The Complete gstack Guide: How to Turn Claude Code into a Virtual Engineering Team

March 25, 2026
Views 8
Comments0
This post is also available in: 日本語한국어

Table of contents

What if you could write over 10,000 lines of production code per day, all by yourself? gstack, released by Y Combinator President Garry Tan, is an open-source tool that transforms Claude Code into a virtual engineering team composed of a CEO, designer, staff engineer, QA lead, security officer, and release engineer. This is backed by a track record of shipping over 600,000 lines of production code in just 60 days.

This guide systematically walks you through gstack's installation, all 28 skills, core workflows, and practical tips.

What is gstack?

gstack is a software factory that runs on top of Claude Code. Rather than a simple coding assistant, it provides 28 specialized skills, each fulfilling a distinct professional role, to systematically manage the entire software development process.

Key Metric Details
Code Shipped Over 600,000 lines in 60 days (35% test code)
Daily Productivity 10,000–20,000 lines/day (part-time work)
Number of Skills 28 specialized skills
License MIT (completely free, no premium tier)

Installation

Prerequisites

  • Claude Code – Anthropic's CLI coding tool
  • Git – Version control
  • Bun v1.0+ – JavaScript runtime (for compiled binary generation)
  • Node.js – Required only on Windows

Step 1: Global Install (30 seconds)

git clone https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
cd ~/.claude/skills/gstack && ./setup

This installs gstack to ~/.claude/skills/gstack, and Claude Code automatically recognizes all skills.

Step 2: Add to Project (Optional)

To share the same configuration across your entire team, copy it into your project repository:

cp -Rf ~/.claude/skills/gstack .claude/skills/gstack
rm -rf .claude/skills/gstack/.git
cd .claude/skills/gstack && ./setup

Codex/Gemini/Cursor Users

# Repo-local install
git clone https://github.com/garrytan/gstack.git .agents/skills/gstack
cd .agents/skills/gstack && ./setup --host codex

# Or global install
git clone https://github.com/garrytan/gstack.git ~/gstack
cd ~/gstack && ./setup --host codex

# Auto-detect
./setup --host auto

Core Workflow: Think → Plan → Build → Review → Test → Ship

gstack's greatest strength is dividing the entire software development lifecycle into structured stages. Each stage is assigned a specialized professional role.

Stage 1: Planning

/office-hours – YC Office Hours

The starting point for every project. Redefines the product through 6 key questions and generates 2–3 implementation alternatives with effort estimates.

/office-hours

Two modes are available:

  • Startup mode: Sharp questions about product-market fit
  • Builder mode: Exploratory and generative approach
/plan-ceo-review – CEO/Founder Mode

Rethinks problems from the user's perspective to uncover the "10-star product hiding inside the request." Supports 4 scope modes:

  • Expansion: Dream big (feature expansion)
  • Selective Expansion: Cherry-pick opportunities
  • Hold Scope: Maximum quality within current scope
  • Reduction: Minimum viable version
/plan-eng-review – Engineering Manager

Locks in architecture before coding. Visualizes system boundaries, data flow, state transitions, failure modes, and edge cases through diagrams. Generated test plans are automatically passed to /qa.

/plan-design-review – Senior Designer

Rates 7 design dimensions on a 0–10 scale. Detects missing elements like empty states, error states, mobile behavior, and loading states, and identifies vague AI-generated descriptions (AI slop).

Automation tip: Run /autoplan to automatically execute CEO → Design → Engineering reviews in sequence. Only important decisions (taste decisions) are presented to the user.

Stage 2: Design

Skill Role Key Features
/design-consultationDesign Partner Competitive landscape research, typography/color/spacing system design, HTML preview generation, DESIGN.md output
/design-reviewDesigner Who Codes 80-item visual audit, CSS-only fixes (safe), up to 30 fixes, before/after screenshots

Stage 3: Code Review

Skill Role Key Features
/reviewStaff Engineer Detects production failure patterns: N+1 queries, race conditions, injection bugs. Auto-fixes obvious issues
/investigateDebugger Root cause tracing. Follows data flow. Re-evaluates architecture after 3 failed fixes
/codexSecond Opinion Independent review via OpenAI Codex. 3 modes: review/adversarial/consultation

Stage 4: Testing & Security

Skill Mode Description
/qaDiff-aware Automatic on feature branches: tests only affected pages based on git diff
Full Comprehensive exploration (5–15 min), 5–10 documented issues
Quick (--quick) 30-second smoke test
Regression Comparison test against baseline
/qa-onlyReport Only Generates bug reports without code changes
/csoSecurity Audit OWASP Top 10 + STRIDE threat modeling. Confidence gate at 8/10+

Stage 5: Ship & Deploy

Skill Role Key Features
/shipRelease Engineer Sync main → run tests → audit coverage → create PR (one-command pipeline)
/land-and-deployRelease Engineer Merge PR → deploy → production health check
/canarySRE Post-deploy monitoring: console errors, regressions, page failures
/benchmarkPerformance Engineer Page load, Core Web Vitals, resource size measurement with before/after comparison

Browser Automation: /browse

gstack includes built-in Chromium-based browser automation. Compared to MCP protocol, there is virtually no token overhead (MCP consumes 30,000–40,000 tokens per session), with response times of approximately 100–200ms per command.

Key Features

  • Persistent Daemon: Long-running Chromium server eliminates cold starts
  • Ref System: Accessibility tree-based element references (@e1, @e2). Operates safely without DOM mutations
  • Cookie Import: Import cookies from Chrome, Arc, Brave, and Edge
  • Localhost-Only Binding: Blocks network access, per-session Bearer token authentication

Commonly Used Browser Commands

# Navigate to page
browse goto https://example.com

# Snapshot (with element references)
browse snapshot -i

# Click element
browse click @e5

# Fill form
browse fill @e12 "search term"

# Take screenshot
browse screenshot

# Execute JavaScript
browse js "document.title"

Safety Guardrails

gstack provides safety guardrails to prevent accidental execution of destructive commands:

Skill Function
/carefulWarns before destructive commands (rm -rf, DROP TABLE, force-push, etc.)
/freezeRestricts edit scope to a specific directory (lock)
/guard/careful + /freeze combined (maximum safety)
/unfreezeRemoves /freeze lock

Parallel Sprints: Conductor Integration

gstack integrates with Conductor (a macOS app) to run multiple Claude Code sessions simultaneously. Each session operates independently in a workspace isolated via git worktree, enabling parallel development without conflicts.

The workflow is simple:

  1. Register a repository
  2. Deploy multiple agents simultaneously (each with an isolated workspace)
  3. Monitor progress in real-time, review code, and merge

Builder Ethos

gstack follows two core principles:

Boil the Lake

When the cost of completion is only a few minutes more than a shortcut, always complete the work. In the AI-assisted era, "it'll take 2 weeks" should be reframed as "it'll take 1 hour." Pursue 100% completion over 90% solutions.

Search Before Building

"Has someone already solved this?" should always come before "let me design from scratch." Leverage three layers of knowledge:

  • Layer 1 (Proven): Widely used, battle-tested patterns
  • Layer 2 (New and Popular): Latest best practices (requires critical evaluation)
  • Layer 3 (First Principles): Original reasoning specific to the problem

Practical Workflow Example

Here's the complete flow for taking a new feature from planning to deployment:

# 1. Validate product idea
/office-hours
→ Redefine product with 6 key questions

# 2. Determine scope with CEO review
/plan-ceo-review
→ Select from 4 modes to finalize implementation scope

# 3. Lock engineering architecture
/plan-eng-review
→ Generate diagrams, edge cases, test plans

# 4. (Or automate) Full planning pipeline
/autoplan
→ Auto-run CEO → Design → Engineering reviews in sequence

# 5. Code review after implementation
/review
→ Detect production failure patterns + auto-fix

# 6. QA testing
/qa
→ Find bugs + auto-generate regression tests

# 7. Security audit
/cso
→ OWASP Top 10 + STRIDE threat modeling

# 8. Deploy
/ship
→ Tests → Coverage → Create PR

# 9. Post-deploy monitoring
/canary
→ Watch for console errors, regressions

Privacy and Telemetry

  • Default: OFF – Nothing is sent without explicit consent
  • Collected (when opted in): Skill name, duration, success/failure, version, OS only
  • Never collected: Code, file paths, repo names, branch names, prompts, user-generated content
  • Instant disable: gstack-config set telemetry off
  • Local analytics: Always available via gstack-analytics

Troubleshooting

Symptom Solution
Skills not showing cd ~/.claude/skills/gstack && ./setup
/browse fails cd ~/.claude/skills/gstack && bun install && bun run build
Stale install Run /gstack-upgrade or set auto_upgrade: true in config
Windows issues Requires Git Bash or WSL. Node.js must also be on PATH
Claude doesn't recognize skills Add gstack skill list to the project's CLAUDE.md

Key Takeaways

  • gstack is an open-source software factory that transforms Claude Code into a virtual team of 28 specialized roles
  • Think → Plan → Build → Review → Test → Ship → Reflect – covers the entire process
  • Structured workflow starting with /office-hours and ending with /ship
  • Built-in Chromium browser supports real UI testing and screenshot automation
  • /careful, /freeze, /guard prevent catastrophic mistakes
  • Conductor integration enables parallel sprints
  • MIT License – completely free, no premium tier
  • Install in 30 seconds: just git clone + ./setup

References

Comments 0items