Added Oct 14, 2026
gstack
Skills that have your AI coding agent question your idea like a startup mentor, then review the plan, review the code and write the docs.
- Ecosystem
- Agent Skills
- Skill category
- Coding
- Creator / organization
- Garry Tan
What it is
gstack is an open-source set of Skills by Garry Tan (GitHub user garrytan), who introduces himself in the README as President and CEO of Y Combinator. The README says gstack turns Claude Code into a virtual engineering team, with each role as a slash command written in Markdown, and it can also be installed for Codex and other AI coding tools. The repo lays out more than twenty roles in sprint order (think, plan, build, review, test, ship, reflect); AI Inspiration World recommends only the planning, documentation and code review ones. /office-hours is the starting point: in the style of YC office hours, six forcing questions help you rethink the product, push back on your framing, offer alternative approaches and end with a design doc. /plan-ceo-review rethinks the request from a founder's angle, with four modes: expansion, selective expansion, hold scope and reduction. /plan-eng-review locks in architecture, data flow, edge cases and a test plan. /plan-design-review scores each design dimension from 0 to 10, explains what a 10 looks like and then edits the plan. /autoplan runs these reviews in one go and brings you only the taste decisions. Once code is written, /review looks for bugs that pass CI but break in production and fixes the obvious ones itself; /document-generate fills in missing reference, how-to, tutorial and explanation docs using the Diataxis framework, and /document-release checks which docs a change has made stale. What sets it apart is that the steps feed each other: the design doc from /office-hours is read by /plan-ceo-review, and the engineering review's test plan is passed on to the testing step, so each step knows what came before.
Who it's for:founders and developers with a real project who want the agent to think the idea and plan through before building. You'll need to be comfortable with an AI coding tool that runs commands, such as Claude Code or Codex; readers who don't write code won't get much from it.
Before you start: it installs as a whole and needs Claude Code, Git and Bun (the README names 1.4.2 as the tested version). The README has you clone the repo into ~/.claude/skills/gstack and run ./setup, which downloads dependencies, compiles binaries, downloads Chromium through Playwright and builds gstack's own browser, keeping its state in ~/.gstack. The README's first step also asks the agent to add a CLAUDE.md section saying to use gstack's /browse for all web browsing and never other browser tools; we suggest leaving that part out. Team mode (--team) checks for updates online at the start of every session and commits its setup into your repo; we haven't reviewed or recommended it. With default settings, usage telemetry is sent only if you opt in (off by default), and every Skill start checks GitHub for a new version, which gstack-config set update_check false turns off. Outside reviews are on by default: /office-hours, the plan reviews, /autoplan, /review, /document-release and others send the review prompt and the relevant code, diff or plan to a second model through your own Codex or Claude Code login, with a notice before the first send. We don't recommend sending your code to an outside model for review; run gstack-config set codex_reviews disabled first. When a planning Skill reaches a step that needs a third-party website, it asks whether to drive a browser for you; we suggest choosing the manual steps. gstack's bundled browser engine includes anti-bot stealth code that disguises the browser to get past sites' bot detection, so we do not recommend /browse, /open-gstack-browser, or /qa, /qa-only, /design-review, /scrape and /skillify, which rely on it when the Aside browser isn't installed. We also do not recommend /setup-browser-cookies, which imports your signed-in cookies from Chrome and other browsers into gstack's browser; /pair-agent, which hands browser control to other AI agents and opens an ngrok tunnel for remote ones; /ship, which pushes code and opens a PR; or /land-and-deploy and /setup-deploy, which merge and deploy to production. We haven't reviewed or recommended the other entry points, including /setup-gbrain and /sync-gbrain (which can sync a code index to an outside database), /design-shotgun, the /cso security audit and the iOS device testing tools. MIT license.
Based on the project's public documentation. Not tested by AI Inspiration World.