gstack: Garry Tan's Claude Code role-play kit for solo shipping

What gstack actually is

gstack is a collection of 23 slash commands for Claude Code, published by Garry Tan, the president and CEO of Y Combinator. The commands are meant to replace a blank prompt with structured roles: a CEO that rethinks the product, an engineering manager that locks architecture, a designer that catches what Tan calls AI slop, a reviewer that finds production bugs, a QA lead that opens a real browser, a security officer that runs OWASP and STRIDE audits, and a release engineer that ships the pull request. It also includes eight power tools, all in Markdown, all MIT licensed. The target user is a technical founder or a staff engineer who wants a repeatable process for shipping without a human team. The README is explicit about who it is for: founders, first-time Claude Code users, and tech leads who want rigorous review on every PR.

The mechanism: slash commands as a virtual team

The core idea is that each slash command encodes a workflow. You invoke /office-hours to describe what you are building, then /plan-ceo-review on a feature idea, then /review on a branch, then /qa on a staging URL. The commands are not standalone programs; they are Markdown instructions that Claude Code interprets. The README gives a quick start sequence: install, run /office-hours, run /plan-ceo-review, run /review, run /qa, and stop there to decide if the tool fits. The design assumes that Claude Code can follow long, structured prompts reliably. The commands are opinionated, meaning they enforce Tan's specific standards for what a CEO review or a QA pass should include. That is both the strength and the risk: you get a process that works for one person, but you inherit their biases.

Installation and team mode

Installation is a single command that you paste into Claude Code: git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup. The setup script then asks you to add a gstack section to CLAUDE.md that lists the available skills and instructs Claude Code to use the /browse skill for all web browsing, never the mcp__claude-in-chrome__* tools. There is a team mode for shared repos: you run ./setup --team from inside your repo, then bin/gstack-team-init required, and commit the .claude/ and CLAUDE.md changes. The README claims this gives auto-update checks throttled to once per hour, network-failure-safe, and silent. You can swap required for optional to nudge teammates instead of blocking them. The team mode avoids vendored files and version drift, but it also means every teammate's Claude Code session will pull the latest gstack automatically, which could break a pinned workflow.

OpenClaw integration and native skills

gstack also works with OpenClaw, an AI agent framework that spawns Claude Code sessions via ACP. The README provides a prompt template for OpenClaw agents: when spawning a Claude Code session for coding work, tell it to use gstack skills. Example dispatches include running /cso for a security audit, /review for code review, /qa for URL testing, and /autoplan followed by /ship for a feature build. There are also four methodology skills that work directly in OpenClaw without a Claude Code session, installable via clawhub install gstack-openclaw-office-hours. This is a meaningful extension because it means gstack is not locked to the Claude Code CLI; it can be driven by a higher-level agent. The README mentions gstack-lite and gstack-full prompt templates in docs/OPENCLAW.md, but the details are truncated in the available material.

The productivity claims and the LOC controversy

The README makes aggressive productivity claims: Tan says his 2026 run rate is about 810 times his 2013 pace, measured in logical code changes, not raw lines. He acknowledges that raw line counts inflate with AI and points to docs/ON_THE_LOC_CONTROVERSY.md for methodology, caveats, and a reproduction script. The claim is that he shipped 3 production services and 40+ features in 60 days while running YC full-time. These numbers are not independently verifiable from the repository, and the README itself admits that the LOC critics have a point about raw counts. The value of gstack does not depend on whether those numbers are accurate; it depends on whether the slash commands produce useful output. Still, you should treat the productivity multipliers as marketing, not as a benchmark you can reproduce.

Limitations and wrong-tool cases

The most obvious limitation is that gstack is only as good as Claude Code's ability to follow the Markdown instructions. If you use a different AI coding assistant, gstack is useless unless you go through OpenClaw. The commands are also opinionated to Tan's workflow, so a CEO review or a design consultation will reflect his standards, not yours. You can fork and modify, but that is maintenance work. Another limitation is the team mode auto-update: it silently updates every session, which could introduce breaking changes to your team's workflow without notice. The README says it is network-failure-safe, but it does not describe how to pin a specific version. If you need deterministic behavior for a regulated environment, this is the wrong tool. Also, the QA command opens a real browser, which means you need a browser automation setup; the /setup-browser-cookies command exists, but it adds setup complexity.

Alternatives and what to compare

The closest alternative is a plain Claude Code setup with your own CLAUDE.md and custom slash commands. The difference is that gstack gives you a pre-built, opinionated set of roles and workflows, while a custom setup requires you to write those prompts yourself. Another alternative is using Claude Code's built-in skills or community skill collections, but those are typically less structured. OpenClaw itself is an alternative if you want a general-purpose agent that can spawn multiple sessions, but it does not come with the role definitions that gstack provides. The real comparison is between adopting gstack's conventions and building your own prompt library. gstack saves you the initial writing effort but locks you into Tan's process. If you want to understand what each command does in detail, you need to read the Markdown files in the repository, which the README does not fully document.

Maintenance, license, and upgrade path

gstack is MIT licensed, so you can fork, modify, and redistribute it freely. The repository is not archived and has no recent releases listed, which suggests it is maintained on a rolling basis. There is a /gstack-upgrade command, which implies that upgrading is a supported workflow. In team mode, the auto-update mechanism means you do not have to manually upgrade, but you also lose control over when changes arrive. The README says there are no vendored files in your repo, so the code lives in ~/.claude/skills/gstack on each machine. That means uninstalling is a simple matter of deleting that directory and removing the gstack section from CLAUDE.md. The maintenance cost is low if you accept the defaults, but if you customize the commands, you will need to merge upstream changes manually. There is no documentation of a changelog or release notes in the available material, so you cannot assess how frequently breaking changes occur.

Editorial conclusion

Adopt gstack if you are a technical founder or solo developer who already lives inside Claude Code and wants structured prompts for planning, review, and QA. Skip it if you need a tool that works outside the Claude Code ecosystem or if you expect it to replace human judgment in architecture or design. Before adopting, verify that the /review and /qa skills actually run against your stack, confirm that the auto-update behavior in team mode matches your repo's governance, and read docs/ON_THE_LOC_CONTROVERSY.md to understand the productivity claims. gstack is a personal workflow made public, not a product with a support contract, so treat its defaults as starting points to fork and modify.

Official sources

1. Official README
2. Project repository