Introduction
gstack is the open-source Claude Code setup that Y Combinator CEO Garry Tan published in March 2026, packaging his personal AI-coding workflow as 23 opinionated, reusable skills. Instead of typing a fresh prompt each time, you invoke slash commands — each one casting Claude Code in a role such as CEO, designer, engineering manager, release manager, doc engineer, or QA — so one developer can drive what looks like an entire engineering team.
The repository (github.com/garrytan/gstack, MIT-licensed) had gathered roughly 121,000 stars and 18,000 forks as of July 2026, making it one of the most-starred repositories tied to Claude Code. It also drew, in TechCrunch's words, "so much love, and hate." This post explains what gstack actually is, the tools inside it, how to install it, and why it split the room.
What is gstack?
gstack is not a new application. It is a configuration layer on top of Claude Code, Anthropic's terminal-based coding agent. Concretely, it is a collection of Claude Code skills — Markdown files that live in ~/.claude/skills and define custom slash commands. Each skill is a structured, opinionated set of instructions that tells Claude how to behave for a specific task, which is exactly the pattern Anthropic formalized when it shipped Agent Skills.
Because the skills are just text, gstack is fundamentally a prompt engineering artifact — a curated library of prompts, wrapped as commands, that encode a founder's whole software-development lifecycle. Tan's framing is that "the skills are the prompts": once the commands exist, you stop hand-writing instructions and start orchestrating roles.
The 23 tools: gstack's virtual company of roles
The organizing idea is a "virtual company." Rather than treating Claude as a generic assistant, gstack gives it named jobs and switches between them. The repository's tagline advertises 23 opinionated tools that serve as CEO, designer, engineering manager, release manager, doc engineer, and QA — with a security officer added through the /cso (chief security officer) command. In total, though, the repository documents more than 45 distinct slash commands: beyond the 23 headline tools, the wider skills directory ships power tools, supporting utilities, and platform-specific commands (for example, an iOS set). Those 23 core tools are the spine.
In practice, this turns solo development into a pipeline of hand-offs, each one an AI agent playing a part:
- CEO / founder decides whether a feature or idea is worth building.
- Designer proposes and reviews the interface.
- Engineering manager plans the change and reviews the approach.
- QA lead tests the result and hunts for regressions.
- Release manager ships, canaries, and documents the deploy.
The result is an agentic workflow: a chain of specialized steps with review gates between them, rather than one open-ended "build this" request.
gstack's slash commands: from /office-hours to /ship
The skills surface as slash commands you type inside Claude Code. Reported commands include:
/office-hours— a founder-style gut check on what to build./plan-eng-review— plan a change and review it before code is written./design-shotgun— generate several design directions at once./qaand/qa-only— run quality-assurance passes on the work./ship,/canary, and/land-and-deploy— take a change to production, gradually./careful,/freeze, and/unfreeze— throttle or lock down risky work./cso— a security review from a chief-security-officer perspective./codex— a cross-model second opinion, calling out to a different model to review Claude's output.
The /codex command is worth flagging: it deliberately brings a second, non-Claude model in to critique the work, an acknowledgment that one model reviewing itself has blind spots. Several commands also lean on browser and tooling integrations in the spirit of the Model Context Protocol (MCP), the open standard for connecting agents to external tools and data. If you want the general playbook these commands operationalize, Anthropic's own Claude Code best-practices repository covers the underlying habits.
How to install gstack
Installation assumes you already have Claude Code set up. From there, gstack is a one-command install that clones the repository into your Claude skills directory and runs its setup script:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup
The setup registers the skills so the slash commands become available in Claude Code, and a /gstack-upgrade command pulls later updates. Because it drops into the standard ~/.claude/skills path, nothing about gstack is proprietary — you can read every skill as plain Markdown, edit it, or delete the ones you dislike.
From a senior-engineer prompt to a whole toolkit
gstack is best understood as the successor to an earlier, smaller artifact. In February 2026, Tan shared a single senior-engineer prompt for Claude Code — one structured plan-mode instruction that forced the model to review architecture, code quality, tests, and performance before writing anything. That prompt did one job well: reviewing a single change like a staff engineer.
gstack generalizes that idea across the whole lifecycle. Where the senior-engineer prompt was one carefully worded message you pasted in, gstack is a library of commands covering planning, design, implementation, QA, security, and release. This is the practical meaning of "I don't prompt anymore": the discipline that used to live in a long prompt now lives in named, reusable skills. If you only want the review step, the earlier prompt is still the tighter tool; gstack is for people who want the entire assembly line.
Is gstack over-engineered? The criticism
gstack was polarizing, and the criticism is worth taking seriously. TechCrunch reported (March 17, 2026) that the setup drew love and hate in roughly equal measure. Skeptics argued it is "a bunch of prompts in a text file," that experienced Claude Code users already maintain their own equivalents, and that copying a founder's personal configuration wholesale is cargo-culting rather than engineering. It is, by design, opinionated: it reflects Tan's stack and taste, leans toward TypeScript web and iOS work, and may not map cleanly onto a Python data pipeline or an embedded project.
Tan's own case for it rests on productivity claims that should be read as his claims, not established facts: he says that in a recent 60-day stretch he shipped 3 production services and 40+ features part-time while running YC, and that his 2026 code output runs at roughly 810 times his 2013 pace. Impressive if accurate, but self-reported and dependent on how you count "logical lines." The honest takeaway is narrower and more useful: gstack is a well-documented, working example of how to turn a coding agent into a role-based team — a reference to learn from and adapt, not a setup to adopt unread.
Conclusion
gstack is Garry Tan's Claude Code setup made public: 23 opinionated skills that recast Claude as a small engineering organization, installable with a single command. Its real value is not the specific commands but the pattern it demonstrates — encoding an entire development lifecycle as reusable AI agent skills, with review and security gates built in. Treat it as a template. Read the skills, keep the ones that fit your stack, rewrite the rest, and remember that the productivity numbers attached to it are one founder's account rather than a benchmark.
For other named practitioners' approaches to agentic coding, see the Ralph technique and Harper Reed's LLM codegen workflow.