How I AIGuillermo Rauch: Vercel CEO on how v0 hit 3,200 PRs merged per day (and lets anyone ship)
CHAPTERS
- 0:00 – 1:22
Vibe coding vs. production reality: iterating safely at scale
Guillermo frames the core problem with “vibe coding”: getting from zero to one is easy, but iterating, reviewing, and deploying safely at company scale is hard. He contrasts old workflows (petitioning engineers or relying on rigid CMS setups) with a new model where anyone can propose changes through v0 while retaining engineering rigor.
- •Zero-to-one demos are common; safe iteration and deployment are the real challenge
- •Non-engineers (e.g., marketing) constantly need website changes
- •Old process: ask engineers or hope the CMS supports the change
- •Goal: reduce friction while keeping production-grade safeguards
- 1:22 – 4:40
What is skills.sh and why it blew up
They introduce skills.sh as an open ecosystem for “skills” used to augment agents and AI applications. Guillermo describes it like an npm-style hub, notes rapid community adoption (tens of thousands of submissions), and emphasizes it was conceived and iterated using v0.
- •skills.sh is a hub for reusable AI/agent “skills” (installed with simple commands)
- •Massive community scale: ~34–35k skills and growing fast
- •Designed to be production-grade with instant transitions and scalability
- •Built via collaborative, prompt-driven iteration inside Vercel
- 4:40 – 6:40
From Git-backed project to new branch inside v0
Guillermo demonstrates v0’s GitHub-centric workflow by creating a new branch directly in the product. The interface mirrors established engineering conventions (contributor/feature naming) while removing the need for manual Git tooling, making branches and previews first-class in a chat-driven environment.
- •skills.sh is connected to Git; v0 reads the real repo state
- •“New Branch” creates a proper Git branch without local setup
- •Branch naming follows standard team conventions (user/feature)
- •v0 aims to embed real engineering workflow details into an easy UX
- 6:40 – 9:05
A full cloud dev environment in the browser (VS Code + running app)
They explore the in-v0 development environment: a VS Code-like editor, dependencies installed, and the actual Next.js app running in a dev/staging mode. Claire highlights how this eliminates setup pain (Homebrew, local tooling) and helps less-technical builders learn by working in a real environment.
- •In-browser VS Code editor paired with a live preview
- •Dependencies and exact framework versions are installed automatically
- •Runs a realistic Next.js dev server in a VM
- •Lowers the barrier for non-engineers to move from prototype to real code
- 9:05 – 11:18
Building a five-star rating system (with abuse prevention in mind)
Guillermo adds a rating feature to skills.sh with a compact, high-level prompt, while explicitly considering production concerns like abuse and rate limiting. v0 inspects the existing architecture and aligns with current data choices instead of inventing new infrastructure.
- •Feature request: sidebar five-star ratings for skills quality signals
- •Production concern: prevent abuse and add rate limiting/controls
- •v0 adapts to existing stack (e.g., Redis via Upstash)
- •UI matches the site’s “terminal core” aesthetic
- 11:18 – 13:20
Quick QA loop: persistence, refresh behavior, and layout shift fixes
They test the new UI in the preview, confirm ratings persist across refresh, and notice a subtle layout shift when data is absent. Guillermo prompts v0 to eliminate the jitter, emphasizing attention to detail as part of shipping-ready AI-assisted development.
- •Validate rating submission and data persistence
- •Refresh tests confirm storage works as expected
- •Identify and fix layout shift via quick follow-up prompt
- •Demonstrates iterative polish, not just initial generation
- 13:20 – 15:25
Creating a pull request and deploying a production-like preview
Guillermo walks through generating a PR from v0, including an auto-written PR description. Then he uses Vercel’s preview deployment to validate the change on production-grade infrastructure (CDN/rendering), building confidence before merge.
- •v0 generates PR content/description as part of the workflow
- •Vercel bot deploys a preview for the PR automatically
- •Preview runs on production-like infrastructure (vercel.sh environment)
- •Branch previews + review provide shippable confidence and rollback safety
- 15:25 – 17:48
How Vercel uses v0 internally: PRs merged per day and repo import
Guillermo shares internal metrics showing rapid adoption of v0-driven PR merges (surging to ~3,200/day in late January). He also shows how teams can import an existing GitHub repo into v0 by pasting the URL, making “prompting on real production code” a normal workflow.
- •Internal adoption: v0-assisted PR merges scaling dramatically
- •v0 turns prompting into a pathway to real merged code
- •Simple repo import: paste a GitHub URL to start working
- •Expectation: this changes how teams operate day-to-day
- 17:48 – 22:04
Organizational impact: democratizing changes without losing rigor
They discuss how v0 enables anyone—especially marketing and product—to propose improvements that are immediately reviewable and deployable. The PR/preview workflow removes the “prioritization humiliation ritual” while preserving enterprise practices like review, measurement, and cautious releases.
- •Non-engineers can propose real changes, not just prototypes
- •Git workflow ensures responsible shipping (review, preview, rollback)
- •Frees teams to debate the merit of ideas, not implementation feasibility
- •Enterprise reality: CTAs, experiments, and release hygiene matter
- 22:04 – 25:17
Favorite non-coding AI uses: image/video generation and internal tooling
Guillermo highlights image generation as a major productivity lever, including Vercel’s internal “Nanobanana” playground for parallel generations. They also discuss automating bottlenecks like Open Graph/social card creation by combining classic rendering with generative models, plus long-horizon research tasks.
- •Image generation for memes, presentations, and visual communication
- •Internal playground enables many generations in parallel, then selection
- •Automation of OG/social card creation to unblock launches
- •Growing excitement about video generation and research agents
- 25:17 – 27:57
AI chess match demo: 3D rendering, model-vs-model, and reliability
They dive into Guillermo’s v0-built 3D chess broadcast concept (ESPN-style), with two models playing and their thinking streamed live. The project showcases Vercel AI infrastructure (model switching via AI Gateway) and durable execution patterns that can keep experiences running despite failures.
- •v0-generated Three.js-style 3D experience with live match flow
- •Streams model reasoning tokens to show decision-making
- •AI Gateway enables swapping/benchmarking models and tracking cost
- •Workflow reliability: designed to keep running through provider hiccups
- 27:57 – 31:44
Teaching kids with AI: physical interfaces, APIs, and learning by building
Guillermo shares building with his kids using a Vestaboard to connect code to the physical world, teaching the concept of APIs. Claire relates with her own kid-friendly hardware projects and they discuss playful ways to encourage reading/typing amid increasing speech-to-text habits.
- •“Physical AI” projects make code feel real and motivating
- •Teaching APIs through a tangible display (Vestaboard)
- •Speech-to-text vs. typing: finding playful ways to teach fundamentals
- •Opportunity for high-quality educational games built with generative tools
- 31:44 – 34:43
When AI gets stuck: escape hatches, multi-model help, and self-made debug tools
Guillermo explains practical troubleshooting: rely on collaboration and Git workflows, switch models to learn missing concepts, and build debugging interfaces with AI assistance. He recounts solving a 3D asset issue (pieces stuck together) by consulting other models and adding visualization/debug toggles.
- •Reduce “stuck” moments via collaboration and Git-based handoff
- •Ask other models to learn domain concepts, then apply in v0
- •Create meta-debug tools (mesh/texture toggles) using AI itself
- •Skills + better frameworks reduce failures as models gain resources
- 34:43 – 43:35
Audience Q&A: validation (customer 0/1), Sandbox VMs, GenUI, and App Store dreams
In Q&A, Guillermo outlines how Vercel validates products via “customer zero” (internal taste/use) plus “customer one” design partners and community feedback. He explains v0’s underlying compute (Sandbox VMs) enabling broader stacks (e.g., React Native), explores “flash” generative UI research, and discusses the long-term goal of App Store deployment becoming as easy as web deploys.
- •Product validation model: customer zero (Vercel) + customer one (partners)
- •Community pressure-tests ideas (e.g., skills distribution/discovery)
- •Sandbox VMs provide powerful general-purpose compute for agents/tools
- •Exploration: on-the-fly GenUI vs. deployed apps; ambition: simpler App Store shipping
