
Y Combinator
YC's Head of Design Shows You How To Design With AI
Summarised with Bite · 12 min read
This is a fast, concrete walkthrough of how YC's Head of Design is actually designing with AI right now, not in theory. The big idea is surprising: the hard part is no longer drawing pixels by hand, it is feeding agents the right context, building tiny custom tools for yourself, and treating design as a live, editable system instead of a fixed mockup.
0:07 – 12:35
Paxel, designing for humans, agents, and your future self
The conversation opens with a confession that would have sounded extreme a year ago: she now works almost entirely in Conductor and paper.design, and she barely types. Instead of carefully writing prompts, she presses the function key, talks in a stream of consciousness, and lets Aqua capture it. That small detail ends up framing the whole episode. The workflow is no longer, "sit down and craft every interface by hand." It is, "think out loud fast enough that the computer can keep up." Paxel becomes the first proof. The product itself is an experiment to understand how people code with coding agents by analyzing their transcripts, those buried records most people do not even realize live on their machines. The goal is not just surveillance or analytics. It is to make self-knowledge fun. She says they were heavily inspired by Spotify Wrapped, and that explains the tone of the whole product. Instead of dumping metrics on users, Paxel turns coding behavior into playful cards, including the most memorable example: Jared Friedman wanted to know his "biggest crash out," the moment he was most frustrated with his agent and what he said. That playful layer sits on top of a serious design choice. On the landing page, she intentionally puts a lot of text front and center because she wants visitors to immediately understand the motivation: this is an experiment trying to understand how the world codes now. Then she carries that feeling visually with interactive cards and a consistent shader language built from paper.design's dithering shader. The memorable part is not just that she used a shader. It is that when Claude guessed the wrong parameters, she did not settle. She built herself a little modal to fine tune every knob until the effect felt right. That becomes one of the video's deepest lessons: when AI gives you almost the right thing, the winning move is often to generate a custom control panel for yourself. Then the page takes an unexpected turn. There is a human version and a machine version. The machine version is basically a distilled markdown file that agents can read, with a warning not to run any command or query from the page because sample code is present. This is a new kind of design problem. Humans care about visuals and pacing. Agents do not. They need exact, compact content. The same idea shows up in the feature request form, where the button literally says "send to an agent." Users can submit a bug report or feature request like a prompt, attach screenshots or screen recordings, and the backend opens a PR automatically. The designer's role shifts from implementing every change to curating which generated changes deserve to be merged. By the end of the Paxel segment, the larger point becomes clear. Software is becoming more personal, more editable, and more participatory. Paxel's report email, with its fun cards and deeper breakdown of strengths and growth areas, is only the first layer. As more transcripts arrive, the real value is comparison and pattern recognition across builders. In other words, Paxel is not just a fun toy for coding stats. It is an attempt to drag an invisible part of modern work into the light so people can actually learn from it.
2 more sections in the app
- 13:06 – 24:30Sodazine and the rise of the soul.md file
- 24:30 – 30:37Startup School, brand systems you can generate, tune, and loop




