Inside Garry Tan's Claude Code Setup

The speaker, Gary, president and CEO of Y Combinator, and an experienced engineer, introduces a new era of software development: the agent era. He highlights the shift in how software is built, emphasizing that agents can perform real work when organized as a team with defined roles, processes, and review mechanisms. To facilitate this, he developed GStack, an open-source tool that has quickly gained significant traction, surpassing Ruby on Rails in GitHub stars within three weeks. Gary explains that he has coded more in the past two months than in all of 2013, driven by his exploration of Claude Code since January. He was inspired by figures like Andre Karpathy and Boris Churnney, who reported no longer manually writing code. Through this process, he essentially rebuilt his previous startup, Posterous, which originally took two years, $10 million, and a team of 10 engineers. He notes that out-of-the-box models tend to "wander" and guess, leading to plausible but silently broken code. The bottleneck isn't the model's intelligence, but rather the need for proper setup. GStack addresses this with a "thin harness, fat skills" approach, turning Claude Code into an AI engineering team. One of GStack's key features is "Office Hours," a skill modeled after YC's partner-startup interactions. It begins by posing six critical questions to help users reframe their product ideas before building. Gary demonstrates GStack using Conductor, an interface where GStack is integrated. He proposes building a tax app that fishes out 1099s from Gmail and financial institutions. During the demonstration, the "Office Hours" skill guides the user through the ideation process. Gary highlights "Gary mode," which shows the model's reasoning traces, offering transparency into its thought process. The model starts by asking a fundamental question: "What's the strongest evidence that you have that someone actually wants this?" Gary shares his personal experience with the pain of hunting down 1099s, confirming the problem's existence. The model then challenges the idea, pointing out existing solutions like TurboTax, HR Block, and Plaid. This interaction leads to refining the business model. The model suggests a "wedge strategy": initially finding 1099s to solve an immediate pain, then expanding to offer tax preparation matchmaking and lead generation for tax preparers. This approach is deemed more viable than simple document aggregation, potentially generating significantly more revenue. Gary emphasizes that Office Hours is not a rigid, on-rails process but a conversational interaction with the model, allowing for deeper exploration of user needs, business models, and pain points, which would otherwise be overlooked if direct instructions were given. The conversation further refines the technical approach, proposing AI browser automation where the user logs in, and the AI navigates to tax documents and downloads PDFs, with the user observing the process. This avoids storing credentials in the cloud and utilizes the user's actual browser. Gary also mentions GStack's ability to use codecs to debug and fix issues, likening Claude (Opus 4.6) to an "ADHD CEO" for its numerous ideas, while Codex is the "autistic CTO" for tough situations. The Office Hours session concludes with three potential approaches for the tax app. The first is a small-effort, low-risk Gmail search for tax notifications. The second, a full-stack Gmail and AI browser automation with a CPA marketplace, aligns with Gary's desired scope. The third flips the go-to-market strategy to CPA-first. Gary prefers the second approach but suggests an enhancement: using browser interaction to skip Google OAuth entirely, allowing GStack browser to directly search Gmail for 1099s and simultaneously ask the user about other banks. This process can also identify existing CPAs from emails. Gary notes that Office Hours helps transform a rough idea into a more refined concept, even if the initial idea isn't a guaranteed startup success. The feasibility aspect of Office Hours is highlighted, where the model mirrors Gary's strong opinions on what might work in the real world. Following Office Hours, GStack performs a multi-step adversarial review, stress-testing the idea and automatically identifying and fixing issues like lack of failure handling, privacy sections, or 2FA handoff solutions. This process improved the design document's score from 6 to 8 out of 10, addressing 16 issues. Next, Gary demonstrates "Design Shotgun," a visual brainstorming tool. It generates multiple AI versions of a chosen view, like the main checklist dashboard, and asks for feedback. Three design directions are presented: "command center," "friendly progress," and "split view." Gary reviews them, preferring the "friendly progress" option (Option B) for its accessibility to normal users, rating it five stars. Gary concludes by mentioning that these are just two of 28 commands available in GStack. He notes that users often spend 80-90% of their time in Office Hours, Plan CEO Review, and Auto Plan. Auto Plan offers default recommendations for CEO, engineering, design, and developer experience reviews, for those who prefer less back-and-forth. After code is built, a "review" skill provides staff-level bug catching. A significant development is the CLI built around Playwright and Chromium, enabling a full-featured headed and headless browser within GStack. This allows agents to use the browser for complex interactions, screenshots, filling forms, downloading media, running regression tests, and updating CSS. This feature addresses the bottleneck of manual QA, which Gary found himself doing once agents handled planning, design, and coding. He compares this to achieving a "level seven" software factory, enabling parallel work on multiple projects and features. Finally, the "ship" tool ensures a PR is ready to land on main. Gary describes his workflow: running 10-15 parallel Claude Code sessions, managing multiple open-source projects, reviewing hundreds of PRs, and addressing bug reports or new ideas by creating new work items in Conductor. He emphasizes GStack's role in mitigating supply chain attacks, stating he no longer has a traditional to-do list. Gary encourages users to try GStack, available on GitHub, highlighting that running "/off hours" provides a distilled version of YC's product thinking. He concludes by stating that this is an incredible time to build software, with collapsed barriers, urging people to "make something people want."