Y Combinator · Video summary

Inside Garry Tan's Claude Code Setup

Leverage AI agents like Claude Code with a structured team approach (roles, process, review) to build software 10x faster, moving from idea to deployable product with AI-assisted planning, coding, and QA.

Y Combinator · April 23, 2026 · 22-min episode · 5-min read

Key takeaways

💡The agent era of software development requires treating AI like a team with defined roles, process, and review to achieve real work, not just wandering models.00:47
🚀Integrate AI agents into your workflow by defining specific roles and processes, treating them as a team rather than just a tool.00:50
🚀The 'thin harness, fat skills' approach, exemplified by GStack, uses minimal scaffolding to turn AI models into a team of specialists.02:17
🤔The 'Office Hours' skill, distilled from YC's founder interactions, forces critical product reframing by asking pointed questions (e.g., 'What's the strongest evidence someone wants this?') before building.02:31
🔧Utilize AI 'Office Hours' or similar reframing tools to critically evaluate your product ideas and business models before significant development effort.02:31
💰Successful startups often employ a 'wedge strategy', starting with a narrow hook (e.g., aggregating 1099s) to solve an immediate pain, then expanding to a more lucrative service (e.g., tax preparation matchmaking).06:36

Chapters

0:09🎯 Introduction and Background
0:56🎯 The Agent Era and GStack
1:20🎯 Experience with Claude Code
1:52🎯 Challenges of Raw AI Models
2:12🎯 Introduction to GStack and Skills Approach
2:33🎯 Using Office Hours Skill
3:52🎯 Practical Startup Idea Development
4:50🎯 Model Reasoning and Web Search
7:57🎯 Browser Automation with GStack
10:51🎯 Building and Automating Workflow
12:29🎯 The Power of AI in Startup Prototyping
13:04🎯 Adversarial Review and Design Tools
14:17🎯 Creating Visual Brainstorming Outputs
18:01🎯 Scaling and Managing Multiple Projects
18:58🎯 Automating QA with Claude in Chrome
19:30🎯 Optimizing Development Workflow
20:24🎯 Clicking and Managing Ideas
21:28🎯 The Future of Building with AI

The digest

High-Level Overview Garry Tan, president and CEO of Y Combinator and a veteran software engineer, presents a detailed walkthrough of his AI-powered development system called GStack, built on Claude Code—an AI coding agent platform. The video illustrates how AI can be harnessed as a collaborative engineering team with specialized roles, dramatically accelerating software creation. This approach is designed for engineers, startup founders, and technologists eager to leverage AI agents to handle complex development tasks, reducing manual coding while maintaining quality through structured processes and review.

The Agent Era in Software Development

Garry Tan emphasizes that the future of building software lies in the agent era, where AI agents function similarly to human teams with distinct roles, processes, and review cycles. Unlike traditional coding, this approach uses AI specialists (skills) that collaborate to build, plan, and review software.

Tan reflects on his experience coding Posterous, which took two years, 10 engineers, and $10 million to build. Using Claude Code and GStack, he has now recreated that functionality within months mostly autonomously.
The key challenge is not the intelligence of AI but the scaffolding and process around it. AI tends to "wander" or make plausible but fragile guesses without proper structure.