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

Summary (tech concepts, product features, and design/process analysis)

Tools & workflow shifts (designer “agentic” setup)

• The speaker (YC head of design) primarily uses Conductor and Paper (paper.design) for end-to-end work.
• For inspiration, she builds a Pinterest mood book / visual references.
• Instead of typing, she uses Aqua (YC company) to talk to the computer and generate outputs from a “stream of consciousness,” which she describes as feeling magical and fast for ideation and implementation.
• She emphasizes “building the muscle” of editable outputs: rather than generating one-off designs, she treats software as something she can repeatedly tweak.

---

Project 1: Paxel (coding-agents transcript analytics + “Spotify Wrapped” style)

What it is / goal

• Paxel is an experiment to understand how people code with coding agents.
• Coding-agent behavior is described as a black box; Paxel aims to surface “tricks and insights” from real transcripts.

How it works (product behavior)

• Users run a terminal command to pull transcripts.
• It analyzes code/agent transcripts and emails users a report.
• The report includes:
  • Fun fact “cards”, such as:
    • Preferences for a model
    • Patterns like commit timing
    • Prompt/usage behaviors like “plan mode”
  • A more detailed section revealing:
    • Coding decision patterns
    • Strengths
    • Growth areas

Design + UX elements on the page

• The landing page is intentionally explicit that Paxel is an experiment, with lots of content and interactive “hoverable cards” and micro-interactions.
• She uses paper.design shaders—specifically a dithering shader—to create a consistent visual language.
• She built a fine-tuning modal to adjust shader parameters (dithering feel, knobs/dials).
• She even made the model public so others can experiment.
• Strong emphasis on “disposable tools” and iterative previews: build a tuning interface, tweak, discard when done.

Human vs. agent-readable content

• The site includes “human vs machine” modes:
  • A human-facing version with visuals and readable layout.
  • An agent-friendly version that is distilled Markdown for efficient consumption.
• There’s a warning for agents not to execute commands from the page (the “machine” version is for safe content consumption).
• The agent-friendly design challenge is framed as content-first, not visual-first.

Interactive “submit to an agent” feature request

• The site has a form that can be used for either:
  • Bug reports
  • Feature requests
• Users treat it as a prompt box to an agent.
• They can attach screenshots and screen recordings for context.
• Upon submission, the system triggers an agent that opens a PR; humans decide whether to merge.
• The CTA text is intentionally framed as “send to an agent.”
• She highlights the benefit of collecting user names for credits—enabling software to become more personalized.

Conceptual product insight

• Paxel makes hidden artifacts visible: coding transcripts exist inside the machine and are hard to access; Paxel brings them to the surface for feedback and learning.
• As transcript volume grows, Paxel will compare users to others and improve insights.

---

Project 2: Sodazine / “Zen” (intentional non-AI art + agent-assisted site iteration)

Physical art constraint

• The “Zen” project includes literal physical zines/books with cover art and internal graphics.
• For that component, she intentionally chooses no AI involvement, returning to traditional Illustrator work for highly intentional, highly detailed craft.

Agent-focused documentation using transcripts

• For every meeting, she records it and dumps transcripts into a single source-of-truth file (a soul.md-like document).
• She argues this is better than jotting loose notes: the “source of truth” doc becomes a comprehensive context bundle for future decisions and for feeding an agent.

Experimentation & iteration loop for website design

• She builds a workflow around:
  • Pinterest mood boards (initial vibe: black-and-white rudimentary vibe)
  • Feeding images + content direction into Claude to one-shot multiple website variations (e.g., “16 different times”)
  • Creating a personal glossary/collection page to organize, compare, and bookmark the best iterations (including “pin”/bookmark behavior)
• This is framed as disposable design tooling: iterate quickly, throw away what doesn’t work, assemble what does.

Using richer context to unlock surprising ideas

• She stresses that the more context included in soul.md (including article titles, manifesto, meeting context), the more the agent can produce non-obvious ideas.
• Example: a party time/date detail gets organically included in generated design outcomes.

Advanced interactive concept

• One successful iteration becomes a fully interactive map of San Francisco:
  • Users can drop pins and create small stories/memories.
  • Submissions are anonymous except for location selection and story content.
  • The experience includes ways to browse submissions and share them externally (e.g., share as PNG with coordinates).

Practical agent UX

• She contrasts this approach with generic results from “Claude/codeex designs”: to get better design outcomes, you should supply real references (Pinterest, Google images, admired websites) and let the agent detect patterns.

---

Project 3: Startup School branding at Chase Center (shaders + automated social assets)

Event + objective

• She describes Startup School (YC’s major annual event) at Chase Center with thousands attending.
• Branding goal: create assets that feel “very YC,” like a variation of YC, with strong visual consistency.

Shader-driven visual system

• They reuse paper.design shaders (movement texture) and fine-tune parameters like graininess, edge behavior, rotation, and scale.
• These shaders are used across multiple channels (social cards, screens, tickets) to maintain consistency.

Agent-assisted asset templating for many speakers

• Workflow:
  • Start in Figma for initial layout.
  • When generating many speaker cards, she asks Claude to create a template that generates cards from confirmed speaker names.
  • She iterates on text layout variations quickly and consistently across all cards.

Looping social media motion

• She builds a tool to precisely manage screen recording timing so motion loops seamlessly for Twitter/Instagram.
• Claude is used to generate the timing logic so the loop starts/ends at the same pixel alignment for a perfect seamless loop.

Personalized “acceptance ticket”

• The acceptance ticket design:
  • Reuses the shader system.
  • Renders personalized elements like the user’s name, city, and event information.
• Tickets are shared publicly, increasing excitement and adoption.

Broader branding thesis

• The talk frames this as a new paradigm: branding isn’t just static design; agents + code + shader parameterization enable consistent visuals across physical screens, digital assets, and iterative production.

---

Main speakers / sources

• Main speaker: E Bufar — Head of Design at Y Combinator (YC)
• Other referenced YC/internal sources:
  • Charlie (mentioned in relation to Conductor feature-request agent flow)
  • Jared Friedman (YC partner referenced for Paxel prompt idea, “biggest crash out”)
  • YC design/tooling context: Conductor, Codeex (and Cloud Code mentioned), Aqua (talk-to-computer capture tool)
• External inspiration/analogs: Spotify Wrapped