Image: GitHub
Why it mattersA reader can clone one repository and have a set of named roles instead of a blank chat. The setup is the message about how Tan thinks a small team should use AI.
A blank chat is the hardest way to use a coding agent. Garry Tan, the Y Combinator president, released gstack as an open Claude Code setup. It bundles 23 agents, each with a named role on a small software team. The repository had 139,922 stars on GitHub on 5 November 2026, per the GitHub API.
The project is ranked 18 on Trendshift's daily list on 5 November 2026. The latest tagged release is v1.96.0.0, published on the same day.
What the 23 roles are
Each role is a prompt the user can call by name inside Claude Code. The README names several. A CEO prompt handles product thinking. A Designer prompt reviews the UI. An Engineering Manager prompt covers architecture and data flow. A Release Engineer prompt handles deployment and pull requests. A QA Lead prompt runs browser tests. A Security Officer prompt runs OWASP and STRIDE checks. A Technical Writer prompt handles documentation. Other prompts cover a design system audit, performance work, iOS testing and a knowledge base.
The setup works as a chain. The CEO prompt rethinks the product. The Engineering Manager prompt turns that into architecture. The release and QA prompts review and test before a merge. Each step reads what the previous step produced, so a review never starts from a blank slate, the README says.
What the repository is and is not
The code is MIT-licensed and free. The setup needs Claude Code, git and Bun v1.4.2 or newer. Team installs follow a stable release channel with two releases a week, on Monday and Thursday, the release notes say.
The repository is a configuration bundle, not a model and not a service. It does not call outside APIs beyond Claude Code's own. A user still pays Anthropic for the model. A user still runs Claude Code as the agent.
The number worth watching is the stars figure. 139,922 stars in about eight months is large, and it carries two signals. First, a lot of developers tried a packaged setup over a blank prompt. Second, Tan's name sells. A fair test is whether teams that did not already know who Tan is get a day-one lift from gstack.
A team that already runs Claude Code can clone the repository and keep its current settings beside gstack's. The simple test is to run the same task twice. First with a blank prompt. Then with the gstack chain: CEO, Engineering Manager, and the review prompts. The question is which run catches the mistakes the team would have fixed in review anyway.
For a founder who wants to try it, the first week is the test. A small change through the full chain takes longer than one prompt. The speed-up has to come from fewer review cycles or fewer bugs in production, not from the first prompt itself.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.