A practical guide to gstack, the open-source Claude Code skill pack from Garry Tan (President and CEO of Y Combinator). Not a marketing pitch. The actual install, the actual commands, what each skill does, where it shines, and where it doesn't. If you want a 5-person product team in your terminal for free, this is the install path.
The problem nobody talks about
Solo builders and small teams hit a ceiling. You can write the code, but you can't also be the staff engineer reviewing your own PR. You can't be the designer who challenges your wireframes. You can't be the QA that catches the real-browser regression. So you either hire (slow, expensive, often wrong people) or you ship slower than you should.
AI agents fix some of this, but most builders roll their own workflows from scratch. Bespoke prompts. One-off scripts. Whatever sticks. The result is inconsistent quality and a lot of repeated work.
gstack ships the workflows. 23 specialized skills, organized into 8 stages of the product lifecycle, each tested in production by Garry Tan across 40+ public and private repos. You install it once. The agent runs the team for you.
Why this changes everything
- One install, full team. 23 specialized AI roles in one repo. The eng review, the design review, the staff engineer audit, the release engineer, the QA in a real browser. All shipped, all free.
- Verified by usage. Tan didn't write gstack as a demo. He uses it across 40+ public and private repos, and his 2026 contribution rate is over 1,200 commits year to date.
- Open standard. gstack follows the Agent Skills format, so the skills work in Claude Code, Cursor, Codex, OpenClaw, and other AI coding tools.
- Free, MIT, no paywall. No premium tier, no per-skill cost, no upsell. The full pack is on GitHub.
Step 1: install gstack
Three things on your machine first:
- Bun v1.0+ (the JavaScript runtime gstack runs on)
- Node.js
- Git
- Claude Code (the agent gstack plugs into)
If you have those, install gstack:
git clone https://github.com/garrytan/gstack.git
cd gstack
./setupThe setup script registers all 23 skills with your Claude Code install. 30 seconds, give or take. Once registered, type /in any Claude Code session and you'll see the gstack commands listed.
Step 2: the 8 workflow stages
gstack organizes the 23 skills into 8 stages of how shipping actually happens. Think of it as a senior engineer's mental model, encoded.
Think. Before you build anything, interrogate the idea. /office-hoursruns the product interrogation. It asks the questions a YC partner would ask in a 1-on-1. What problem are you solving? Who's the user? What's the smallest test?
Plan. Once the problem is clear, you plan the build. /plan-eng-review runs an engineering review on your spec. /plan-ceo-review runs a business review. /plan-design-review covers UX. /plan-devex-review covers developer experience.
Design. /design-consultation walks you through the design space. /design-shotgun generates multiple variants in parallel and ranks them. /design-html turns the winner into rendered HTML.
Build. This is where Claude Code does its normal thing, with all the gstack context loaded. Your CLAUDE.md plus the gstack skills means the agent already knows the conventions and the constraints.
Review. /review is a staff engineer doing a PR audit. /design-review is a designer auditing the rendered output. They flag things you missed and suggest specific fixes.
Test. /qa runs real-browser tests via Playwright with anti-bot stealth. /benchmark runs performance benchmarks. /canary runs a canary check on the new version against the old.
Ship. /ship is a release engineer. It packages, tags, and prepares the release. /land-and-deploy runs the actual deploy.
Reflect. Once it's shipped, /retro runs a retrospective. /investigate runs a debugging session if something broke. The output feeds back into the next /office-hours session.
Step 3: the power tools
Beyond the 8-stage workflow, gstack ships a few power tools that don't slot neatly into any single stage.
/browse. Real Chromium browser via Playwright with anti-bot stealth. Use it when you need the agent to actually render a page, click buttons, take screenshots. Most coding agents fake browser access. /browsedoesn't.
/pair-agent.Multi-AI coordination. Run two agents on the same problem with different mental models, see where they agree and where they don't.
/codex. Cross-model code review. The same PR gets reviewed by Claude AND OpenAI Codex. Two perspectives, one PR. Disagreements between them flag the actual risky parts.
/cso. Security audit. Reviews your changes for common security holes, surfaces things a security-focused engineer would catch.
Step 4: parallel sprints with Conductor
This is the feature most builders won't appreciate until they need it. gstack ships with Conductor, a parallel session manager. You can run 10 to 15 concurrent Claude Code sessions, each on a different feature or bug. Conductor handles the coordination, prevents merge conflicts via dependency tracking, and surfaces a unified view of what's in flight.
Use case: you're a solo founder shipping a v1. You have a frontend bug, a backend feature, a copy revision, and a database migration. With Conductor, you spin up four parallel sessions, give each one its scope, and walk away. Twenty minutes later, all four are ready for review.