Y Combinator CEO Garry Tan just open-sourced gstack — a collection of 23 Claude Code slash commands that turn a single developer into a virtual engineering team. CEO product review, design critique, code review, QA with real browser testing, security audits, and automated releases — all from Markdown skill files.
The Premise: One Builder, Team-Scale Output
Tan’s pitch is backed by his own numbers: 3 production services and 40+ shipped features in 60 days, part-time, while running YC full-time. He cites Andrej Karpathy’s admission of not typing code since December 2025 and Peter Steinberger building OpenClaw (247K GitHub stars) essentially solo with AI agents.
The claim is not that AI writes perfect code. The claim is that with the right role specialization — giving the LLM different personas for different tasks — one person can cover the functions that traditionally require a team of designers, reviewers, QA engineers, and release managers.
The 23 Specialists
gstack organizes Claude Code into specialized roles via slash commands:
Product and Strategy:
/office-hours— describe what you are building, get structured feedback/plan-ceo-review— product-level review of any feature idea/plan-eng-review— architecture and engineering review/plan-design-review— design critique before implementation
Design:
/design-consultation— interactive design session/design-shotgun— rapid design exploration across multiple directions/design-html— generate production-ready HTML/CSS
Engineering:
/review— rigorous code review on any branch/ship— prepare changes for merge/land-and-deploy— automated release workflow/canary— canary deployment validation/benchmark— performance testing
QA and Security:
/qa— opens a real browser against your staging URL/browse— web browsing skill for research and testing/connect— integration testing
Each command is a Markdown file in ~/.claude/skills/gstack/. No dependencies beyond Claude Code, Git, and Bun.
Why Role Specialization Works
The key insight is not giving Claude more tools — it is giving Claude different mindsets for different tasks. When you run /review, Claude acts as a senior engineer looking for production bugs. When you run /plan-design-review, it acts as a designer looking for AI slop and UX inconsistencies.
This mirrors how Karpathy’s CLAUDE.md works — behavioral constraints in Markdown that shape LLM output. gstack takes it further by creating an entire roster of specialized behaviors instead of a single set of principles.
The architecture is simple:
- Skill files are Markdown documents with role descriptions and behavioral rules
- Slash commands activate specific skill files
- Claude Code reads the active skill and adjusts its behavior accordingly
- No custom tooling, no API wrappers, no infrastructure
What This Means for Solo Builders
The 810x productivity claim (normalized logical lines of code per day) is provocative and debatable. But the underlying pattern is real: AI coding agents with structured roles let individuals cover more surface area than ever before.
For founders, this means:
- Pre-funding: Ship an MVP without a team. Use
/plan-ceo-reviewto pressure-test the product,/reviewfor code quality,/qafor testing. - Early stage: One technical founder can maintain the velocity of a small team while focusing on customers and fundraising.
- Growth stage: Use gstack as a template for how your team structures AI-assisted workflows.
How to Install
# Clone into Claude Code skills directory
git clone --single-branch --depth 1 \
https://github.com/garrytan/gstack.git \
~/.claude/skills/gstack
# Run setup
cd ~/.claude/skills/gstack && ./setupThen add gstack to your project’s CLAUDE.md to make the slash commands available.
My Take
The individual tools are useful, but the real value is the pattern: decomposing the software development lifecycle into specialized AI roles, each with explicit behavioral constraints. Whether you use gstack directly or build your own skill library, the approach of role-based LLM specialization is becoming the standard for AI-assisted development.
The MIT license means you can fork it, strip what you do not need, and add your own roles. I would start with /review and /qa — those cover the highest-value gap for most solo developers.
Related Reading
- Karpathy’s CLAUDE.md: 4 Rules That Fix LLM Coding
- GitHub Copilot Agent Mode
- AI Coding Agents and Platform Engineering
- Context Architecture for AI Agents
- GBrain: Garry Tan’s Self-Wiring Knowledge Graph for AI
Master Claude Code
Want to build your own gstack-style workflow? These resources will get you there:
- 🎓 Claude Code Masterclass (Free on Udemy) — 13 lectures covering CLAUDE.md, hooks, MCP servers, and autonomous coding patterns
- 📘 Claude Code Companion Book (Leanpub) — Deep dive into enterprise patterns, team workflows, and production Claude Code setups
- 📝 5 Claude Code Mods That Make It Production-Ready — Ralph Loop, Ultracode, Context7, and more
