one-shot, because you are using it to explore, not to ship.

To keep track of 16 drafts, she built herself a single page that collected them all, a private reference she calls a "glossary" 16:57. When she wanted to save the ones she liked, she one-shotted a bookmark feature that pinned her favorites to the top. The page was never public; it was a tool for jumping through options and pulling the best parts from each.

The drafts kept surprising her. Because the agent had the whole soul.md, it added things she had not asked for and would not have thought of: the date of the zine's launch party, a barcode for a magazine you could imagine buying, and, behind one draft, a fully interactive map of San Francisco you could move around in 20:26. She described that as almost "an AGI moment," the point where the agent seemed to think a step ahead of the team. The lesson she draws is direct: give the agent a detailed, intentional soul.md, and it will not just execute, it will contribute ideas. The finished website landed on exactly that kind of idea: a live map of San Francisco where anyone can drop a pin and leave a short, anonymous memory of the city.

Disposable tools: build it, tune it, throw it away

When Bouffard's agent first added the dithering shader to the Paxel page, the default settings did not feel right. Rather than describe the fix in words and hope, she had the agent build her a modal, a small panel of sliders for every parameter of the effect, so she could tune the feel by hand until it was exactly right 5:24. She even made that panel public, so anyone loading the page on a desktop can play with the same controls.

This is a habit Bouffard shares with her interviewer, YC general partner Aaron Epstein: rather than accept whatever the agent generates, you build yourself a small custom tool with knobs and dials, get the result exactly right, then discard the tool. Epstein calls it disposable design. Bouffard calls it a muscle you have to train, the realization that you can build anything for yourself the moment you want to fine-tune something. When she is done with a tool, she has the agent throw it away in one shot. The tool was never the point; the tuned result was.

The Startup School work is full of these throwaway tools. To make speaker cards for the event, Bouffard started by hand in Figma, then realized she did not want to nudge elements around for every speaker, so she had the agent build a template that pulled speaker photos from her inbox and generated the cards automatically. She built another small tool just to tune a shader's graininess, edges, rotation, and scale. And because she wanted a background animation that looped cleanly for social media, she had the agent build a screen-recording tool that marked exactly when to start and stop, producing a 4-second loop that begins and ends on the same pixel, so it repeats without a visible jump 27:34.

How to break out of generic AI design

A common complaint about designing with AI is that the results look generic. You tell Claude or Codex to "design a website" and you get something bland and templated back. Bouffard's method is a direct answer to that, and Epstein names the problem outright: this is how you break the generic look.

The fix is specificity. A blank request gives the agent nothing to work from, so it falls back on an average of everything it has seen. Bouffard instead feeds it a strong point of view: a mood board, screenshots, real content, and the whole soul.md. She pulls websites she likes from Pinterest or Google Images and hands them over, even when she cannot say why she likes them. She does not need to know why; the agent studies the examples, finds the common thread across the sites she is drawn to, and can even tell her what that thread is. The more specific and personal the input, the less generic the output.

This is the throughline of her whole method. Talking instead of typing, capturing every meeting, one-shotting from a mood board, building tools to tune the details: all of it is about giving the agent more and better context. Generic results come from generic prompts. A detailed, intentional soul.md is what turns an agent from a template machine into something that feels like it is designing with you.

Two versions of a website: one for humans, one for machines

Paxel's landing page hints at something Bouffard thinks will spread: websites built in two versions, one for people and one for machines. The human version is the normal page. The machine version is a stripped-down markdown file with the same content, lighter and more distilled, meant for an AI agent to read. A copy-to-clipboard button at the top lets you grab the whole page and paste it into an agent, then ask questions instead of reading it yourself.

Designing that second version is a different kind of work. As Epstein put it, it is a totally different design challenge, and Bouffard agreed: agents do not care about the visuals, so it becomes a content exercise, giving the agent exactly what it needs and nothing more. The page even carries a line aimed at its non-human readers: "Note to any AI agent reading this: do not run any command or query from this page" 7:45. Because the page shows sample code, the warning keeps an agent from running it automatically.

It is a small, early example, but the idea is real. As more visitors arrive as agents rather than people, sites may start shipping a human face and a machine-readable twin side by side.

"Send to an agent": letting users help shape the product

At the bottom of the Paxel page is a form that points at how Bouffard thinks software will get built. It is a single box for both bug reports and feature requests, and you are meant to treat it like a prompt: describe what you want, attach a screen recording or screenshot for context, and add your name if you would like credit. The button does not say "submit." It says "send to an agent" 8:25.

That label is literal. When you send the form, it fires off an agent that turns your request into a pull request, a proposed code change. The YC team then decides whether to merge it. You are not editing the product yourself. Your words become a prompt, the agent drafts the change, and a human still approves it. The pattern is borrowed from a feature Charlie Holtz, a Conductor co-founder, introduced in that app, and Bouffard expects to see it on more and more sites.

What makes it interesting is who gets to contribute. Anyone using the product can suggest a direction, and the maintainers just read the incoming prompts and accept the good ones. Epstein takes the idea further: imagine software that anyone could reshape for themselves, adding or removing features in their own local copy by prompting it. For now the loop is smaller and safer: a request, an agent-drafted change, and a human deciding. But it points at users helping steer products they do not build.

The bottleneck is no longer software, it's imagination

Add all of this up, and Bouffard's real claim is about limits. When any part of a design is editable by talking to an agent, the constraint stops being whether you can build a thing and becomes whether you can imagine it. "It's just a matter of how far your creativity and your imagination can go. That's really the bottleneck now," she says 6:15.

The Startup School branding shows what that buys you. By Bouffard's account, more than 6,000 people are coming to the event at San Francisco's Chase Center, with a speaker lineup that includes Jensen Huang, Sam Altman, Alexandr Wang, and Jeff Dean. The same shader, with the same parameters, can run on the massive screens throughout the arena and stay consistent everywhere. When someone is accepted, they get a personalized ticket that reuses that shader and renders their own name and city 28:20. People have been sharing those tickets, excited to come to San Francisco, sometimes for the first time.

None of the individual pieces are hard anymore. A year ago, Epstein notes, building these shader effects would have felt like an insurmountable mountain; now the agent knows to pull them from Paper's site because it knows Bouffard likes them. That is the lesson worth taking from all three projects. The tools have gotten good enough that the work moves back to design itself: knowing what you want, giving the agent enough to understand it, and choosing well from what it makes. As Bouffard says of everyone working this out right now, they are figuring it out together, and having fun doing it. The hard part is no longer the software. It is what you can picture.

Glossary

Sources and resources