May 25, 2026·5 min read

gstack: Y Combinator's President Open-Sourced His Own Claude Code Setup

TL;DR

gstack packages 23 slash commands into Claude Code that act like a virtual engineering team: CEO, designer, eng manager, release manager, doc engineer, and QA roles.
It crossed 10,000 GitHub stars in its first 48 hours and sat at over 94,000 stars two months after release.
Garry Tan claims it let him ship 600,000+ lines of production code in 60 days while running YC full time, though that number leans on a metric his own team had to redefine to make credible.
It's built for people who want structure imposed on them. Experienced engineers who already know what to build tend to find the planning overhead slows them down instead.

Y Combinator’s president spends his working hours evaluating other people’s startups, not shipping his own code. In March 2026, Garry Tan published his personal Claude Code configuration on GitHub anyway, and it picked up 10,000 stars before most of the industry had finished their morning coffee.

That speed is the story before any feature list is. A toolkit built by an investor for his own use, released for free, outgrew most funded developer-tool startups’ entire lifetime star count in two days.

What is gstack?

gstack bundles 23 opinionated slash commands into Claude Code, each one standing in for a role a real engineering team would normally fill: /office-hours for early-stage design discussion, /plan-eng-review for turning a plan into a test strategy, /qa for verification, /ship for release. Tan describes the effect as turning a single developer into a virtual software team, and the commands are built to hand work to each other in sequence rather than operate as isolated tools.

It isn’t Claude Code exclusive. The same command set works across seven other AI coding agents, which matters more than it sounds like: the workflow, not the specific model behind it, is the actual product.

Why did it go viral this fast?

Part of the answer is who published it. Tan was an early member of Palantir’s founding team and later co-founded Initialized Capital before taking over as YC’s president in 2023, so when he says a workflow works, thousands of founders he’s funded are inclined to believe him before reading a single line of the README. The other part is that this wasn’t packaged as a product pitch. It was his own daily setup, refined through his own use, published as-is.

Tan backs the release with a specific productivity claim: 600,000-plus lines of production code shipped in 60 days, alongside running YC full time. He’s careful to frame the more defensible number as logical code changes rather than raw line count, since AI-assisted coding is well known for inflating line counts without adding proportional value. By that stricter measure, he claims a 2026 run rate of roughly 11,417 logical lines a day against 14 a day in 2013, a jump he states as around 810 times. That’s Tan’s own accounting of his own work, not an independently audited figure, and it’s worth reading with the same skepticism due to any self-reported metric this large.

What does /office-hours actually do?

/office-hours runs first, at the point where a developer has an idea for a feature but hasn’t specified it yet. Describe a new API endpoint to it, and it produces a design document. That document then feeds /plan-eng-review, which turns the design into a test plan, which /qa picks up later to verify the implementation actually meets it. The chain is the point: each command reads what the previous one produced instead of starting from a blank prompt.

Who should skip it?

The same structure that helps a non-expert avoid missing steps can slow down someone who already knows exactly what they’re building. Planning-heavy commands like /office-hours and /plan-ceo-review consume a meaningful chunk of context and tokens before any code gets written, and a user without a clear direction already in mind can burn through a session’s token limit in the planning phase alone.

Experienced engineers report a specific frustration: repeated planning loops that never quite convert into shipped code, because the tool keeps asking for more startup-style validation than the task actually needed. Anyone billing tokens by usage, or anyone who’s already run this exact kind of planning process in their head a hundred times, is likely to find the overhead more costly than the structure is worth.

How do you install it?

Point Claude Code at https://github.com/garrytan/gstack and ask it to install the toolkit globally, or read through the repository’s own README for a more selective, manual setup. Either path assumes Claude Code is already installed and working in a terminal.

gstack works best as training wheels for people who’ve never run a software project before and want the discipline of a real team without hiring one. For anyone who already has that discipline, the training wheels just get in the way.