Skip to content
Search

Gstack

lastmod 2026-04-12created 2026-03-233 edits

A virtual engineering team built with Claude Code

About

  • An open-source software factory made of AIs, built and used by YC CEO Garry Tan, set up so that one person works like a 20-person team
  • Made of slash commands covering the whole sprint in the order Think → Plan → Build → Review → Test → Ship → Reflect; each skill automatically passes context to the next skill
  • Start with /office-hours to validate product assumptions, lock in the architecture with /plan-ceo-review · /plan-eng-review, then automate everything from bug fixes to PR creation with /review · /qa · /ship
  • Each command acts like a role-specific expert: CEO review, Eng manager design, designer audit, QA lead, release engineer, etc.
  • Works not only in Claude Code but in any agent that supports the SKILL.md standard, such as Codex, Gemini CLI, and Cursor
  • MIT license

Target users

  • Founders·CEOs — technical founders who still want to ship code themselves
  • Claude Code beginners — users who need a structured, role-based workflow instead of a blank prompt
  • Tech Leads·Staff Engineers — senior engineers who need rigorous review, QA, and release automation on every PR

Core sprint skills

  • /office-hours - validates product assumptions with 6 forcing questions in YC office-hours format, generates a design doc and automatically passes it to downstream skills
  • /plan-ceo-review - reframes the problem and explores the '10-star product'; 4 modes: Expansion·Selective Expansion·Hold Scope·Reduction
  • /plan-eng-review - defines architecture, data flow, ASCII diagrams, edge cases, test matrix, and security concerns
  • /plan-design-review - rates each design dimension from 0 to 10 and explains what a 10 looks like, includes AI Slop detection, runs interactively with one AskUserQuestion per design decision
  • /design-consultation - builds a complete design system from scratch and generates realistic product mockups
  • /review - finds bugs that pass CI but blow up in production, auto-fixes obvious issues, flags completeness gaps
  • /investigate - iron rule: no fix without investigation; traces data flow, tests hypotheses, stops after 3 failures
  • /design-review - runs the same audit as /plan-design-review, then fixes the issues found directly, with before/after screenshots
  • /qa - tests the app in a real browser, finds and fixes bugs, auto-generates a regression test for each fix
  • /qa-only - same methodology as /qa but only produces a bug report, with no code changes
  • /cso - OWASP Top 10 + STRIDE threat model audit; 17 false-positive exclusion rules, a confidence gate of 8/10 or higher, a concrete exploit scenario for each finding
  • /ship - syncs main, runs tests, audits coverage, pushes, creates the PR; auto-bootstraps a test framework if there is none
  • /land-and-deploy - PR merge → wait for CI·deployment → verify production health, all in one command
  • /canary - post-deployment monitoring loop for console errors, performance regressions, and page failures
  • /benchmark - measures baselines for page load time, Core Web Vitals, and resource sizes, with before/after comparison per PR
  • /document-release - updates all project docs to match what was shipped, auto-detects stale READMEs
  • /retro - weekly retrospective; per-person breakdowns, shipping streaks, test health trends; /retro global for a unified retrospective across all projects and AI tools (Claude Code, Codex, Gemini)
  • /browse - a real Chromium browser, real clicks, real screenshots, about 100ms per command
  • /setup-browser-cookies - imports cookies from Chrome·Arc·Brave·Edge into the headless session to test authenticated pages
  • /autoplan - automatically runs CEO → design → engineering review in order, surfacing only taste decisions to the user

Power tools

  • /codex - independent code review from the OpenAI Codex CLI; 3 modes: review (pass/fail gate)·adversarial challenge·open consultation; cross-model analysis when both /review and /codex are run
  • /careful - warns before running destructive commands such as rm -rf, DROP TABLE, and force push; enabled by typing "be careful"
  • /freeze - restricts file edits to a specific directory, preventing out-of-scope changes while debugging
  • /guard - /careful + /freeze combined, the maximum safety setting for production work
  • /unfreeze - removes the /freeze boundary
  • /setup-deploy - one-time setup for /land-and-deploy, auto-detects the platform, production URL, and deploy commands
  • /gstack-upgrade - upgrades gstack to the latest version, detecting and syncing both global and vendored installs

Parallel sprints

  • With Conductor, multiple Claude Code sessions can run simultaneously in isolated workspaces
  • E.g. one session on /office-hours, another on /review, a third implementing a feature, and a fourth on /qa, all at the same time
  • Founders, lead engineers, and PMs can fork the AI development factory to experiment with and extend it themselves

See also

  • Claude Code
  • GBrain - a personal knowledge management (Knowledge Brain) tool built by YC CEO Garry Tan in the style of Karpathy's LLMWiki
  • Everything Claude Code - a hackathon winner's recipe for an AI development team

Favorite site

author Zer0 sitemap