gstack: Y Combinator CEO’s AI Programming Team OS
This week’s most talked-about open source project on GitHub isn’t a new framework or a new model — it’s a skill set that lets your AI coding assistant switch cognitive modes on demand.
gstack, open-sourced by Y Combinator President & CEO Garry Tan, accumulated over 18,000 stars in just one week, making it one of the fastest-growing projects on GitHub this week.
The Problem It Solves
Anyone who has used Claude Code or similar AI coding tools knows the frustration: the AI is too “one-size-fits-all”.
Ask it to do product planning and it starts writing code. Ask it for a code review and it starts brainstorming new features. Planning, reviewing, shipping, testing — these require completely different mindsets, but AI always responds in the same mushy generic mode.
Garry Tan’s solution: give the AI different “brains” and switch between them on demand.
12 Expert Roles, Full Development Lifecycle
gstack provides 12 carefully designed slash commands, each corresponding to a specialist role:
| Command | Role | Core Capability |
|---|---|---|
/plan-ceo-review | Founder / CEO | Look beyond the literal request, find the real product |
/plan-eng-review | Eng Manager | Lock in architecture, data flow, edge cases, test matrix |
/plan-design-review | Senior Designer | 80-item design audit, AI Slop detection |
/review | Paranoid Staff Engineer | Find bugs that pass CI but blow up in production |
/ship | Release Engineer | Sync main, run tests, push, open PR — all in one go |
/browse | QA Engineer | Give AI eyes to click through and test your app |
/qa | QA Lead | Read git diff, auto-identify affected pages and test them |
/retro | Engineering Manager | Analyze commit history, generate team retrospectives |
/document-release | Technical Writer | Auto-update README and architecture docs post-release |
The Killer Feature: Giving AI Eyes
The most groundbreaking feature in gstack is /browse and /qa.
The biggest blind spot of traditional AI coding assistants: they can’t see what your app looks like. They have to guess about UI state, auth flows, and page layouts.
gstack solves this with a built-in persistent Chromium instance powered by Playwright:
18 tool calls, ~60 seconds, a complete QA pass.
AI Slop Detection: Does Your Site Look AI-Generated?
/plan-design-review includes a feature that hits close to home for many developers: AI Slop Detection.
It identifies 10 telltale signs of AI-generated websites:
- Blue-to-purple gradient heroes
- 3-column icon grids
- Uniform bubbly border-radius on everything
- Centered text on every section
- Decorative floating blobs in the background
One Person, Ten Parallel AI Agents
Garry Tan describes his own workflow in the README:
“This is the setup I use. One person, ten parallel agents, each with the right cognitive mode for its task. That is not incremental improvement. That is a different way of building software.”
Combined with Conductor, you can run multiple Claude Code sessions simultaneously:
- One session running
/qaon staging - Another doing
/reviewon a PR - A third implementing a feature
- Seven more working on other branches
Each workspace gets its own isolated browser instance — no port collisions, no shared state.
Installation
Just paste one command into Claude Code:
Claude handles the rest automatically.
Requirements: Claude Code, Git, Bun v1.0+
Why This Project Matters
gstack’s viral success isn’t just about its features — it represents a new AI programming paradigm:
- Specialization over generalization — different tasks need different mindsets
- AI needs perception — an AI that can “see” your app is a complete AI
- Workflow is the product — great AI tools integrate into your flow, not interrupt it
- Solo teams become viable — with AI expert roles, individual developers can have full team capabilities
As Garry Tan puts it: “This is not a prompt pack for beginners. It is an operating system for people who ship.”