Design After the Bottle­neck.

Eve Bouffard, YC's Head of Design, shows how three projects rewired the process — soul files, dual audiences, and sixteen one-shots before breakfast.

▶ Play video

Photo: Y Combinator / Design Review

What this episode is really about.

Most design process videos are about tools. This one is about a reordering. Eve Bouffard, YC's Head of Design, walks through three projects — Paxel, SOTA Zine, and Startup School 2026 — and the cumulative argument, never quite stated outright, is that something structural has changed. The bottleneck in design used to be execution. Build time, handoff latency, iteration cost, the gap between "I can see it" and "I can ship it." Those costs have collapsed. What's left is imagination: the capacity to define what the thing should feel like before the agent makes it.

The projects span a spectrum. Paxel is a data product with a distinct aesthetic (dithered shaders, Spotify-Wrapped-style stat cards). SOTA Zine is an editorial celebration of San Francisco built with artists and writers. Startup School 2026 is a large-scale event brand deployed across screens, physical materials, and an arena stage. In each case, Eve's story isn't about prompting. It's about what she did before prompting — and what she built to make the fine-tuning feel right when the prompts fell short.

Three kinds of latticework edits are on offer. Some established Farnam Street models come out amplified: first principles, activation energy, feedback loops, the classic argument for creative constraints. Some get complicated: specialization, the idea that tools shape thinking, the ownership model of creative work. And a handful of new models emerge — small, portable, and surprisingly general.

Old models, sharper edges.

The episode's clearest illustration of first-principles thinking is the soul.md file. Eve describes a common designer workflow: meetings, stakeholder conversations, scattered notes, a summary document that somebody maybe reads. Her replacement is blunt — record everything, filter nothing, and dump it all into one file that becomes the authoritative context for every future decision on the project. Not a map of the project. The project's soul. The distinction matters: a summary compresses by the judgment of whoever wrote it; a soul file compresses by accumulation. The AI needs context, not conclusions.

Creative constraints make an appearance in the one-shot fleet. Before picking a direction for SOTA Zine, Eve fed a mood board into Claude and asked it to generate 16 distinct websites from the same prompt — one-shotted, all different. The constraint is the fixed brief; the variation is the output space. She then built herself a navigation tool to scroll through all sixteen. The point isn't efficiency; it's coverage. Range before depth. You can't find the edge of a design space if you only generate one candidate.

Activation energy runs underneath the whole episode — specifically, the cost of going from "I want to tweak this" to "I have tweaked this." Eve describes building custom fine-tuning modals, recording-trigger scripts, and glossary navigators — tools that exist only to lower that cost for herself. Each one is small. Each one changes what's worth iterating on. The phrase she returns to is "training the muscle": every bespoke tool you build for yourself makes it easier to build the next one. Feedback loops tighten into something that approaches real-time.

Models that do not survive intact.

The classic argument for specialization — hire the frontend engineer, the motion designer, the brand strategist, the QA tester — gets quietly reversed. Eve isn't hiring specialists. She's building specialized tools for herself, then deploying them. The human stays generalist; the scaffolding specialises. It's the same two-layer pattern that Gary Tan called thin harness, fat skills on an earlier Lightcone, applied to individual design workflow rather than agentic architecture. The implication is that specialization migrates from people to instruments — and instruments are now cheap to build.

The broader Whorfian idea — that the tools you have shape the thoughts you can think — gets complicated when you can make your tools. The old version of this is: "I'm a Figma designer, so I think in components and frames." The new version is murkier. If every design decision can be instantiated in code, and code can be generated from description, then the constraint on thinking is less the tool and more the quality of your imagination and your prompts. Tools no longer define the design space; they execute it. Which means the limiting factor shifts back to something older and harder to teach: taste.

Finally, the traditional model of creative ownership bends. Eve describes moments where the agent surprised her — hover effects she didn't specify, visuals it scraped and surfaced that she wouldn't have found herself. The designer didn't produce those things. She accepted them, refined them, directed them. The right analogy might be closer to editing than authoring. Attribution stays with Eve; the process of getting there is collaborative in a way the model of "the designer makes" doesn't quite capture.

New entries for the latticework.

The most durable new model in the episode is the soul.md — a single canonical context file that replaces the scattered notes, summary docs, and stakeholder-meeting takeaways of traditional design briefs. Its logic: an AI doesn't need your conclusions, it needs your context. The more unfiltered the input, the better the output. The soul file grows continuously, gets passed to every new conversation about the project, and becomes the memory the model otherwise lacks. Generalisable far beyond design: any long-running project with evolving requirements and multiple collaborators benefits from one place where everything lands, unedited.

The second model is bespoke scaffolding — the deliberate practice of building small tools for yourself before you need them at scale. Eve's shader modal, her recording-trigger script, her 16-iteration glossary navigator: none of these are products. They're instruments. The pattern: if you find yourself repeatedly adjusting a parameter by hand, or navigating through outputs tediously, build the thing that makes it fast. Then optionally publish it. The key insight is that the cost of building a bespoke tool has dropped far enough that the break-even threshold is now a single frustrating afternoon.

The Paxel project introduces dual-audience design — the idea that every significant digital artifact now has two readers: a human and a machine. The human version is visual, rich, animated. The machine version is a markdown file: same information, stripped of presentation, structured for parsing. Eve built both deliberately for Paxel. The implication is that "the design" is no longer singular. It forks at publication. What the agent reads should be different from what the person reads — and the designer is responsible for both.

The episode closes with a vision Eve calls prompts-as-product: the feature request form in Paxel, which doubles as an agent submission interface — a user's natural-language request that fires off code changes in their own local copy. The idea generalises: what if every piece of software shipped a prompt interface alongside its GUI? The design challenge is no longer "how do I arrange these controls" but "how do I make my system prompt-addressable." That's a new discipline, and the latticework doesn't have a model for it yet.

When to reach for which.

The latticework, after Design Review.

The video's central provocation is quiet: if execution is no longer the bottleneck, what changes? The answer, it turns out, is the shape of the designer's job. Eve still controls taste. She still defines the soul of each project. But she's building instruments instead of hiring specialists, generating fleets instead of drafting single concepts, and shipping two versions of every artifact instead of one. The role didn't shrink. It forked.

The biggest bottleneck is no longer software. It's imagination. — Eve Bouffard, YC Design Review

Munger's latticework gains a few new entries: soul.md as canonical truth, bespoke scaffolding as cost-collapse response, dual-audience design as structural artifact discipline, and prompts-as-product as a discipline that doesn't have a name yet. The episode is most useful for what it implies about the things already in the latticework: activation energy, creative constraints, feedback loops — all of them now operate at a timescale where the gain from tightening them is immediate rather than theoretical.