By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.

YC's Head of Design Shows You How To Design With AI

AI isn't just changing the tools designers use. It's changing the way they think, prototype, and build.

Aaron EpsteinhostEve Bouffardguest

Jul 10, 202630mWatch on YouTube

Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.

CHAPTERS

Jump to publisher timestamps

Chapter summaries

Eve’s current AI-first design stack: Conductor, Paper Design, Pinterest inspiration, and voice input
Aaron asks Eve what tools she’s using now, and she explains that her workflow has consolidated around Conductor and Paper Design for end-to-end projects. She also describes how she gathers visual direction via Pinterest mood boards and increasingly interacts with her computer by speaking rather than typing.
- •Eve says she now works almost entirely in Conductor and Paper Design for full projects
- •She uses Pinterest mood boards to set visual inspiration and direction
- •Eve describes using voice because she thinks faster than she types
- •She uses Aqua to capture her spoken stream-of-consciousness instructions to the computer
Paxel’s purpose: making coding-agent transcripts understandable and fun
Eve explains Paxel as an experiment to understand how people are coding with agents, since practices are changing quickly and feel like a black box. The team’s goal is to surface patterns and share insights, while keeping the experience playful.
- •Paxel is framed as an experiment to learn how people code with coding agents today
- •Eve says the goal is to uncover tricks, insights, and patterns worth sharing
- •Playfulness is a core objective, inspired by the idea of “Spotify Wrapped” for coding sessions
- •Eve shares that office interviews shaped which fun “cards” to include, including a suggestion from Jared Friedman about identifying a user’s “biggest crash out” moment
How Paxel works: pulling transcripts and generating ‘fun facts’ cards
Eve walks through the product flow: a terminal command pulls transcripts and analyzes them to produce a report. The output highlights behavioral patterns—like preferred models or timing habits—and evolves over time as more data is collected.
- •Users run a terminal command that pulls transcripts and analyzes them
- •Eve lists sources/features it reads across tools and outputs fun facts
- •Example insights include model preference, late-night commits, plan mode usage, and common prompts
- •The current version is described as “single player mode,” with comparisons planned once more transcripts exist
Designing the Paxel landing page: clarity of motivation, interactive cards, and consistent visual language
Eve explains that the landing page prioritizes being upfront about Paxel’s experimental motivation, even if that means more text than typical. She ties the page’s structure to the playful “cards” concept, using hover micro-interactions and a unified visual style.
- •Eve wanted the landing page to clearly state Paxel’s motivation as an experiment
- •She acknowledges the page includes unusually prominent explanatory text
- •Interactive hover cards and micro-interactions reinforce the playful “Wrapped-like” concept
- •A consistent visual language is emphasized across the site
Shaders and ‘build-a-tool’ iteration: dithering with Paper Shaders and fine-tuning via custom knobs
Eve describes using Paper.Design shaders—specifically a dithering shader—implemented via Claude, then building a private tuning tool to refine parameters. Aaron highlights the broader pattern: turning designs into living, adjustable systems rather than static assets.
- •Eve used Paper Shaders (Paper.Design) and asked Claude to implement a dithering shader
- •She built a custom tool to fine-tune shader parameters until it felt right
- •Eve says they often build small internal ‘models’ to perfect details, then discard them when done
- •Aaron notes the shift from static graphics to dynamic, controllable systems with knobs and dials
‘Human vs. machine’ website modes: agent-friendly markdown, copy-to-clipboard, and new content constraints
Eve explains the idea of offering separate website versions for humans and for agents. The agent-oriented version is a distilled markdown representation, designed for easy consumption and safe usage, including a prominent copy-to-clipboard flow.
- •Eve predicts websites may increasingly offer ‘for humans’ and ‘for machines/agents’ modes
- •They created an agent-friendly markdown version with distilled content
- •A copy-to-clipboard feature lets users paste the full page into tools like Claude Code/Codex to ask questions
- •Aaron and Eve note the design challenge shifts from visuals to precise content and safety considerations (e.g., not auto-running sample commands)
“Send to an agent” feature requests: prompts as product feedback that can generate PRs
Eve describes a form inspired by Conductor that treats bug reports and feature requests as agent-ready prompts. Submissions can include screenshots or recordings, and the backend triggers an agent to open a PR that the team can choose to merge.
- •The form is designed for both bug reports and feature requests
- •Users are encouraged to write requests like prompts, and attach screenshots/screen recordings for context
- •The CTA explicitly says “Send to an agent,” matching what happens behind the scenes
- •Eve says submissions can open PRs, with humans deciding whether to merge; names can be collected to give credit later
Why Paxel matters long-term: inbox reports, deeper pattern analysis, and surfacing hidden transcripts
Eve shows how Paxel delivers a report with fun cards and a deeper breakdown of patterns, including strengths and growth areas. She frames Paxel as a way to bring buried coding transcripts to the surface so people can learn from them, especially as the dataset grows.
- •Paxel sends an inbox report with fun cards plus a more detailed view of decision patterns
- •Eve says the report can include strengths and growth areas for users to focus on
- •She expects more valuable comparative insights as more transcripts are collected
- •Eve argues transcripts are deeply hidden on machines and many people don’t realize they can be analyzed
SOTA Zine origin and principle: celebrating San Francisco, starting with a physical, non-AI-crafted zine
Eve introduces SOTA Zine as ‘State Of The Art,’ inspired by an idea from Garry to celebrate San Francisco with local artists and writers. She emphasizes that the physical zine’s core graphic design and art were intentionally created without AI involvement.
- •SOTA means ‘State Of The Art,’ with a goal of celebrating San Francisco
- •Eve says the project involves collaborating with artists and writers in the city
- •The zine is a physical object, not just a website
- •For the zine’s cover and some interior art, they deliberately used a no-AI approach, returning to Illustrator for highly intentional, detailed work
Soul.md as a source of truth: recording meetings, capturing exhaustive context, and structuring project memory
Eve describes recording every meeting about the zine and placing transcripts into a Soul.md file that serves as the project’s source of truth. She argues that capturing maximal context—including a manifesto—improves agent-driven decisions across design and execution.
- •Eve says they recorded all zine meetings and dumped transcripts into a Soul.md file
- •She treats Soul.md as an exhaustive glossary and source of truth for the project
- •They also wrote a manifesto and added it to Soul.md to increase context for the agent
- •Eve notes Soul.md can be one file or a hierarchy (e.g., separate files for design, manifesto, written content), but the key is sharing rich context with agents
One-shot website exploration: mood board to 16 iterations, navigation glossary, and pinning favorites
Eve explains a rapid iteration workflow: assemble a Pinterest mood board, feed images and project context into Claude, and generate many one-shot website variations. She then builds a personal navigation page to compare versions and pin the best options.
- •Eve starts with a Pinterest mood board to define a rudimentary black-and-white vibe
- •She downloads reference images and asks Claude to one-shot website concepts using the vibe plus project content
- •She repeats this process 16 times to explore a wide design space quickly
- •Eve builds an internal ‘glossary’ page to browse iterations and adds a pinning feature to keep favorites at the top
Breaking out of generic AI design: feeding richer inputs and letting the agent surprise you
Reviewing the iterations, Eve highlights how detailed context in Soul.md leads to more interesting and unexpected outputs, like including launch party details or playful zine cues. She and Aaron discuss how reference images and examples help avoid generic AI design results.
- •Eve uses article titles and other details from Soul.md to ground one-shot designs in real content
- •She notes the agent included unexpected elements (like party details and a barcode), which helped brainstorming
- •Eve says high-quality iterations require shepherding the agent with intentional direction and lots of context
- •Aaron observes many people get generic outputs; Eve suggests supplying Pinterest/Google Images/examples of websites you like so the agent can infer what you’re drawn to
Where SOTA Zine landed: an interactive SF memory map and shareable story cards
Eve describes the final web direction: an anonymous interactive map where people pin locations and share short memories from San Francisco. The site also links out to related zine content and includes a share flow that exports a submission as a PNG with location details.
- •The final concept is an interactive map of San Francisco for anonymous pinned stories and memories
- •Eve frames it as a way to understand how people experience the city through small, intimate moments
- •The site includes entry points to zine-related assets like digital posters and a Substack for articles
- •A sharing feature downloads a selected submission as a PNG and includes coordinates and the street tied to the story
Startup School branding workflow: automated speaker cards, shader parameter tuning, and perfect looping recordings
Aaron transitions to Startup School assets, and Eve explains how she moved from manual layout to a template-driven tool that generates many speaker cards automatically. She details experimenting with text layouts, fine-tuning Paper.Design shader parameters, and building a helper tool to record seamless looping clips for social media.
- •Eve describes Startup School as YC’s biggest event of the year and outlines the need for shareable speaker lineup assets
- •She started in Figma but switched to asking Claude to create a template/tool to avoid repetitive manual work across many speakers
- •Her tool generates speaker cards and supports quick iteration on text layout variations
- •Eve fine-tuned shader grain, edges, rotation, and scale, and built a loop-timing screen-recording tool to capture a smooth, seamless animation loop
Personalized acceptance tickets and arena-scale consistency: reusing shaders across the full event experience
Eve explains designing personalized tickets that render each attendee’s name and city using the same shader style as other assets, and she highlights how people shared them on social media. She closes by describing how agents make it easier to keep branding consistent, even across massive screens at the venue.
- •They created personalized acceptance tickets using the same shader aesthetic as other Startup School assets
- •The tickets include individualized details like the recipient’s name and city, plus event information
- •Eve says it’s been rewarding to see attendees share tickets and excitement publicly
- •She contrasts the difficulty of building shaders previously with how agents now make it easier, and emphasizes consistent shader parameters across venue-scale displays

Publisher timestamps

Get more out of YouTube videos.

Read along on your phone, search it beside the video, or ask your favorite AI app what was said.