In This Guide
What is gstack?
gstack is a skill pack for Claude Code that installs nine slash commands, each tuned for a distinct phase of software development. Instead of getting one generic AI assistant that handles everything with the same mushy approach, you get nine specialists: a founder who pressure-tests product direction, an engineering manager who locks down architecture, a paranoid staff engineer who hunts production-killing bugs, a release engineer who handles the tedious shipping mechanics, a QA engineer with a real browser, and more.
Created by Garry Tan, President and CEO of Y Combinator, gstack was born from a simple frustration: AI coding tools work in one mode. They take your request literally, implement exactly what you said, and never ask whether you are building the right thing. The result is a mediocre blend of planning, coding, reviewing, and shipping all mashed together.
gstack fixes that by giving you explicit gears. You tell the model what kind of brain to use right now. Founder taste. Engineering rigor. Paranoid review. Fast execution. Each mode activates a different cognitive frame, and switching between them is as simple as typing a slash command.
Why Cognitive Modes Matter
When you ask a generic AI assistant to "review this PR," it does not know whether you want a nitpick pass on code style, a deep structural audit for race conditions, or a quick gut check before merging. It guesses. Sometimes it gives you three paragraphs about variable naming. Sometimes it catches a real security issue. The inconsistency is the problem.
Cognitive modes solve this. Each gstack skill locks Claude Code into a specific mindset with explicit constraints:
- Planning mode thinks in terms of product vision, user empathy, and long time horizons. It actively resists taking requests literally.
- Engineering mode thinks in architecture diagrams, failure modes, state machines, and test matrices. It resists sprawling ideation.
- Review mode thinks like a paranoid staff engineer who has been burned by production incidents. It looks for bugs that pass CI.
- Shipping mode thinks in checklists and execution. No more talking. Land the plane.
- QA mode thinks with a real browser, clicking through your app and checking every page your code changes touch.
The result is that each mode is dramatically better at its specific task than a generic approach. When you use /plan-ceo-review to evaluate product direction, the model actively pushes back on weak ideas. When you switch to /review, it hunts for N+1 queries and race conditions instead of debating product strategy. Each mode excels because it is not trying to do everything at once.
All 9 gstack Claude Code Skills
Rethink the problem from the user's point of view. This skill activates what Garry Tan calls "Brian Chesky mode" — it refuses to take your request literally and instead asks what the 10-star product hiding inside the request looks like. When you say "add photo upload," it asks whether the real feature is helping sellers create listings that actually sell.
The model challenges assumptions, expands scope when the real opportunity is bigger, and narrows scope when the team is over-building. It thinks with taste, ambition, user empathy, and a long time horizon.
Read the full /plan-ceo-review guide →Once the product direction is locked, this skill becomes your best technical lead. It nails architecture, system boundaries, data flow, state transitions, failure modes, edge cases, trust boundaries, and test coverage. The key unlock: it produces diagrams. Sequence diagrams, state diagrams, component diagrams, and data-flow diagrams that force hidden assumptions into the open.
The goal is not to make the idea smaller. It is to make the idea buildable — with a technical spine that can carry the product vision without collapsing under edge cases.
Read the full /plan-eng-review guide →A structural audit for bugs that pass CI but blow up in production. This skill looks for N+1 queries, stale reads, race conditions, bad trust boundaries, missing indexes, escaping bugs, broken invariants, and tests that pass while missing the real failure mode. It also integrates with Greptile to triage automated review comments, separating valid issues from false positives.
This is not a style nitpick pass. It is a paranoid reviewer imagining the production incident before it happens.
Read the full /review guide →The last-mile execution skill. When the branch is ready, /ship syncs main, runs tests, resolves any Greptile review comments, pushes the branch, and opens the PR. It handles the boring release hygiene that kills momentum — the stuff humans procrastinate when the interesting work is done.
This skill is for a ready branch, not for deciding what to build. It takes care of syncing, testing, versioning, pushing, and PR creation in one command.
Read the full /ship guide →Gives Claude Code eyes. A persistent headless Chromium browser that starts in about 3 seconds and responds in 100–200ms per command after that. The agent can navigate URLs, fill forms, click buttons, take screenshots, read console errors, and verify deployments — all without you opening a browser.
Built on Playwright by Microsoft, it maintains persistent sessions so cookies, localStorage, and login state carry over between commands. After a push, you can tell Claude to QA staging end-to-end.
Read the full /browse guide →Where /browse gives the agent eyes, /qa gives it a testing methodology. It reads your git diff, identifies which pages and routes your changes affect, spins up the browser, and tests each one. No URL required. No manual test plan. It figures out what to test from the code you changed, then fixes any bugs it finds with atomic commits and re-verifies.
Three tiers: Quick smoke test (30 seconds), Standard (diff-aware), and Exhaustive (full app exploration). Delivers before/after health scores and a ship-readiness summary.
Read the full /qa guide →Report-only QA testing. Uses the same methodology as /qa but never fixes anything. It produces a pure bug report with screenshots, health scores, and categorized issues. Use this when you want a clean assessment without any code changes — perfect for auditing someone else's branch or getting a baseline quality score.
Imports cookies from your real browser (Comet, Chrome, Arc, Brave, Edge) into the headless Playwright session. This means /qa and /browse can test authenticated pages without you logging in manually every time. An interactive picker lets you choose exactly which domains to import, and cookie values are never displayed.
First import per browser triggers a macOS Keychain prompt. Decryption happens in memory and cookie values are never written to disk in plaintext.
Read the full cookie import guide →A data-driven engineering retrospective. Analyzes commit history, work patterns, and shipping velocity to write a candid retro. It identifies every contributor, gives you the deepest treatment on your own work, then breaks down each team member with specific praise and growth opportunities — the kind of feedback you would give in a real 1:1.
Computes metrics like commits, lines of code, test ratio, PR sizes, and fix ratio. Detects coding sessions from timestamps, finds hotspot files, tracks shipping streaks. Saves JSON snapshots for trend tracking over time.
Read the full /retro guide →The Complete Workflow: One Feature, Five Modes
The real power of gstack Claude Code skills is not any single slash command. It is how they compose into a complete development workflow. Here is how Garry Tan builds a feature using all five modes in sequence:
Start in plan mode. Describe the feature. Use /plan-ceo-review to pressure-test whether you are even building the right thing. Only after the product direction is locked, use /plan-eng-review to nail the architecture with diagrams, failure modes, and test coverage.
Implement the plan. Then use /review to find bugs that pass CI — race conditions, N+1 queries, trust boundary violations. Fix them. Use /ship to sync with main, run tests, push, and open the PR. Finally, /qa reads your diff and tests every affected page in the browser.
You: [describe the feature in plan mode]
You: /plan-ceo-review # Is this the right product?
You: /plan-eng-review # Lock the architecture
You: [exit plan mode, implement]
You: /review # Find production-killing bugs
You: /ship # Sync, test, push, open PR
You: /qa # Browser-test every changed pageSix interactions. Each one activates a specialist that is excellent at exactly one thing. That is a fundamentally different way to use AI for software development.
Parallel Sessions with Conductor
gstack is powerful with one Claude Code session. It is transformative with ten. Conductor runs multiple Claude Code sessions in parallel, each in its own isolated workspace. One session runs /qa on staging. Another does /review on a PR. A third implements a feature. Seven more work on other branches. All at the same time.
Each workspace gets its own isolated browser instance automatically — separate Chromium process, cookies, tabs, and logs stored in .gstack/ inside each project root. No port collisions, no shared state, no configuration needed. /browse and /qa sessions never interfere with each other, even across ten parallel workspaces.
Who gstack Is For
gstack is for developers who already use Claude Code heavily and want consistent, high-rigor workflows instead of one mushy generic mode. You want to tell the model what kind of brain to use right now — founder taste, engineering rigor, paranoid review, or fast execution.
This is not a prompt pack for beginners. It is an operating system for people who ship. If you are landing 10 PRs a day and want each one to be reviewed, tested, and shipped with real rigor, gstack is what you are looking for.
Installation Overview
gstack requires Claude Code, Git, and Bun v1.0+. The /browse skill compiles a native binary that works on macOS and Linux (x64 and arm64).
Installation is a single command pasted into Claude Code. Claude clones the repository, runs the setup script, builds the browser binary, registers all nine skills as symlinks, and updates your CLAUDE.md. The whole process takes about 30 seconds.
# Paste this into Claude Code and it handles everything:
Install gstack: run git clone https://github.com/garrytan/gstack.git \
~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setupEverything lives inside .claude/. Nothing touches your PATH or runs in the background. For the complete walkthrough with project-level installation and team setup, see the full installation guide.
Upgrading is just as simple: type /gstack-upgrade in Claude Code. It detects your install type, upgrades, syncs any project copies, and shows what changed.
Open Source and MIT Licensed
gstack is fully open source under the MIT license. The source code is available on GitHub. You can read every skill prompt, inspect the browser binary source, fork it, modify it, and contribute back.
The skill files are Markdown prompts — human-readable instructions that tell Claude Code how to behave in each mode. The browser automation is a compiled Bun binary built on Playwright. Everything is transparent and auditable. When you install gstack into your project repository, real files get committed (not a submodule), so git clone just works for your teammates.
Created by Garry Tan because he does not want AI coding tools stuck in one mushy mode. Contributions are welcome — see CONTRIBUTING.md for setup and development workflow.