Guillermo Rauch: How v0 Went from Prototyping Tool to 3,200 PRs Merged Per Day

Vibe coding’s dirty secret: going from zero to one is easy. The hard part is iterating at scale and deploying safely. Guillermo Rauch’s answer is to bring the full Git workflow into v0, turning it from a prototyping tool into a production engineering environment where anyone in the company can open branches, submit PRs, and ship to production.

The Podcast

Guillermo Rauch, CEO of Vercel, joins Claire Vo on the “How I AI” podcast for a live demo session. Rather than talking abstractly about AI’s potential, Rauch walks through real workflows: building a feature on skills.sh (a viral skills marketplace with 34,000+ community submissions), pushing it through branch previews, and merging to production. The conversation is part product demo, part philosophy on how vibe coding changes organizational dynamics.

Skills.sh: From Slack Chat to Viral Platform

The most concrete case study in the conversation is skills.sh, a hub for AI skills that went viral with 500 new submissions per hour at its peak.

The entire product started as casual Slack messages. Rauch and his team chatted about the idea, prompted v0 with rough concepts, and iterated collaboratively. The VP of design took the initial v0 prompts further, then the CTO and product lead refined it. The design aesthetic they landed on, which Rauch calls “terminal core,” came from a single suggestion: “Wouldn’t it be cool if we make the top of the website look like a terminal?”

Despite being vibe-coded, skills.sh runs at production scale. With 35,000+ skills, Rauch demonstrates instant page transitions and production-grade performance, all running on Vercel infrastructure.

The motivation for skills.sh came from concrete user feedback: people were asking why models like Opus 4.5 “kind of know the latest Next.js but not really.” Skills solve this by letting the community package best practices that models can consume, essentially crowdsourced context injection for AI.

The Git Workflow Revolution in v0

The core announcement is v0’s new Git integration. Previously, v0 users lived in prototype land, what Rauch calls “La La Land.” Now the tool supports the full engineering workflow:

1. Branch creation: One button click creates a named branch following engineering conventions (contributor/feature-name)
2. Full dev environment: VS Code editor, installed dependencies, real Next.js dev server running in a virtual machine
3. Branch previews: Changes deploy to Vercel’s production CDN for testing in a production-like environment
4. PR generation: v0 automatically creates pull requests with descriptions

Rauch demonstrates this by adding a five-star rating system to skills.sh. His prompt is deliberately casual: three incomplete sentences plus “make it make sense within the style of this skills website.” He also adds a production concern, asking v0 to be “mindful that we should rate limit or prevent abuse on the scores.”

The result fits the terminal core design, persists data correctly, and stores ratings in the project’s existing Redis/Upstash backend. When Rauch notices a layout shift on page refresh (a detail he calls out as triggering his “neurosis”), he prompts a fix and it’s resolved in seconds.

The growth numbers are striking: PRs merged via v0 went from near zero in early January to 3,200 per day by late January, a 100x increase in weeks. And this was still in early preview.

The End of “Petitioning the Government”

The organizational implications are where this gets interesting. Rauch frames the old workflow as “petitioning the government”: marketers who wanted to change a page had to beg engineers or pray the CMS supported their vision.

Now anyone can open a production page in v0, prompt changes, and submit a PR. The engineering team still reviews and merges, but the creation bottleneck is gone.

“The humiliation ritual of prioritization goes away and you can actually focus your time on defending the merits of an idea on the actual idea as opposed to the hypothesis of the idea that then has to be implemented.”

This changes what gets debated. Instead of arguing whether an idea is worth engineering time to build, teams argue about the idea itself, because the prototype already exists.

Rauch shares that before the Git workflow existed, he was creating v0 prototypes with no mechanism to hand them off. Now there’s a common substrate: anyone’s idea can become a PR. He demonstrates his own workflow: whenever he has a product improvement idea, he creates a v0 prototype and hands it off as a pull request to the engineering team.

Customer Zero, Customer One

When asked about product-market fit, Rauch shares a mental model he calls “Customer Zero” and “Customer One.”

Customer Zero is the team itself. He references Rick Rubin’s confidence in taste: they build what they believe should exist, trusting their own judgment first.

Customer One is a tight design partner loop. Claire Vo, for example, is on a direct text chain with Vercel’s CPO, sending bug reports and feature requests in real time. This tight feedback loop, rather than broad surveys, shapes product direction.

There’s also a philosophy about internal tooling: every day, the team asks itself “how can we build an agent that takes over a task that we were previously giving to a person?” And typically the person who was doing that task becomes the one building the agent.

“Every day we’re basically asking ourselves how can we build an agent that takes over a task that we were previously giving to a person, and typically the person that was working on that task is now the one creating the agent.”

Physical AI and Teaching Kids

Rauch’s personal v0 projects reveal his thinking about where things are heading. His 3D chess game (v0-chess-match.vercel.app) pits two AI models against each other on a 3D board, streaming their thinking tokens so viewers can watch the models reason about moves. It runs on Vercel’s workflow engine for durability, meaning the game can theoretically run forever, surviving network failures and compute hiccups.

Building it taught Rauch about 3D development through a multi-model workflow: when stuck on 3D mesh issues, he’d ask another model to explain the concepts, then paste those explanations into v0 as context. He also had v0 build him debugging tools (mesh visualization, texture toggling), so the AI could effectively help debug itself.

With his kids, he vibe-coded a Vesta Board integration at the office, teaching them the concept of APIs by showing how typing code could change a physical display in the real world. Two of four kids were genuinely mind-blown (one was on his iPad, not paying attention).

On the topic of whether to teach kids typing, Rauch shares that his father in Argentina tricked him with a soccer typing game: to score goals, you had to type fast. He’s looking for the modern equivalent while his kids are getting into speech-to-text.

What’s Coming

Rauch hints at several items on Vercel’s roadmap:

• Text-to-SVG: Recraft models through Vercel’s AI gateway for creating scalable, high-DPI assets
• Video generation: something unspecified dropping soon
• Generative UI (“Flash v0”): research into spontaneous UI generation where agents render interfaces on the fly without deploying full applications
• React Native / App Store deployment: a “long-held dream” to make pushing to app stores as simple as pushing to the web, leveraging the same React-based deployment platform
• Agent building in v0: using v0 to create workflow-based agents that combine sandboxes, durable execution, and existing Vercel infrastructure

Afterthoughts

The 3,200 PRs/day number is the real headline. This isn’t prototype generation; it’s production code flowing through Git, reviewed and merged. That’s the gap between “AI can make demos” and “AI changes how companies ship.”

• The “petitioning the government” metaphor names something real about organizational friction. The bottleneck was never the idea; it was the translation cost from idea to implementation. When that cost drops to a prompt, who contributes to a product changes fundamentally.
• Rauch’s approach to getting unstuck is worth noting: use another model as a domain expert consultant, copy its explanations as context for v0, and have the AI build its own debugging tools. Not elegant, but practical.
• The Customer Zero / Customer One framework is a useful heuristic for early-stage product development. Build for yourself first (taste), then pressure-test with a tight design partner (reality). Skip the broad surveys until you have something concrete.
• Skills as a distribution mechanism for AI best practices may matter more than any individual tool. If the model “kind of knows Next.js but not really,” skills close that gap. It’s crowdsourced context injection, and with 500 submissions per hour, the flywheel is spinning fast.