When the CEO of Y Combinator ships 40 features and 3 production services in 60 days, part-time, while running the most influential startup accelerator in the world, the natural question is not how he did it. It is whether a solo founder with the same tooling could do it too. Garry Tan just answered that question by open-sourcing gstack, his complete personal Claude Code workflow and the repo already has 122,610 GitHub stars and 18,331 forks. It is one of the fastest-growing developer tool repos of 2026, and it is a blueprint that any solo builder can fork in 30 seconds.
What gstack Actually Is
gstack is not a code generation tool. It is a methodology encoded as software. Tan describes it as a set of 23 opinionated AI agent tools that serve as a virtual engineering team. The slash-command specialists cover every role a software team needs: a CEO who rethinks the product, a designer who catches AI slop, a staff engineer who finds production bugs, a QA lead who opens a real browser and clicks through flows, a security officer who runs OWASP and STRIDE audits, and a release engineer who ships the PR. Each specialist runs as a Claude Code skill, chaining into the next in a structured sprint cycle that mirrors how a real engineering team operates.
The sprint cycle is the key architectural insight: Think, Plan, Build, Review, Test, Ship, Reflect. Each skill feeds into the next. Run /office-hours and an AI-powered YC partner pushes back on your product framing. The output design doc feeds into /plan-ceo-review, which challenges scope and extracts the real problem. That plan goes to /plan-eng-review for architecture diagrams and edge case analysis. Then the build phase produces code that /review audits for bugs it auto-fixes. The /qa skill opens a real browser and tests the staging environment. Finally /ship runs tests, bootsraps missing coverage, opens a PR, and /land-and-deploy merges and verifies production health. Nothing falls through the cracks because every skill inherits context from the one before it.
Tan reports that his 2026 coding velocity measured at roughly 810 times his 2013 pace, when he built Bookface at YC. On logical code change, not raw lines, he shipped 11,417 logical lines per day in 2026 compared to 14 per day in 2013. Year to date through April, 2026 had already produced 240 times the entire 2013 year across 40 public and private repos. The methodology behind the measurement is fully documented and reproducible in the repo for anyone who wants to verify the numbers.
Why It Resonated at 122K Stars
Developer tools go viral when they solve a pain that everyone feels but nobody has articulated clearly. gstack's pain is AI agent bloat. As coding agents like Claude Code and Codex grow more capable, they also grow more verbose. They generate unnecessary boilerplate, add libraries nobody asked for, and over-engineer solutions that a senior engineer would dispatch in a single function. gstack reverses this by constraining each agent to a specific role with specific outputs. The CEO specialist does not write code. The engineer specialist does not design. Each agent stays in its lane, producing higher quality output than a single generalist agent could.
The 122K star count in under four months is a signal, not a vanity metric. It tells us that the developer community has been waiting for a structured alternative to the blank-prompt approach. Every Claude Code user has experienced the frustration of asking for a feature and getting back a sprawling, over-engineered solution. gstack provides guardrails for the agent, not constraints on the human. The slash commands are prompts, not policies. You can skip any step, run them out of order, or override any decision. The structure is optional, which is exactly what makes it so widely adopted.
Tan's README leads with a quote from Andrej Karpathy that has become a rallying cry for the AI-native coding movement: "I don't think I've typed like a line of code probably since December." Karpathy's sentiment captures the zeitgeist shift. The question is no longer whether AI agents can write code. It is how to direct them so the code they write is actually good. gstack is the most comprehensive answer to that question that has been published under an open-source license this year.
What Makes It Different From Other Agent Workflows
The most obvious comparison is to OpenClaw, Peter Steinberger's 247K-star open-source agent orchestration framework. OpenClaw spawns multiple Claude Code sessions and coordinates them. gstack takes a different approach: it runs inside a single Claude Code session but layers 23 specializations as slash commands. The two are complementary rather than competitive. In fact, gstack includes an OpenClaw integration guide. When OpenClaw spawns a Claude Code session, it can tell that session to load gstack skills, combining multi-agent orchestration with role-based specialization.
The second comparison is to Ponytail, the 85K-star agent skill that reduces AI code generation by 54% through YAGNI enforcement. gstack and Ponytail solve related but distinct problems. Ponytail is a single skill that prevents over-engineering at the prompt level. gstack is an entire workflow that covers the full product lifecycle from ideation to production deployment. A power user would likely use both: Ponytail to keep individual agent outputs lean, and gstack to orchestrate the overall engineering process.
What makes gstack genuinely different is its focus on product thinking, not just code generation. The /office-hours skill does not write code. It asks six forcing questions that reframe your product before you write a single line. The /plan-ceo-review skill has four modes: Expansion, Selective Expansion, Hold Scope, and Reduction. These are product strategy tools, not coding tools. Tan has essentially open-sourced the decision-making framework he uses as a YC partner to evaluate thousands of startups. That is why gstack feels less like a developer tool and more like a how-to-ship manual from someone who has done it at every scale.
How to Get Started in 30 Seconds
Getting started with gstack takes one command pasted into Claude Code. The setup clones the repo, installs dependencies, and registers all 23 slash commands. The install also works on 10 different AI coding agents including OpenAI Codex CLI, Cursor, Hermes, and Kiro, not just Claude Code. The team mode feature allows auto-updating for shared repos so every teammate gets gstack automatically without vendored files or version drift.
The recommended first steps after install are simple. Run /office-hours and describe what you are building. The AI pushes back on your framing, challenges your premises, and generates implementation alternatives. Run /plan-ceo-review on any feature idea. Then run /review on a branch with changes. Run /qa on a staging URL. In five commands, you get the full gstack experience without committing to the entire workflow. Tan explicitly designed the onboarding path so users can try the most valuable skills independently and adopt the full cycle when they are ready.
Who This Is For
gstack is designed for technical founders and CEOs who still want to ship code but do not have the time to manage an engineering team. It is also valuable for first-time Claude Code users who find the blank prompt overwhelming and want structured roles to guide them. For tech leads and staff engineers, gstack provides rigorous code review, QA, and release automation on every PR. The MIT license means there are no barriers to adoption. Fork it, modify it, contribute back. Garry Tan has published the exact workflow that let him ship at 810 times his 2013 pace, and any solo founder can download it and start using it in the time it takes to brew coffee.