How I AIGuillermo Rauch: Vercel CEO on how v0 hit 3,200 PRs merged per day (and lets anyone ship)
CHAPTERS
- 0:00 – 1:35
Why “vibe coding” needs production-grade iteration and safe deploys
Guillermo frames the core problem: prompting a prototype is easy, but scaling iteration, review, and safe deployment is hard. He contrasts the old workflow—non-engineers “petitioning” engineers or relying on rigid CMS setups—with a new approach where anyone can propose changes through v0 with proper safeguards.
- •Zero-to-one demos are easy; shipping changes safely at scale is the real challenge
- •Non-engineers historically depended on engineers or inflexible CMS workflows
- •v0 aims to reduce friction to change while keeping production rigor
- •Preview + review + merge are positioned as the scalable path for AI-made changes
- 1:35 – 3:32
skills.sh explained: a viral hub for reusable agent “skills” built with v0
They introduce skills.sh as an npm-like ecosystem for sharing skills used by agentic frameworks. Guillermo highlights its scale (tens of thousands of skills) and the speed/quality of the production experience, emphasizing that it was conceived and iterated via v0.
- •skills.sh as an open hub for community-submitted skills (like npm)
- •Rapid community growth and viral distribution
- •Site performance and production-grade UX despite massive content
- •Built with a prompt-driven, collaborative workflow inside Vercel
- 3:32 – 4:51
How Vercel ideated skills.sh: prompt-driven design collaboration (“terminal core”)
Guillermo describes the internal creative process: ideas emerge in Slack, get translated into v0 prompts, and evolve quickly through iteration. He gives a concrete example of shaping the look and feel—making the site’s top resemble a terminal UI.
- •Prompting as a collaborative design language across the team
- •Fast iteration from inspiration to functioning product surface
- •Design direction example: “style terminal core” header concept
- •v0 as the substrate for moving from idea to shippable implementation
- 4:51 – 9:07
v0 + GitHub workflow: create a branch from v0 and run the full project in-browser
Guillermo demonstrates importing the skills.sh repo-backed project, creating a new branch directly inside v0, and launching a full dev environment. The UI reflects familiar engineering conventions (project/branch naming) while bundling editor + running preview together.
- •New Branch makes Git flow a first-class experience inside v0
- •Branch naming mirrors real team conventions (contributor/feature)
- •In-browser VS Code editor paired with a live running preview
- •Dependencies and exact project versions are installed automatically
- 9:07 – 11:17
From prompt to feature request: adding a five-star rating system with abuse controls
They choose a real feature request driven by community needs: a rating system to help identify high-quality skills amid high submission volume. Guillermo models “production thinking” in the prompt by calling out abuse prevention and rate limiting from the start.
- •Community signal: users want quality/vetting indicators
- •Feature plan: five-star ratings in the skill sidebar
- •Operational risk awareness: abuse, spam, and rate limiting
- •Prompting style: concise intent + constraints + fit with existing design
- 11:17 – 12:22
Watching v0 implement with the existing stack (Redis) and matching site styling
Guillermo explains what v0 is doing as it scans the repo, respects the existing data source (Redis), and implements UI + persistence. They validate the feature in the preview and note small UI issues that emerge during real interaction.
- •v0 inspects the codebase instead of inventing a new architecture
- •Uses the existing Redis/AppStash setup for storage
- •UI output matches the site’s terminal-core styling
- •Persistence validated via refresh and repeated interaction
- 12:22 – 13:21
Polish and UX rigor: fixing layout shift with iterative prompting
They catch a subtle layout shift on refresh and immediately iterate with a small, targeted prompt to eliminate it. The segment highlights how “micro-feedback” and detail orientation translate into fast UI improvements without leaving the workflow.
- •Identifying layout shift as a quality regression
- •Using quick, specific prompts to address UI stability
- •Skeleton/loading behavior as a common polish area
- •AI-assisted iteration supports high design standards
- 13:21 – 15:26
Testing like an engineer: moving from dev preview to PR + Vercel preview deployment
Guillermo connects the dots between local-like dev confidence and production-like assurance via Vercel’s PR previews. v0 auto-generates the PR description, the Vercel bot deploys a preview, and they validate the change on production-grade infrastructure.
- •Dev environment is useful but PR preview adds stronger confidence
- •v0 generates a PR description automatically
- •Vercel bot produces an instant shareable preview URL
- •Preview runs on production-grade CDN/rendering for realistic validation
- 15:26 – 18:18
How Vercel uses v0 internally: measuring PRs merged and accelerating delivery
Guillermo shares internal metrics via a data assistant: PRs merged through v0 have surged to thousands per day. He also shows how easy it is to start from any existing GitHub repo by pasting a URL and continuing work conversationally.
- •Internal adoption tracked through “PRs merged via v0” metric
- •Rapid growth to high daily PR merges in early rollout
- •Workflow vision: trigger work from anywhere (phone/Slack/v0.app)
- •Import by pasting a GitHub URL to start iterating immediately
- 18:18 – 22:28
Org adoption and culture: letting marketers ship safely without bypassing review
They explore how v0 changes who can contribute while keeping engineering controls. Marketers can propose edits directly on real pages, but PRs, previews, and rollback keep changes responsible—shifting debate from prioritization politics to idea merit.
- •Before: marketers had to request changes or rely on CMS limitations
- •Now: anyone can propose changes via v0 on the actual codebase
- •Safety rails: PR workflow, previews, code review, rollback
- •Enterprise rigor needs feature flags, experiments, and measurement
- 22:28 – 25:18
Favorite non-coding AI workflows: image/video generation and automated social cards
Guillermo highlights non-coding uses: an internal image playground (Nanobanana) for memes, presentations, and rapid ideation, often paired with v0 for implementation. They also discuss automating OG/social card generation as a real productivity unlock.
- •Image generation as a daily utility for communication and storytelling
- •Parallel generation + selection workflow, then implement in v0
- •Emerging excitement about video generation
- •Automating bottlenecks like OG/Twitter card creation with agents
- 25:18 – 28:11
Showcase project: agent-vs-agent 3D chess match built with v0 and Vercel AI infra
They dissect a holiday project: a 3D, ESPN-style chess broadcast where two models play while streaming reasoning tokens. Guillermo explains how it leverages AI Gateway and durable workflows so the game can keep running despite provider or runtime failures.
- •Three.js 3D rendering generated/assembled via v0
- •Two models compete; reasoning tokens are streamed for transparency
- •AI Gateway enables model swapping and cost/usage reporting
- •Workflows provide reliability and retries for long-running processes
- 28:11 – 31:42
Teaching kids and “physical AI”: APIs, real-world displays, and typing debates
Guillermo shares experiences bringing kids to the office to build with a Vestaboard and teaching the concept of APIs—code that changes the physical world. They discuss the shift to speech-to-text and strategies for sneaking literacy/typing into games.
- •“Physical AI” as the next frontier beyond screens
- •Vestaboard demo to teach APIs and real-world actuation
- •v0 lowers the barrier by letting kids express intent in English
- •Typing and reading skill-building through game-like constraints
- 31:42 – 35:04
When AI gets stuck: escape hatches, model triangulation, and AI-built debugging tools
Guillermo explains practical troubleshooting: fall back to local dev, collaborate via Git, consult other models to learn missing domain concepts, and even ask v0 to generate debugging tools. He gives a concrete example from 3D asset manipulation and mesh visualization.
- •Escape hatch: clone locally or bring in collaborators via Git workflow
- •Use other models as “experts” to unblock domain-specific issues
- •Ask v0 to create debugging utilities (e.g., mesh/texture toggles)
- •Skills + smarter models reduce stuck states over time
- 35:04 – 43:35
Audience Q&A: validating ideas (customer zero/one), Sandbox, agents, GenUI, and App Store dreams
In Q&A, Guillermo outlines Vercel’s validation model: build for themselves (customer zero) and pressure-test with design partners (customer one). He explains Sandbox as the general-purpose compute powering v0, previews building durable agents/workflows, and discusses explorations in generative UI and eventual web-like App Store deployment.
- •PMF approach: internal taste + external design partners and community feedback
- •Sandbox VMs as a general-purpose environment (beyond Next.js)
- •v0 will help build reliable agents using workflows and background execution
- •Exploring “flash v0” generative UI and longer-term App Store deployment
