gstack: Garry Tan’s Claude Code Setup That Turns One Dev Into a Team

Garry Tan open-sourced gstack, his Claude Code configuration of 23 specialist slash commands, and the MIT-licensed project drew 141,000 GitHub stars within days of release. gstack adds a persistent cross-session memory layer called GBrain, managed by the /learn command, and version 1.57.6.0 introduces gbrain dream-stage caching that pre-computes likely next-session context in the background. Tan claims 3 production services and 40+ shipped features in 60 days while running Y Combinator full-time, though the article estimates real-world gains at 3x to 5x rather than his 810x productivity claim.

Garry Tan open-sourced his complete Claude Code configuration this week, and 141,000 developers starred it before most people finished reading the README. The project is called gstack , it ships 23 specialist slash commands, and it is the most complete public example of what a structured agentic coding workflow actually looks like in production. If you have been using Claude Code as expensive autocomplete, this is the missing workflow layer. Role Specialization, Not More Tools The core idea in gstack is deceptively simple: instead of one AI assistant doing everything, you give it explicit roles for each development phase. Tan’s framing is worth quoting directly — “Planning is not review. Review is not shipping. Founder taste is not engineering rigor. If you blur all of that together, you get a mediocre blend of all four.” gstack solves this by installing 23 skill files, each of which switches Claude Code into a different cognitive mode on demand. No new APIs, no custom tooling, no infrastructure — just structured Markdown files that change the model’s behavior. The key insight: different mindsets for different tasks, not more capabilities piled onto one. The Specialist Roster The 23 commands fall into four categories. Product and strategy starts with /office-hours — YC-partner interrogation mode that pressure-tests your scope before you write a line of code — and /plan-eng-review , which locks architecture and produces diagrams. Engineering gives you /review a paranoid staff engineer hunting N+1 queries, race conditions, and trust violations and /ship automated PR creation, test runs, sync . QA and security includes /qa for find-fix-verify cycles with real Playwright https://playwright.dev/ browser sessions, and /cso for OWASP Top 10 and STRIDE threat modeling. The standout is /codex , which runs Claude and OpenAI Codex in parallel and surfaces cases where they disagree — a cross-model signal that neither model provides alone. GBrain: The Part Most Reviewers Missed What separates gstack from a prompt dump is GBrain , the persistent cross-session memory layer managed by the /learn command. When an agent resolves a design debate or makes an architectural decision, GBrain logs it. The next sprint, agents start with that context intact rather than re-arguing closed questions from scratch. Version 1.57.6.0 adds gbrain dream-stage caching, which pre-computes likely next-session context in the background. It is a small feature with an outsized effect on workflow continuity across a multi-day project. Real Results and the 810x Distraction Tan claims 3 production services and 40+ shipped features in 60 days, running YC full-time alongside. A CTO who adopted gstack reported that /review caught a live XSS vulnerability on day one. The community debate has largely fixated on Tan’s 810x productivity claim — critics, reasonably, point out that “normalized logical lines of code per day” is a metric you can game. According to reporting on the community split https://newclawtimes.com/articles/garry-tan-gstack-claude-code-ai-coding-developer-debate/ , the backlash was partly driven by status YC CEO rather than technical merit. That debate is a distraction. The real value of gstack is coverage : one developer can credibly run strategy review, security audit, and QA in a single afternoon sprint without manually context-switching between four different mental modes. That is not 810x. It is probably 3x to 5x in real-world conditions. Still significant. Getting Started: Three Commands First gstack is MIT-licensed, free, and installs in one command. Requirements: an active Claude Code subscription, Git, and Bun v1.0+ https://bun.sh/ . The binary is 58MB with no node modules at runtime. git clone https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && \ cd ~/.claude/skills/gstack && ./setup Do not try to adopt all 23 commands at once. Start with three: /plan-eng-review before you write a feature, /review before you open a PR, /qa before you ship. Add the rest once those three are habit. The full gstack repository https://github.com/garrytan/gstack and a detailed breakdown of all 23 specialists https://lucaberton.com/blog/garry-tan-gstack-claude-code-ai-factory-2026/ are worth bookmarking.