Key Moments
How to Make Claude Code Your AI Engineering Team
Y CombinatorKey Moments
An open-source tool called GStack leverages AI agents like Claude Code to automate software engineering tasks, significantly accelerating development but requiring careful management of AI 'skills' for effective team collaboration.
Key Insights
Garry Tan coded more in the past two months than in all of 2013, highlighting the accelerated development pace with AI coding tools.
GStack has achieved over 70,000 GitHub stars, surpassing established tools like Ruby on Rails, indicating strong community adoption.
The 'Office Hours' skill in GStack is a distilled version of Y Combinator's partner sessions, designed to pressure-test startup ideas with six forcing questions before coding begins.
GStack's 'design shotgun' feature, utilizing OpenAI's DALL-E for image generation, can produce multiple visual design options for a UI element in about 60 seconds.
The `SLQA` and `SL browse` tools, built around Playwright and Chromium, automate browser interactions, testing, and QA, addressing a bottleneck Garry experienced with AI-driven development.
Garry Tan currently manages 10-15 parallel Claude code sessions and has approximately 400 open PRs to review, illustrating the scale of parallel development enabled by GStack.
The agent era demands a team-based approach to AI development
Garry Tan, president and CEO of Y Combinator, introduces GStack, an open-source toolkit designed to harness the power of AI agents, specifically Claude Code, by treating them as a cohesive engineering team. Tan emphasizes that the 'agent era' of building software requires a structured approach mirroring human teamwork, complete with defined roles, processes, and review mechanisms. He shares his personal experience, noting he has written more code in the past two months than in the entirety of 2013, highlighting the dramatic increase in development velocity enabled by AI. GStack's core philosophy is the 'thin harness, fat skills' model, providing a lightweight framework to manage specialized AI capabilities. The project has rapidly gained traction, accumulating over 70,000 GitHub stars, outperforming established tools like Ruby on Rails, underscoring the significant interest and potential of this AI-driven development paradigm.
Office hours: pressure-testing ideas like a YC partner session
One of GStack's foundational skills, 'Office Hours,' is modeled directly after the rigorous sessions YC partners conduct with startups. This skill is designed to critically evaluate a product idea before any code is written, employing six forcing questions to reframe the concept and identify potential weaknesses. Tan demonstrates this by simulating an office hours session for a tax app idea designed to extract 1099s from Gmail. The AI, in 'Gary mode,' reveals its reasoning process, showing how it searches for context, identifies pain points, and probes the viability of the idea. A key question it poses is, 'What's the strongest evidence you have that someone actually wants this?' This deep dive aims to uncover the true user need and business model potential, moving beyond a superficial solution to a more robust and scalable offering. The process simulates the personalized, adversarial feedback that founders receive at YC, helping to refine the idea and identify its core value proposition.
Navigating the 'adverse review' and conceptualizing business models
The simulated office hours session prompts a deep dive into the startup idea, revealing its potential as a wedge strategy. While the initial idea focuses on aggregating tax documents, the AI points towards a more lucrative expansion: matchmaking and lead generation for tax preparers. This represents a classic business model evolution, moving from a low-margin service ($2-5 per year) to a potentially 10x higher revenue stream based on transaction percentages. The AI challenges the premise, questioning why existing solutions like TurboTax or Plaid aren't sufficient, and pushes for a more comprehensive understanding of the user's needs. The interaction highlights how AI can act as a critical thinking partner, uncovering hidden opportunities and potential pitfalls. Tan even notes that the process is so engaging and insightful that he might build the app for himself simply for the learning experience, emphasizing that GStack fosters a conversational and exploratory approach to development.
Browser automation as an unexpected solution
During the office hours session, a critical consideration emerges: how to access and download tax documents. While initial thoughts might involve complex integrations like Plaid, the AI proposes a more unconventional but effective approach using browser automation, integrated into GStack's 'browser' skill. This method involves the user logging into their account, after which the AI takes over, navigating to tax document portals, locating the 1099s, and downloading the PDFs. A key advantage is that the process occurs within the user's visible browser, not in the cloud, enhancing transparency and control. This solution bypasses the need for stored credentials or deeper API integrations, relying instead on simulating user actions. Tan expresses surprise and admiration for this approach, noting that such creative solutions might not have been considered even a few months prior, underscoring the evolving capabilities of AI in software development.
Iterative design with 'design shotgun' and AI-generated previews
After the initial idea refinement in office hours, GStack offers tools for rapid design visualization. Tan demonstrates 'design shotgun,' a feature that leverages OpenAI's codecs and image generation capabilities to produce multiple visual mockups of a user interface in approximately 60 seconds. For the tax app, three distinct design directions are generated: a 'command center friendly' version, a more 'progress-oriented' friendly view, and a 'split view' option. Tan reviews these options, scoring them and selecting 'Option B' for its user-friendly, card-based approach with progress indicators. The ability to quickly iterate on visual design and receive AI-generated feedback significantly accelerates the design phase, allowing developers to lock in a direction before committing to extensive coding. This rapid prototyping capability is crucial for agile development, enabling quick adjustments based on AI-driven insights and user preferences.
Automating QA and the role of Codex as the 'autistic CTO'
Tan distinguishes between the creative, idea-generating aspect of AI, which he likens to an 'ADHD CEO' (like Claude Opus 4.6), and the meticulous, detail-oriented execution needed for coding. For the latter, he advocates for tools like Codex, describing it as the 'autistic CTO' – essential for debugging and ensuring code quality. GStack incorporates this by offering a 'review' skill that performs staff-level bug catching after code generation. Furthermore, Tan developed specialized tools like `SLQA` and `SL browse`, which wrap Playwright and Chromium. These tools enable AI agents to perform complex browser interactions, take screenshots, run regression tests, and identify real-world browser issues, mirroring human QA processes. This automation is crucial, as Tan found himself spending excessive time on manual QA, a task he deems the 'least fun part of software development' once AI handles the majority of coding.
Scaling parallel development and managing a high volume of PRs
GStack is designed to support a highly scalable development workflow, enabling users like Tan to manage an enormous volume of work. Tan reveals he runs 10 to 15 parallel Claude Code sessions simultaneously, often working across multiple open-source projects concurrently. This leads to a substantial number of open Pull Requests (PRs), currently around 400, which he reviews in waves. The tools within GStack, including office hours, adversarial review, and automated QA, streamline the process of evaluating and integrating community contributions. He emphasizes GStack's role in mitigating risks, particularly concerning supply chain attacks, providing a layer of security and integrated review. The platform transforms the development process from a sequential to a highly parallel one, allowing for rapid iteration and efficient management of a large-scale open-source contribution pipeline. This allows him to potentially handle 10, 15, 20, or even 50 PRs per day, depending on his meeting schedule.
The future of software development: collapsing barriers
Garry Tan concludes by asserting that we are witnessing the most incredible time in history to build software, with the barriers to entry drastically reduced by AI tools like GStack. The core challenge is no longer the technical difficulty but rather identifying 'what to build.' He encourages developers to leverage these advancements, stating, 'It's time to let it rip. Go make something people want.' The GStack toolkit, available on GitHub, offers a practical gateway into this new era, providing AI-powered capabilities that mimic the rigorous product thinking and development processes found at Y Combinator, democratizing access to sophisticated engineering support.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●People Referenced
GStack Workflow for Building with AI Agents
Practical takeaways from this episode
Do This
Avoid This
Common Questions
GStack is an open-source tool that transforms AI coding models like Claude Code into an AI engineering team. It uses specialized 'skills' and a 'thin harness fat skills' approach to manage AI agents, enabling them to perform complex software development tasks.
Topics
Mentioned in this video
GStack is an open-source repository Gary built to function as an AI engineering team, using specialized skills and following a thin harness fat skills approach. It has gained significant traction on GitHub.
Codex is described as the 'autistic CTO' that complements Opus 4.6, suitable for when complex tasks require focused execution.
Plaid connects to banks and is mentioned in the context of why existing solutions aren't fully solving the 1099 aggregation problem.
GStack has more GitHub stars than Ruby on Rails, highlighting the rapid growth and interest in agent-based software development.
Conductor is the platform through which GStack is accessed, and it provides features like quick start and Gary mode for visualizing AI reasoning.
The tax app demo uses Gmail to find 1099 forms and other tax documents.
Playwright is a tool wrapped by GStack at the CLI level to enable browser automation, including screenshots, complex interactions, and downloading media.
Chromium is mentioned as the browser that GStack's CLI tool, built around Playwright, interacts with.
Opus 4.6 is described as the default Claude model, characterized as an 'ADHD CEO' with many ideas but needing the 'autistic CTO' (Codex) for when the going gets tough.
Gary built the first version of Bookface, which is Y Combinator's internal social platform and knowledge base.
Gary co-founded Posterous, a microblogging platform that was sold to Twitter. He also mentions building it with AI agents took a fraction of the time compared to the original manual build.
GStack has accumulated more GitHub stars than Ruby on Rails, indicating its popularity and adoption.
Posterous, a microblogging platform co-founded by Gary, was sold to Twitter.
Gary, the speaker, is the president and CEO of Y Combinator; the organization's internal social platform and knowledge base, Bookface, was built by him.
HR Block is mentioned as a service that has 1099 import features, but it's implied that these don't fully solve the user's problem.
OpenAI's codecs are used by GStack for image generation tasks during the design phase.
More from Y Combinator
View all 582 summaries
44 minHow Stripe Built Their New Website
50 minRobots Are Finally Starting to Work
21 minThis Startup Wants To Catch Cancer Before It Spreads
32 minThis AI Company Catches Fraud Across the Internet
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Get Started Free