Y Combinator (YC) president Garry Tan has open-sourced the entire development environment he works in. It's called gstack. It's only been on GitHub a short while, and it has already passed 54,000 stars.
Tan goes back to the days when Coinbase, Instacart, and Rippling were one- or two-person teams. He was an early engineer, PM, and designer at Palantir, then founded Posterous and sold it to Twitter. Today he runs YC full-time while writing 10,000 to 20,000 lines of production code a day on the side — more than 600,000 lines in 60 days.
How one person works like a team of 20. He's just released the tool that makes it possible.
What is gstack?
It's a collection of slash commands that turns Claude Code into a "virtual engineering team." Twenty-eight specialized roles are defined in Markdown files, each one handling a single stage of the development process.
This isn't just a set of prompt templates. It's the entire sprint process, structured end to end.
Think → Plan → Build → Review → Test → Ship → Reflect
Each stage's command reads the output of the one before it. /plan-ceo-review reads the design doc that /office-hours wrote; /qa runs the test plan that /plan-eng-review produced; /ship confirms the fixes for bugs that /review caught. Nothing falls through the cracks between stages.
What roles are there?
The core idea of gstack is to assign the AI an "expert role." Instead of agonizing over what to put in a blank prompt, you summon the specialist that fits the situation.
Planning. /office-hours works like a YC office-hours session. Tell it "I want to build a daily briefing app" and it reframes the request: "What you've actually described is a personal-assistant AI." It challenges assumptions, offers alternatives, and writes a design document.
/plan-ceo-review re-examines the feature from a CEO's point of view, with four modes: expand, optionally expand, hold scope, or cut. /plan-eng-review draws out the architecture, data flow, edge cases, and test matrix as ASCII diagrams. /plan-design-review scores each design dimension from 0 to 10, explains what a 10 would look like, and then makes the fixes itself.
Build and review. /review hunts for the bugs that pass CI but blow up in production. It auto-fixes the clear-cut ones and asks about anything requiring judgment. /investigate does systematic root-cause debugging and stops after three failed fix attempts. It has a built-in "iron rule" that forbids jumping to a fix before investigating.
Testing. /qa is the tool that made the biggest difference, Tan says. It opens a real browser, clicks around, finds bugs, fixes them, generates regression tests, and verifies the fixes. That tool, he says, is what let him scale from 6 parallel workers to 12.
Shipping. /ship handles syncing the main branch, running tests, auditing coverage, pushing, and opening a PR — all in one go. If there's no test framework, it builds one from scratch. /land-and-deploy takes you from merging the PR through waiting on CI, deploying to production, and running health checks, all in a single command.
Security and safety. /cso runs the OWASP Top 10 and a STRIDE threat model. It has 17 false-positive exclusion rules and reports only findings with a confidence of 8/10 or higher. /careful warns you before destructive commands like rm -rf, DROP TABLE, or force-push. You activate it by saying "be careful."
Docs and retros. /document-release reads every documentation file in the project, compares it against the changes you just shipped, and automatically updates anything that's gone stale. /retro runs a weekly retrospective showing commit count, lines of code, test health, and opportunities to grow.
How to install it
It takes 30 seconds. With Claude Code already installed, just paste in this command.
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup
Add a gstack section to your project's CLAUDE.md and you can start using it right away. To share it with teammates, copy it into the project folder and commit it.
It works not only with Claude Code but also with OpenAI Codex, the Gemini CLI, and Cursor. Just install with ./setup --host codex or ./setup --host auto.
A real workflow, step by step
- /office-hours — you say, "I want to build a calendar-based daily briefing app."
- The AI pushes back: "What you've described isn't a briefing app — it's a personal-assistant AI." It extracts 5 core features, challenges 4 assumptions, and proposes 3 implementation approaches.
- /plan-ceo-review — reads the design doc and reviews scope across 10 sections.
- /plan-eng-review — generates a data-flow diagram, state machine, error paths, and test matrix.
- Approve the plan and it writes 2,400 lines across 11 files. About 8 minutes.
- /review — auto-fixes 2 issues, finds a race condition, and asks for approval.
- /qa https://staging.myapp.com — tests in a real browser, finds and fixes bugs.
- /ship — tests go from 42 to 51 (+9 new), and a PR is created.
Eight commands, from idea to deployment.
Running in parallel: 10 to 15 sprints at once
gstack's real power is in parallel execution. In one session you're refining a new idea with /office-hours; in another you're reviewing a PR with /review; in a third you're testing staging with /qa. Six, ten, even fifteen at the same time.
Parallelism works because there's a process. Run ten agents with no process and you get ten kinds of chaos. With a defined order — Think → Plan → Build → Review → Test → Ship — each agent knows what to do and when to stop. It's the same way a CEO manages a team: sign off on the important decisions, and let the rest run.
Features worth a closer look
Real browser control. /browse opens an actual Chrome browser so you can watch the AI click and type in real time. A faint green glow at the top of the screen tells you which window gstack is controlling. If a CAPTCHA or MFA prompt comes up, you can hand off with $B handoff to solve it yourself and return with $B resume.
Cross-model review. /codex gets you an independent code review through OpenAI's Codex CLI. When /review (Claude) and /codex (OpenAI) review the same branch, you get a cross-model analysis comparing what each AI found.
Design exploration. /design-shotgun generates several AI design variants and opens a comparison board in the browser, iterating until you pick a direction. It has a built-in "taste memory" that remembers your preferences.
Guardrails. /careful, /freeze, and /guard provide, respectively, blocking of destructive commands, restriction of edit scope, and a full safe mode. /investigate automatically freezes the module it's looking into.
Who should use it
Tan names three audiences. Founders and CEOs who still want to write code themselves. Claude Code beginners who need structured roles instead of a blank prompt. And tech leads and senior engineers who want systematic review, QA, and release automation on every PR.
There's a takeaway for non-developers too. What gstack demonstrates is that the effectiveness of an AI tool depends less on "how well you write prompts" and more on "how well you structure your process."
When Andrej Karpathy (an OpenAI co-founder) said he hadn't written a single line of code by hand since December, Garry Tan wanted to know "how?" gstack is his answer — proof that one person with the right tools can move faster than a team of 20. The tool is free and MIT-licensed.
