// the find
garrytan/gstack
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA
A collection of Claude Code slash commands that simulate a virtual engineering team — CEO, designer, eng manager, QA lead, etc. — each implemented as a markdown prompt file that guides the AI through a specific sprint phase. Created by Garry Tan (YC CEO) and open-sourced as his personal AI workflow. The browser automation layer is the most technically substantial part; the rest is structured prompt engineering.
- The sprint pipeline (Think → Plan → Build → Review → Test → Ship) is genuinely well-structured — each skill writes output that downstream skills read, so context accumulates rather than resetting with every prompt.
- The browser automation in /browse and /qa is real code with a serious test suite (150+ tests in browse/test/), including adversarial security tests, prompt injection defense with an ML classifier, and CDP allowlist enforcement. Not a thin wrapper.
- The /pair-agent cross-agent browser coordination is novel — scoped tokens, tab isolation, and ngrok auto-tunneling for remote agents are actual engineering, not just a concept.
- The /careful and /freeze guardrails solve a genuine problem: they give you a practical way to scope AI edits to one directory or warn before destructive commands, which matters when running 10 parallel sessions.
- The core 'specialists' are markdown prompt files. The CEO, eng manager, and designer are the same model reading different SKILL.md instructions — there is no actual capability difference between them, just framing. The org chart metaphor is marketing.
- The productivity metrics in the README ('810× my 2013 pace', '240× the entire 2013 year') are self-reported and unauditable. The LOC controversy section exists because the numbers invite scrutiny, not because the methodology is sound.
- Heavy installation friction: requires Bun, Node.js, Claude Code, and the setup is a 'paste this into Claude' flow that trusts the model to run arbitrary git clone + shell commands. The 'works on 10 AI agents' claim requires per-agent setup that isn't well-tested.
- GBrain (persistent memory) adds Supabase or PGLite as a required dependency for what amounts to a searchable notes file — a Postgres instance to remember that you prefer more whitespace is an expensive abstraction.