
YC's CEO opens up his own "Claude Code" software factory: gstack packs a development team into a single command line.
Y Combinator's Garry Tan shares his daily workflow, which includes using 23 different commands to manage a virtual team of roles like CEO, reviewer, QA specialist, security officer, and release engineer – all while leveraging free MIT resources.
Related tags
YC Founder Opens His Claude Code Software Factory: gstack Turns a Development Team Into a Slash Command Pipeline
When you tell Claude Code, “I want to build a daily report app,” gstack will first challenge your idea – “What you’re describing is essentially a personal AI assistant.” This conversation, driven by product intuition, comes from Garry Tan, the CEO of Y Combinator, who has recently released his personal workflow: gstack. It’s a collection of 23 slash commands that turn a development team into an “AI software factory,” with MIT license, free, no paid tiers, and over 300 stars in just one day.
Not Just a Collection of Tools, But a Sprint Pipeline
The core principle of gstack is written in the README: Turn Claude Code (and other coding agents) into a virtual engineering team. The commands are organized in a sprint-like order – Think → Plan → Build → Review → Test → Ship → Reflect – with each step’s output feeding into the next: Design documents written in /office-hours are read by /plan-ceo-review, testing plans written in /plan-eng-review are taken over by /qa, and bugs found by /review are verified by /ship.
The roles are clearly defined: /office-hours simulates YC’s office hours, using six mandatory questions to reverse engineer your product framework; /review is a senior engineer who specializes in catching bugs that “pass CI but fail in production”; the /qa QA manager uses a real browser to reproduce issues, writes regression tests, and fixes them; /cso the security officer performs OWASP + STRIDE audits; and /investigate has a strict rule: no investigation, no fix. If a bug is not fixed after three attempts, the process automatically stops.
The Governance Details Are More Interesting Than the Features
What’s truly remarkable about gstack is that it incorporates “governance for AI work” into its core: /captain allows one controller to run multiple coding agents in parallel, only sending READY blocks without direct merging; merging requires a verified merge train; /runaway-stopper specifically intercepts runaway agent threads that consume tokens; /eval-plan requires upfront registration of budget and stop-loss rules. All outbound activities, including “external review” where code is sent to Codex for review, are logged in the gstack-egress list audit report.
Persistent memory is handled by the accompanying GBrain knowledge base (/setup-gbrain), which allows you to accumulate project patterns and bug fixes across sessions. The team mode uses a single command to guide the entire repository, with teammates automatically receiving gstack; stable versions are released every Tuesday and Thursday, at most once every 72 hours, plus verified security hotfixes.
It’s Not Just Claude Code
gstack supports installing multiple tools: OpenAI Codex CLI, Cursor, Gemini CLI, GitHub Copilot CLI, OpenClaw, Hermes, and others all have corresponding files (full versions have been certified by real workflows, experimental versions have undergone consistency tests, and instruction-only versions simply print instructions). Installation takes about 30 seconds with git clone and ./setup.
The founder of a leading VC opens his personal factory – which includes Claude Code – for everyone to see. This act has significance that is no less than the code itself: when everyone is talking about how much code AI agents can write, Garry Tan’s answer is: first, give the agent a team with a defined process, auditing, and safety measures.
Repository address: https://github.com/garrytan/gstack
More News
Want to see more quality content?
Browse all news