← all repositories
garrytan/gstack

A YC CEO's 23-person AI engineering team, compressed into slash commands

Garry Tan open-sourced the exact Claude Code prompts he uses to ship 810× faster while running Y Combinator full-time.

108.1k stars TypeScript Coding AssistantsAgentsLLMOps · Eval
gstack
Velocity · 7d
+1224
★ / day
Trend
steady
star history

What it does

gstack installs 23 specialist “skills” into Claude Code (and 9 other AI agents). Each skill is a Markdown prompt that plays a role — CEO, eng manager, designer, QA lead, security officer, release engineer — invoked with a slash command like /office-hours or /ship. The skills chain together: a design doc written by /office-hours gets read by /plan-ceo-review, which feeds /plan-eng-review, which sets up tests for /qa to run against a real browser.

The interesting bit

The value isn’t the prompts themselves — anyone can write “act as a senior engineer.” It’s the process glue. Tan explicitly designed this as a sprint pipeline (Think → Plan → Build → Review → Test → Ship → Reflect) where each stage consumes the previous stage’s output. The /office-hours skill is the unusual one: it interrogates your product idea with six forcing questions before any code is written, and the README shows it reframing a “daily briefing app” into a “personal chief of staff AI” — then generating the implementation plan.

Key highlights

  • 23 role-based slash commands plus 8 “power tools” (browser control, canary deploys, security audits, etc.)
  • Multi-agent support: Claude Code, OpenAI Codex, Cursor, OpenCode, and 6 others via auto-detected ./setup
  • Team mode: adds a .claude/ bootstrap to repos so teammates auto-install gstack without vendoring files
  • OpenClaw integration: native skills that dispatch to Claude Code sessions for coding tasks, or run directly as chat skills
  • MIT license, all Markdown, no external services or API keys required

Caveats

  • Requires Claude Code (or another supported agent), Git, and Bun v1.0+; Windows needs Node.js too
  • The 810× productivity claim is based on “logical lines of change” vs. raw LOC, with methodology documented but not independently verified
  • README is truncated in the source; full skill descriptions and some setup details are cut off

Verdict

Founders and tech leads who already use Claude Code and want structured, repeatable AI-assisted workflows should try the 30-second install. If you’re looking for a no-code tool or don’t use AI coding agents, this is just a folder of Markdown prompts.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.