AI & DevelopmentOpen SourceDeveloper Tools

gstack: Garry Tan’s Open-Source Claude Code Team in 2 Commands

Y Combinator CEO Garry Tan claims he shipped 600,000 lines of production code in 60 days — solo — while running one of the most prominent startup accelerators in the world. The tool behind that claim is now free, open source, and installable in one command. gstack bundles 23 opinionated Claude Code slash commands into a virtual engineering team. Skip the hype for a moment: even if you discount the headline numbers entirely, the underlying workflow is the most practical AI coding setup available today.

What gstack Actually Does

Most developers who use Claude Code have some version of a custom CLAUDE.md file — a few lines instructing the model how to behave on their project. gstack is that idea taken to its logical extreme, made opinionated, and actively maintained.

The key insight is not the prompts themselves. It is that gstack encodes the process of good engineering into each role. When you run /office-hours, Claude doesn’t just help you plan — it adopts the mindset of a skeptical YC partner who assumes your feature idea is probably wrong. It asks six forcing questions and writes a design doc that every downstream command reads. The model stops being a helpful assistant and starts acting like someone whose job is to tell you no before you waste a week building the wrong thing.

That framing extends across the full toolkit. /review looks for bugs that pass CI but fail in production. /qa opens a real browser, walks the app like a user, and converts findings into regression tests. /land-and-deploy merges, waits for CI, and verifies production health before it declares success. Each command has a specific job and a specific point of view — which is what separates it from a list of random prompts.

The Commands Worth Knowing First

gstack ships 54 skills total. The five that matter most for getting started:

  • /office-hours — Start here. Forces you to articulate what you are building and why before writing code. Output is a design doc the rest of gstack references.
  • /autoplan — Runs CEO, design, DX, and eng reviews in sequence. Use after /office-hours to pressure-test the plan from multiple angles.
  • /review — Staff engineer code review before you merge. Flags regressions and risk, auto-fixes obvious issues.
  • /qa — Browser-based end-to-end testing. The model navigates your app and files bug reports.
  • /land-and-deploy — Handles the full deploy-and-verify cycle after a one-time /setup-deploy configuration.

Installing It Takes Two Commands

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack && ./setup

Requirements: Claude Code, Git, and Bun 1.4.2 or newer. After setup, type / in Claude Code and every gstack skill appears in autocomplete. Start with /gstack for an overview, then jump into /office-hours on whatever you are working on.

The November Update Added Guardrails

The v1.96.0.0 release on November 5 made one significant change: Claude Code can no longer invoke /land-and-deploy or /modernize autonomously. If you say “land it,” the model now points you to the command instead of running it. You have to type it yourself.

This is not a minor housekeeping update. It is a direct response to the agentic AI safety moment playing out across the industry — autonomous agents merging PRs, pushing to production, and causing irreversible damage without explicit human approval. The /careful command warns before destructive operations. /guard combines /careful with a branch lock. Root-level recursive deletes and default-branch force-pushes are hard-denied, not just warned about.

What to Actually Believe

The 600,000 lines in 60 days figure is Garry Tan’s own claim from the README. It is unverified. Lines of code is a notoriously poor measure of meaningful output — a significant portion of that count is likely scaffolding, tests, and generated boilerplate. Read the claim as “this setup dramatically accelerates output” rather than a benchmark to replicate.

Three other caveats worth knowing:

  • Claude Code only (mostly). Cursor, Kiro, and Codex CLI are listed as experimental. Safety skills are enforced only on Claude Code; on other hosts they are advisory.
  • Conductor dependency for peak throughput. Tan’s headline numbers depend on Conductor, a separate parallel-agent tool not included in gstack.
  • Not for regulated environments. The speed that makes gstack useful becomes a liability in banking, insurance, or defense contexts requiring multi-level human review.

Worth Trying

If you use Claude Code and you have not set up a structured workflow, gstack is the fastest path to one that works. You do not have to adopt all 54 commands. Start with /office-hours before your next feature and /review before your next merge. The workflow pays for itself the first time it catches an assumption you hadn’t questioned.

The MIT license means you can fork it, strip parts you don’t need, and add your own roles. The 66K stars and active weekly release cadence suggest this project is not going away. The fact that the YC CEO published his actual setup — not a demo, but the thing he runs every day — is at minimum a meaningful signal about where developer productivity is headed.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *