← Back to archive

Repo of the Day

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

Published: Oct 2, 2026

Open repository ↗

Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA - garrytan/gstack

Summary

gstack is an MIT-licensed, TypeScript-distributed collection of slash-command skills that turn a Claude Code session into a role-based engineering workflow. The README documents 23 "specialist" skills — including /office-hours, /plan-ceo-review, /plan-eng-review, /review, /qa, and /ship — that run in a Think → Plan → Build → Review → Test → Ship → Reflect sequence, each skill passing artifacts to the next. It targets founders, tech leads, and first-time Claude Code users who want opinionated review, QA, and release steps instead of an empty prompt.

What it is useful for

  • Reframing a product idea before code. /office-hours runs six "forcing questions" against the user's pain; the README's worked example shows a user asking for a "daily briefing app" and the agent reframing it as a "personal chief of staff AI," extracting capabilities and recommending the narrowest wedge to ship first.
  • Locking architecture and design. /plan-eng-review produces ASCII data-flow, state-machine, and error-path diagrams; /plan-design-review scores each design dimension 0–10 and edits the plan toward a 10, including AI-slop detection.
  • Review, test, and ship. /review auto-fixes obvious issues and flags completeness gaps, /qa opens a real browser against a staging URL (or local API, CLI, job, or webhook) to reproduce bugs and write failing regressions, and /ship produces a PR.
  • Onboarding teammates. A team-mode bootstrap commits a .claude/ hook that requires gstack for AI-assisted work in the repo, with an auto-update check throttled to once per hour.
  • Running inside other agents. Claude Code is the only "full" tier. OpenAI Codex CLI, OpenCode, Cursor, Factory Droid, Kiro, and GitHub Copilot CLI ship as "experimental"; Slate, OpenClaw, Hermes, and GBrain are "instruction-only" via a 2KB AGENTS.md digest.

How engineers can use it

  1. Install prerequisites. The README requires Claude Code, Git, and Bun v1.0+; Node.js is required only on Windows. On macOS 15+, the Aside browser is recommended for browser skills; without it, ./setup builds a bundled Chromium.
  2. Run the documented install. From Claude Code, paste the command from the README:
git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack
./setup

Then add the "gstack" section to CLAUDE.md as the README specifies so future sessions load the skills. 3. Walk the sprint. Run /office-hours on an idea, then /plan-ceo-review and /plan-eng-review on the design doc, implement, then /review, /qa <staging-url>, and /ship. 4. For shared repos, the README documents a team-mode command that bootstraps a required (or optional) install and commits the hook so teammates auto-update. 5. For other agents, ./setup --host <name> installs per host; --host auto installs for every detected agent, and --status shows the current state.

Documented limitations to plan around: /cso is currently static-assessment only because no qualified runtime/scanner profile is published, and its project-test completion stays self_reported. Codex, OpenCode, Cursor, Factory Droid, Kiro, and Copilot are experimental with advisory (not enforced) safety skills. Copilot ignores skill hooks, so /careful and /freeze only advise there, and COPILOT_HOME other than ~/.copilot is not supported yet. Full host details live in docs/ADDING_A_HOST.md.