Nine opinionated workflow skills that transform a single AI assistant into a team of specialists — from founder-level product thinking to automated QA testing and one-command shipping.
View on GitHubMany developers treat AI as a universal tool, giving it vague, multi-task instructions within a single context. The result is unfocused output that lacks the depth a real engineering team would deliver. GStack was built on a different philosophy: AI should not stay in one generic mode — it needs explicit cognitive gears.
GStack's core idea is to transform a single AI assistant into a virtual software development team. Instead of letting AI handle everything in one mushy mode, GStack provides clearly defined roles — each activated by a slash command. These roles are called Skills, and they cover the full software development lifecycle, from product vision to post-launch retrospective.
As Garry Tan puts it:
"GStack just hooks everyone up with that 'I know kung fu' moment."
Planning is not review. Review is not shipping. Founder taste is not engineering rigor. If you blur all of that together, you get a mediocre blend of all four. GStack gives you explicit gears — the ability to tell the model what kind of brain you want right now.
Each GStack skill represents a distinct role in a software team. Invoke them as slash commands inside Claude Code to switch cognitive modes on demand.
For a comprehensive overview of all nine skills with detailed examples, see the complete GStack skills guide.
Rethink the problem from the user's point of view. Instead of implementing the obvious feature, this skill finds the 10-star product hiding inside the request — asking what the product is actually for, not just how to add the feature.
Read the full guide →Lock in architecture, data flow, state transitions, edge cases, failure modes, and test coverage. This skill produces diagrams — sequence, state, component, and data-flow — because diagrams force hidden assumptions into the open.
Read the full guide →Find the bugs that pass CI but blow up in production. A structural audit looking for N+1 queries, race conditions, trust boundary violations, missing indexes, broken retry logic, and stale reads. Also triages Greptile code review comments automatically.
Read the full guide →Sync main, run tests, resolve Greptile review issues, push the branch, and open a pull request. This is for a ready branch, not for deciding what to build. It handles the repetitive release hygiene that kills momentum.
Read the full guide →Give the agent eyes. It logs in, clicks through your app, takes screenshots, reads them, checks console errors, and catches breakage. A full QA pass in 60 seconds. Built on a persistent Chromium daemon with sub-second command latency powered by Playwright.
Read the full guide →Systematic QA testing with a find-fix-verify cycle. On a feature branch, it auto-analyzes your git diff, identifies affected pages, tests them, fixes bugs with atomic commits, and re-verifies. Four modes: diff-aware (automatic), full exploration, quick smoke test, and regression baseline comparison. Outputs a health score from 0 to 100.
Read the full guide →Report-only QA testing. Same methodology as /qa but never fixes anything. Use this when you want a pure bug report — a clean handoff to your team without the agent touching your code.
Read the full guide →Import cookies from your real browser (Comet, Chrome, Arc, Brave, Edge) into the headless session. Test authenticated pages without logging in manually. An interactive picker lets you choose which domains to import — no cookie values are ever displayed.
Read the full guide →Team-aware engineering retrospective. Analyzes commit history, coding sessions, shipping velocity, test ratios, and PR sizes. Provides a deep-dive on your work plus per-contributor praise and growth opportunities. Saves JSON snapshots for trend tracking across weeks.
Read the full guide →GStack skills are designed to be used in sequence, following the natural rhythm of how software gets built — from product vision to deployed code. Here is how a typical feature flows through the GStack workflow.
/plan-ceo-review — Pressure-test the ideaDescribe what you want to build. GStack asks whether you are even solving the right problem. For example, "let sellers upload a photo" becomes "help sellers create listings that actually sell" — auto-identifying products, pulling specs, drafting descriptions, and suggesting hero images. Learn more.
/plan-eng-review — Design the architectureOnce the product direction is locked, GStack becomes your technical lead. It produces architecture diagrams, defines async job boundaries, identifies failure modes, and builds a test matrix — turning the vision into something buildable.
Exit plan mode and write the code. GStack stays out of the way during implementation — these skills are about choosing the right cognitive mode, not micromanaging every line.
/review — Find production-grade bugsThe paranoid review pass: race conditions, N+1 queries, trust boundary violations, orphaned uploads, broken invariants. This is where GStack imagines the production incident before it happens. Learn more.
/ship — Land the branchSync main, run tests, push, open PR — six tool calls, done. GStack handles the boring release hygiene that kills momentum after the interesting work is finished. Learn more.
/qa — Verify everything worksGStack analyzes your branch diff, identifies affected routes, spins up a browser, and tests each page. No manual test plan needed. Before/after health scores and a ship-readiness summary. Learn more.
GStack is more than Markdown prompts. The /browse and /qa skills are powered by a compiled binary that talks to a persistent Chromium daemon — engineered for the sub-second latency and persistent state that AI-driven browser interaction demands.
First call starts the browser in approximately 3 seconds. Every subsequent command completes in 100–200ms via HTTP. Cookies, tabs, and localStorage carry over between commands. Auto-shuts down after 30 minutes idle. Learn more about browser automation.
A single 58MB executable compiled with Bun. No node_modules at runtime, no PATH configuration. Native SQLite for cookie decryption, native TypeScript execution, and a built-in HTTP server.
Elements are addressed via accessibility-tree refs (@e1, @e2) — not CSS selectors. Uses Playwright Locators externally, avoiding DOM mutation. Works with CSP-restricted sites, React hydration, and Shadow DOM.
Localhost-only binding. Bearer token auth per session. Cookie values never written to disk in plaintext. Keychain access requires explicit user approval. Database opened read-only. Shell injection prevented via hardcoded registry.
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 with a separate Chromium instance, cookies, tabs, and logs. No port collisions, no shared state, no configuration needed.
One person, ten parallel agents, each with the right cognitive mode for its task. One session running /qa on staging, another doing /review on a PR, a third implementing a feature, and seven more working on other branches — all at the same time. That is not incremental improvement. That is a different way of building software.
Requirements: Claude Code, Git, and Bun v1.0+. The /browse skill compiles a native binary that works on macOS and Linux (x64 and arm64).
Open Claude Code and paste this command. Claude will handle the rest.
Share GStack with your team. Real files get committed to your repo — teammates just run setup once.
~/.claude/skills/gstack/, symlinks for each skill, browser binary at browse/dist/browse (~58MB), and node_modules. Everything lives inside .claude/. Nothing touches your PATH or runs in the background.
For the complete walkthrough including team setup, troubleshooting, and upgrading, see the full installation guide.
Greptile reviews your PRs automatically, catching race conditions, security issues, and bugs that pass CI. GStack adds a triage layer on top: /review and /ship read Greptile's comments, classify each one, and take action.
Valid issues get added to critical findings and fixed before shipping. Already-fixed issues get an auto-reply acknowledging the catch. False positives get pushed back with an explanation. The result is a two-layer review where nothing falls through the cracks.
GStack also learns from false positives — every confirmed FP is saved to a history file, and future runs auto-skip known patterns for your codebase. The /retro skill tracks Greptile's signal-to-noise ratio over time.
GStack is not a theoretical framework. It was born from hands-on daily use by its creator, Garry Tan, who publicly shared his productivity metrics. Using this exact GStack setup, Tan averaged 10,000 lines of code and 100 pull requests per week over a 50-day period — demonstrating what is possible when AI is given structured roles rather than used as a generic assistant.
Accumulated within weeks of its open-source release. GStack surpassed 10,000 stars in the first 48 hours alone, making it one of the fastest-growing developer tools on GitHub in 2026.
Active community adoption with developers forking and customizing GStack for their own teams and workflows. The project maintains 14 open issues and 45 pull requests from community contributors.
Garry Tan's personal throughput using GStack with Claude Code — averaging 10K lines of code and 100 PRs per week across a 50-day stretch, all maintained at production quality.
Built primarily in TypeScript (79.6%) with Go (18.3%) for performance-critical components. The codebase emphasizes type safety, maintainability, and developer experience.
Most AI coding tools — including GitHub Copilot — focus on code generation and autocomplete within the editor. GStack takes a fundamentally different approach: instead of making code suggestions, it provides structured, role-driven workflows that cover the entire software development lifecycle. Here is how GStack differs from conventional AI pair-programming tools.
Traditional AI assistants operate in a single mode — they respond to whatever you ask, with no awareness of your current stage in the development process. GStack explicitly defines nine cognitive modes (CEO, engineering manager, QA engineer, etc.), each with a distinct set of priorities and evaluation criteria. This means the AI's output quality is calibrated to the specific task at hand, not diluted by trying to be everything at once.
GitHub Copilot and similar tools operate within the code editor — they cannot see or interact with a running application. GStack's /browse skill provides a persistent, high-performance Chromium session that completes commands in 100–200ms after initial startup. This is reported to be 20 times faster than Claude for Chrome MCP, with no context bloat. The agent can log in, navigate pages, take screenshots, verify layouts, and catch console errors — closing the loop between writing code and verifying it works.
Where traditional tools assist primarily with code generation and completion, GStack covers product planning (/plan-ceo-review), architecture design (/plan-eng-review), structural code review (/review), automated shipping (/ship), visual QA (/qa), and team retrospectives (/retro). It models the full workflow of a functioning software engineering organization, not just the coding phase.
GStack is MIT-licensed and installed locally as Markdown skill files and a compiled binary. Every skill is readable and editable. There is no SaaS dependency, no telemetry, and no vendor lock-in. Developers can fork the project, modify any workflow, or extend it with entirely new skills. This is in contrast to proprietary services where behavior customization is limited to configuration toggles.
Each GStack skill has a dedicated guide with detailed examples, architecture explanations, and best practices.
Complete guide to all nine skills with workflow examples and cognitive mode explanations.
CEO and engineering review modes for product thinking and architecture design.
Structural audit for N+1 queries, race conditions, and trust boundary violations.
One-command release: sync, test, push, and PR creation.
Persistent Chromium daemon with sub-second commands for visual QA testing.
Diff-aware testing, health scores, find-fix-verify cycles.
Run ten Claude Code sessions simultaneously with isolated workspaces.
How GStack compares to GitHub Copilot and other AI coding assistants.
Complete installation guide for personal and team setup.
GStack is an open-source skill pack for Claude Code that transforms a single AI assistant into a virtual software development team. It provides nine specialist roles — from founder-level product review to automated QA testing and one-command shipping — all invoked as slash commands inside Claude Code.
GStack is for developers who already use Claude Code heavily and want consistent, high-rigor workflows instead of one generic mode. It is not a prompt pack for beginners — it is an operating system for people who ship. You want to tell the model what kind of brain to use right now: founder taste, engineering rigor, paranoid review, or fast execution.
GStack works on macOS and Linux (both x64 and arm64 architectures). It requires Claude Code, Git, and Bun v1.0 or above. The /browse skill compiles a native binary for headless browser automation.
Yes. GStack is open source under the MIT License. The source code is freely available on GitHub. You need an active Claude Code subscription to use the skills, as they run within Claude Code sessions.
Run /gstack-upgrade inside Claude Code. It detects your install type, upgrades both global and project-vendored copies, and shows what is new. You can also set auto_upgrade: true in your config to upgrade automatically whenever a new version is available.
Yes. Use /setup-browser-cookies to import cookies from your real browser (Chrome, Arc, Brave, Edge, or Comet) into the headless session. Then /browse and /qa can test pages behind login without you logging in manually. Cookie decryption happens in-memory via the macOS Keychain — values are never written to disk in plaintext.
GitHub Copilot excels at code generation and in-editor autocomplete. GStack operates at a different level: it provides structured, role-driven workflows covering the entire development lifecycle — not just writing code, but planning products, reviewing architecture, running visual QA in a real browser, and automating the release process. GStack also includes headless browser automation that is 20x faster than Chrome MCP tools, enabling visual QA testing that code-only assistants simply cannot perform. The two tools are complementary rather than competing.
GStack's /browse skill uses a persistent Chromium daemon built on Playwright. The first command starts the browser in about 3 seconds. Every subsequent command — navigation, clicking, screenshots, form filling — completes in 100–200 milliseconds. The browser maintains cookies, tabs, and localStorage between commands with an automatic 30-minute idle shutdown. Each workspace gets its own isolated browser instance, so parallel sessions never interfere.
Yes. GStack supports up to ten parallel sessions via Conductor, each in its own isolated workspace with a separate Chromium process. Random port selection (10000–60000) ensures zero port collisions. One person can have sessions running QA, code review, feature implementation, and shipping simultaneously — each in the right cognitive mode for its task.
Results vary by workflow, but GStack's creator Garry Tan publicly reported averaging 10,000 lines of code and 100 pull requests per week over a 50-day period using this exact setup. The key is not the volume — it is the structured quality that comes from using the right cognitive mode for each stage of development. Planning stays in planning mode, review stays paranoid, and shipping stays mechanical.