What This Project Does

gstack is a complete Claude Code configuration that Garry Tan, the president of YC, made public. It packs 23 opinionated tools and skills, and by opinionated the project means each one carries strong defaults instead of asking you to decide everything up front. The core idea is to let the agent play different roles, CEO, designer, engineering manager, release manager, documentation engineer, and QA, so that a single solo AI session becomes a virtual team that divides work and collaborates.

At its heart it is a process, not a pile of tools. The skills flow in the order of a real sprint, following the line Think, plan, build, review, test, ship, reflect, and each step's output is consumed by the next. For example, the design doc written by /office-hours is read by /plan-ceo-review. The chain covers product definition, design, engineering review, QA, all the way to release and retro, and it also ships extras like design mockups, docs, PDF and diagram generation, browser QA, web scraping, spec writing, and cross session memory. In practice you are not installing a feature, you are adopting a way of running projects where the boring middle steps are handled for you. The 23 skills are not a grab bag, they are a curated set that assumes you want a repeatable routine rather than one off tricks.

What Problem It Solves

After AI coding agents went mainstream, a common pain showed up, weak engineering discipline, a gap between requirements and implementation, and hard acceptance checks. Early agents were good at writing the next function but poor at keeping a product coherent across a whole build. gstack closes that gap with a structured product development process, and the appeal is that discipline is built into the commands, so you get the benefits of a process without hiring the people.

/office-hours uses six probing questions to reshape your product positioning. /plan-ceo-review re-examines scope from a CEO view. /plan-eng-review locks down architecture and tests. /qa actually opens a browser to click through and fix bugs. /ship verifies the fix and opens a PR. It moves the agent from mere code completion to pushing a whole project forward like a team. For a similar skill based way of thinking, see What Is mattpocock/skills and How to Use It, which explains breaking a workflow into reusable skills.

Language and Architecture

  • Main language is TypeScript. The skills ship as slash commands to Claude Code and many hosts, including Copilot, Antigravity, Gemini CLI, and OpenClaw. Each host's setup lives in docs/hosts.md, and a 2KB agents digest is provided for agents that follow read only rules.
  • Skills are wired together through artifact files. Design docs, test plans, and specs sit in $GSTACK_STATE_ROOT/projects/$SLUG/, and downstream skills read them automatically.
  • It bundles a Chromium browser with a sidebar, anti scraping stealth, and automatic model routing, and supports Aside browser integration. Standalone gstack-* CLI tools cover egress audit, context billing, model benchmarks, review and evidence ledgers, plus an iOS QA daemon.
  • Output tools include /make-pdf, which renders Mermaid and excalidraw offline to vector graphics, /diagram, which produces mermaid source, excalidraw, and an SVG/PNG triple, and /design-html, which generates 30KB dependency free production grade HTML that ships as is. Because the artifacts live in a predictable folder, you can inspect or even hand the intermediate files to another tool, which keeps the system open instead of locked in.

License

MIT

Garry Tan is the president of YC, so the project arrived with high attention. It rides the 2026 wave of agent skills and the AI software factory, turning a celebrity backed product engineering process into a ready to use skill pack. The repo has roughly 140,000 stars in total and gained about 612 stars that day, a very large base that keeps drawing new eyes. For a developer still figuring out AI collaboration, this copy the homework style of starting point has a very low cost.

Where to Find It

https://github.com/garrytan/gstack

Getting Started

Install the config into Claude Code first, then pick a small real project on your desk to practice. Do not point it at a production database on day one. Start with /office-hours to get product positioning clear, then move through /plan-ceo-review and /plan-eng-review so scope and architecture settle first. By the /qa and /ship stages it really opens a browser to click through, so run the first pass in a test environment and confirm permissions and routing before going further. If you are not on Claude Code, check docs/hosts.md for your host's setup, since the same skill files adapt to each environment. Spend a few runs watching which artifacts land in the state folder, because understanding that handoff is the fastest way to trust the system. Give it two or three small tasks before anything serious, because watching the handoff between steps teaches you where the guardrails are.

Summary

gstack upgrades "one person telling an AI to write something" into "a virtual team moving through a process". For individuals or small teams that want engineering discipline out of an AI coding agent, it removes the work of building the process and the rules themselves, and it is an open source config worth a try. The bet is that most solo builders fail not for lack of coding help, but for lack of a steady process, and this config supplies exactly that.

FAQ

Do I need to know TypeScript to use it? No. The config is written in TypeScript, but you only use its skills by following the slash commands, so being able to run a CLI is enough.

Is it only usable inside Claude Code? Not exactly. It supports many hosts like Copilot, Antigravity, Gemini CLI, and OpenClaw, and the exact setup is in docs/hosts.md.

Where do the artifact files live, and can they get corrupted? Design docs, test plans, and specs live in $GSTACK_STATE_ROOT/projects/$SLUG/, read automatically by downstream skills, and cross session memory lives there too, so it is best not to edit that directory by hand.