If an AI programming tool's configuration file can garner 96,900 stars, it signals two things: first, it's genuinely cool; second, many people are actively searching for the "right" way to work with AI programming.
gstack, open-sourced last week by Garry Tan (CEO of Y Combinator), is exactly this kind of configuration.
What Is It
gstack is not a framework, not a library, and not a tool. It is a collection of configuration files—Garry Tan's exact setup for Claude Code, made public.
The project's README features a single line:
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA.
23 tools, 6 roles.
6 Roles, 23 Skills
gstack's design philosophy is intriguing. Instead of organizing skills by tech stack (frontend, backend, database), it organizes them by team roles:
CEO Role – Responsible for overall project direction and priority assessment. This isn't a joke. There's a skill in gstack that prompts Claude Code to ask itself before writing any code: "Does this feature really need to be built? What problem does it solve?" It sounds a bit meta, but in AI programming scenarios, the ability to prevent an Agent from doing useless work is far more important than the ability to write code.
Designer Role – Handles UI/UX decisions, including color schemes, layout principles, and responsive design standards. AI coding tools often generate functionally correct but visually disastrous code; the Designer role fills this gap.
Eng Manager Role – Oversees code quality, architectural decisions, and technical debt management. This skill requires Claude Code to consider scalability, maintainability, and team conventions before writing code.
Release Manager Role – Manages version control, release workflows, and changelogs. It ensures every commit has a purpose and every release is documented.
Doc Engineer Role – Handles documentation. This skill forces Claude Code to automatically generate or update relevant documentation after writing code.
QA Role – Responsible for testing, including unit tests, integration tests, and edge case testing.
Why 96,900 Stars
gstack didn't go viral because of technological innovation. Its technical complexity doesn't even match many projects with a few thousand stars. It went viral for sociological reasons:
Celebrity Effect. Garry Tan is the CEO of YC and one of Silicon Valley's most influential figures. Open-sourcing his Claude Code setup is noteworthy in itself.
The Value of Being "Opinionated". The project description deliberately uses the word "opinionated." This means the configurations aren't neutral, "anything-goes" defaults; they are specific choices Garry Tan has made based on years of experience. For developers figuring out how to best use Claude Code, these strongly opinionated setups are far more useful than open-ended tutorials.
Resonance of Role-Based Design. Organizing Agent skills by role rather than by technology isn't a brand-new idea (RPG class systems work similarly), but this is its first large-scale application in the AI programming tool space. It solves a real problem: When faced with an AI that knows everything, you don't know what to make it focus on.
Actual Structure
The gstack repository structure is quite straightforward:
agents/– Agent configuration directory, with one agent file per roleautoplan/– Auto-planning configurationsbenchmark/andbenchmark-models/– Benchmarking relatedbin/– Command-line toolsbrowse/– Browsing/search configurations- etc.
Recent commits (from 4 hours ago) are still updating: v1.37.0.0 feat: split-engine gbrain (remote MCP brain + local PGLite ...). This shows the project is iterating rapidly.
278 commits for a configuration project is substantial, indicating that Garry Tan and his team are continuously optimizing these setups.
Key Feature: gbrain
Recent commits reveal that gstack is introducing a feature called gbrain, which uses a "split-engine" architecture: a remote MCP brain + local PGLite storage.
This means gstack isn't just static configuration files; it has an intelligent decision-making layer. gbrain likely coordinates and makes decisions across multiple agents—for example, deciding whose opinion to follow when the CEO agent and Eng Manager agent disagree.
How to Use It
Installing gstack should be straightforward (refer to the project documentation for specifics). The general idea is to integrate it into your Claude Code configuration, after which the skills for each role will automatically activate in their respective contexts.
Don't enable them all at once. Loading all 23 skills simultaneously will significantly increase the Agent's context overhead. Recommendations:
- Start with the 2-3 roles you need most (e.g., Eng Manager + QA)
- Use them for a while to experience the actual impact of each skill
- Gradually enable other roles as needed
Industry Signals
The emergence of gstack reflects a trend: the use of AI programming tools is shifting from "personal experimentation" to "standardized configurations."
Just as every programmer accumulates their own IDE settings, dotfiles, and code templates, in the AI programming era, every developer will accumulate their own Agent skill configurations. gstack is among the first "dotfiles for AI" to be shared at scale.
By open-sourcing his own configuration, Y Combinator's CEO is setting a precedent for the entire industry. In the coming months, we'll likely see more "XX's AI Programming Setup" repositories appear.
96,900 stars is not the finish line, but the starting point. When AI programming configurations become as standard for developers as dotfiles, projects like gstack will define the best practices for this field.