Run Garry Tan's GStack workflow in your Operator agent

When you hand an agent a vague "build me X," it tends to take the request at face value and write whatever the first reading suggests. That is fine for a one line script and a problem for anything real, where the gap between what you asked for and what you meant is exactly where the bugs and the wasted afternoons live. GStack is a set of skills that forces the agent to slow down at the right moments instead.

It comes from Garry Tan, the president and CEO of Y Combinator, who put his own AI coding setup on GitHub under an MIT license. The framing he opens with is a line from Andrej Karpathy on the No Priors podcast: "I don't think I've typed like a line of code probably since December." GStack is one answer to how a single person ships at that pace.

It turns a coding agent into a small product team made of slash skills: one that interrogates the product idea, one that plans the architecture, one that reviews the diff for the bugs that pass CI and break in production, one that opens a real browser to QA the result, and one that opens the pull request. GStack supports OpenClaw directly, and the OpenClaw integration doc explains how the methodology skills and the full coding pipeline fit together. Since Operator.io runs on OpenClaw, your agent can install it and run the workflow itself.

Garry Tan walks through GStack himself on Y Combinator's channel, demoing office hours, code review, and browser QA on a live task, the same skills your agent loads when you send the prompt below.

How it works

You send the prompt and the agent installs GStack, then runs building work through its stages instead of in one shot.

• It installs the four GStack skills that run natively in your chat from ClawHub, so the methodology is available without spawning a separate session.
• It clones the full GStack pack and runs the setup, then adds a short section to its own AGENTS.md telling it to load GStack for coding tasks.
• When you hand it something to build, it starts with office hours to pressure test what you are actually asking for, plans the work, writes it, then reviews and tests before anything ships.
• It tells you what it found at each step, and it only opens the pull request once the review and QA passes are clean.

The reason the order matters is that each skill feeds the next. Office hours writes down the real requirements so the plan is grounded in them, the plan defines the tests, the review checks the diff against both, and the ship step verifies the tests still pass. Nothing falls through because every step knows what came before it.

On OpenClaw, the full coding pipeline often runs inside a spawned Claude Code session through the ACP runtime, while the four native skills run conversationally in your Telegram or Discord chat. The OpenClaw doc describes that split: you stay in chat for office hours and planning, and the agent dispatches the implementation, review, and ship loop to a coding session with GStack loaded at ~/.claude/skills/gstack.

The setup step is worth a look before you let the agent run it. The git clone in the prompt pulls the default branch, and ./setup is a shell script that writes skills into ~/.claude/skills, edits your agent's AGENTS.md, and can install tooling, so read it the way you would read any setup script you were about to run yourself. If you want a known version, clone a tagged release instead of the default branch so the script the agent runs is the one you reviewed. It is ordinary setup for a coding pack, and it runs with your agent's full permissions, which is the reason to glance at it first.

The skills

Four of GStack's skills run conversationally inside an OpenClaw agent, which is what makes them a fit for Operator out of the gate. The rest of the pack is aimed at writing and shipping actual code.

Ship bootstraps a test framework first if the project does not have one. The full set, with the philosophy behind each, is in the GStack skill reference. The MarkTechPost overview walks through how /autoplan, /review, /qa, and /ship map to planning, review, browser testing, and release.

GStack also includes security and deployment skills such as /cso for OWASP and STRIDE audits and /browse for headless browser access, all documented in the gstack README. You can invoke them by name in plain language once the pack is installed.

It is worth being clear about what these skills touch, because the coding side acts on real things in your repo and browser. QA opens a real browser and drives it through your flows, so any site it loads runs against whatever sessions and cookies that browser already holds. Ship works against your actual git remote and opens pull requests with the credentials your agent has, and /browse gives it headless browser access on the same footing.

That reach is also why the prompt has you decide up front whether the agent may ship without showing you the plan, and why it pays to point GStack at the repositories and accounts you are comfortable handing it rather than your most sensitive ones on the first run.

The prompt

This is the instruction the agent acts on:

Set yourself up with GStack, Garry Tan's open source pack of engineering
workflow skills, so when I ask you to build or fix something you work like a
small product team rather than in a single pass. Four of the skills run
directly here in chat: install them from ClawHub with clawhub install
gstack-openclaw-office-hours gstack-openclaw-ceo-review
gstack-openclaw-investigate gstack-openclaw-retro, so you can pressure test a
half formed idea with office hours, challenge the scope with the CEO review,
debug from root cause with investigate, and run a retro on what we shipped.
For the full coding workflow, install the rest of GStack by running git clone
--single-branch --depth 1 https://github.com/garrytan/gstack.git
~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup, then add a
short section to your AGENTS.md that says to load GStack for coding tasks and
use autoplan to plan, review and qa to catch bugs, and ship to open the pull
request. From then on, when I hand you something to build, reframe it with
office hours first, plan it, write it, review and test it, then ship, and tell
me what you found at each step. Before you start, ask me what I mostly build,
the language and stack I work in, and whether you should ever ship without
showing me the plan first.

The same prompt is saved in the prompts library, so you can send it to your agent without retyping a word.

Using it day to day

Say you ask for a daily briefing tool. With office hours installed, the agent does not start coding. It asks where the current pain actually is, points out that what you described is closer to a personal chief of staff than a briefing, and gives you a couple of scoped ways to start before you commit to the three month version.

You pick one, it plans the piece, writes it, runs review and QA, and shows you the result with the bugs it already caught and fixed. The skills also respond to plain language, so "run a security check" or "test the website" reaches the right one without you remembering a slash command.

You stay the decision maker the way a founder stays close to a team: you approve the plan and the taste calls, and you let the mechanical steps run. The prompt asks up front whether the agent may ever ship without showing you the plan, so you set that boundary once. For the methodology skills that run in chat, you can also just talk to the agent normally and ask it to take you through office hours or a retro when you want the structure.

If your agent has GitHub connected, the ship step can open a pull request against your repo once review and QA pass. The GitHub connection guide covers the token or MCP setup that step needs.

Where it fits in Operator

The four ClawHub skills run inside your Operator agent's own chat, so the planning and review methodology is useful even when no code is involved, for example pressure testing a launch plan or running a weekly retro on your projects. The full coding pack matters most when your agent is actually building software in its workspace, where the plan, review, QA, and ship loop has real diffs to work on. GStack is free and MIT licensed, so there is no tier to clear before you try it.

To set it up, open the prompts library and send the GStack prompt to your agent. It asks what you build, the stack you work in, and whether it can ship without showing you the plan, and then it runs the next thing you build through the full loop.