Skip to content

lee-jet/gstack-ai-roles

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

4 Commits
analysis
analysis
Β 
Β 
assets
assets
Β 
Β 
en
en
Β 
Β 
skills
skills
Β 
Β 
zh
zh
Β 
Β 
README.md
README.md
Β 
Β 
README_zh.md
README_zh.md
Β 
Β 

Repository files navigation

🌌 gstack Specialist Roles & Prompt Library: Your AI Software Factory by Garry Tan


"I am shipping more code than I ever have before. In the last 60 days, I've written over 600k lines of production code part-time." β€”β€” Garry Tan, CEO of Y Combinator

gstack is the definitive open-source "AI software factory" by Garry Tan. This repository provides deep-extracted prompt templates and specialist roles to transform AI assistants like Claude Code, Codex, and Gemini into a full-scale Virtual Engineering Laboratory.

πŸ—οΈ Core Workflow: The AI Sprint & LLM Orchestration

gstack is more than just tools; it's a rigorous software development process.

    [ Requirement ]
         |
         v
    [/office-hours] (Product Ideation) ----> [ Output: DESIGN.md & Plan ]
         |                                     |
         +-----------------|-------------------+
                           v
                  [ Triple-Audit Gate ]
         +-----------------+-----------------+
         | [/plan-ceo]     | [/plan-eng]     | [/plan-design]
         | (Strategy)      | (Architecture)   | (Visual Soul)
         +-----------------+-----------------+
                           |
                           v
                    [ AI Development ] <-------- [/debug] (Investigation)
                           |           <-------- [/design-review] (UI Fix)
                           v
                    [ Funnel Audit ]
         +-----------------+-----------------+
         | [/review]       | [/cso]          | [/qa]
         | (Logic)         | (Security)      | (Browser Test)
                           |
                           v
                    [ /ship & /retro ]

🧠 Builder Ethos

The soul of gstack lies in these three core tenets:

Important

1. Boil the Lake

In the AI era, the cost of absolute completeness is nearly zero. If a feature requires 100% test coverage and exhaustive error analysis (Boil the Lake), AI makes "perfect implementation" extremely cheap. Never take shortcuts; build the full robust system because the machine can handle the fatigue.

Tip

2. Search Before Building

A 1000x engineer's first instinct is to find a proven pattern rather than reinvent the wheel. Always look for time-tested patterns or emerging trends (search-first) before first-principles analysis.

πŸ“Š Compression Ratio Table

Task Type Traditional Team AI-Powered Builder Boost Ratio
Boilerplate 2 Days 15 Minutes ~100x
Unit Tests 1 Day 15 Minutes ~50x
Implementation 1 Week 30 Minutes ~30x
Architecture 2 Days 4 Hours ~5x

🎭 The 13 AI Specialist Roles & Prompt Templates

This repository contains the deep-extracted prompt templates for the core engineering roles.

Category Role & Link Shortcut Mission
Strategy 01. CEO Product Thinker /office-hours Refine vision via "6 Soul Questions."
02. CEO Product Reviewer /plan-ceo-review Audit for "10-star experience" & scope control.
03. Engineering Manager /plan-eng-review Technical rigor, architecture & test plan.
Skill Master Orchestrator /autoplan Built-in: One-click CEO -> Design -> Eng Review.
Design 04. Visual Designer Auditor /plan-design-review High-fidelity UI/UX audit & state coverage.
05. Visual Designer Consultant /design-consultation Build bold, unique design systems & visual soul.
Review 06. Code Senior Reviewer /review PR audit, logic safety & performance.
07. Investigation Specialist /debug Systematic root-cause (No Guessing allowed).
Skill Adversarial Auditor /codex Built-in: Cross-model audit using OpenAI Codex.
Implementation 08. Design Implementation /design-review Pixel-perfect UI/CSS visual fixes & offsets.
Automation 09. QA Lead Tester /qa Automated Chromium headless tests & bug fix.
Skill Browser Ninja /browse Built-in: Ultra-fast headless interaction & cookie import.
Governance 10. Chief Security Officer /cso OWASP/STRIDE audit & secret/leak scanning.
11. Release Engineer /ship Versioning, changelog, land-and-deploy & canary.
12. Tech Doc Engineer /document-release Keep README/Architecture/Diagrams in sync.
13. Performance Analyst /retro Weekly data-driven retrospect & global heatmap.

πŸ’Ž Ready-to-use Agent Skill: gstack-ai-roles

In addition to the static prompt templates, this repository includes a ready-to-use Agent Skill engine located in the skills/ directory.

This engine (compliant with the Agent Skills Spec) orchestrates the 13 specialist roles into a 6-phase "AI Software Factory" pipeline:

  • Phase 1-2 (Think/Plan): Strategic Product Alignment & Architectural Blueprint.
  • Phase 3 (Build): Precision Implementation & Atomic Commits.
  • Phase 4-5 (Review/Test): Adversarial Audit & Functional Stress Testing.
  • Phase 6 (Ship): Release Governance & Retrospective.

πŸ‘‰ View the Skill Documentation: skills/README.md | skills/SKILL.md

πŸ€– Multi-Agent & Cross-Model Support

gstack runs seamlessly across multiple AI agents for maximum adversarial review.

Note

1. Using /codex for "Second Opinion"

This is a killer feature: Cross-model adversarial review. Call OpenAI Models inside Claude Code for an independent audit.

  • Review Mode: Pass/Fail gate for code quality.
  • Adversarial Mode: Specifically finds flaws and challenges your path.
  • Consult Mode: Discusses best solutions across multiple models.

2. Global Dashboard (/retro global)

When working across multiple projects or models, use /retro global. It scans all gstack logs to summarize your total contributions, heatmap, and test health across Claude, Codex, and Gemini.

3. I want a GUI, any alternatives?

If you prefer a visual interface, here are tools that integrate gstack ethos:

Name Interface Core Feature
Cursor GUI + Sidebar Most mature. Supports Composer. Place gstack in .agents/skills.
Windsurf GUI (Flow) Silindered "Agent Flows" where AI moves between files.
Trae GUI ByteDance product, very friendly UI with strong localization.
Void Open Source The open-source alternative to Cursor for privacy focus.

🌟 Market Feedback & Community Insights

gstack has sparked intense debate across GitHub, Reddit, and X.

1. Positive: The Developer's "Nuclear Option"

  • Productivity Leap: While 600k lines in 60 days is Garry-level, many report single-handedly doing the work of a 3-5 person team via /ship and /qa automation.
  • Anti-Hallucination: By enforcing independent role audits (e.g., Engineer cannot review own code, must use /review or /codex), it stops the "AI Compliance Spiral."
  • Founder's North Star: /office-hours is praised for challenging business assumptions with a YC partner's perspective, forcing a return to "Real User Needs."

2. Controversies & Challenges

  • "Black Tech" vs "Wrapper"?: Veterans note the tech stack (Playwright, MD templates) isn't secret. Its power comes from exquisite Prompt Engineering and deep process understanding.
  • LLM Bottlenecks: Large projects can lead to "Attention Decay." AI might miss tiny details in long DESIGN.md docs.
  • Vibe Coding Debate: Critics worry about the decay of fundamental engineering skills or rushing into "Overengineering" too fast.

3. Community Ecosystem

  • gstack++ (C++ Deep Refactor):
    • Developed by bulyaki (GitHub).
    • Replaces Web stack with CMake/GTest/Clang-tidy/Valgrind.
    • Focuses on memory safety and data races in C++.
  • The YC Standard: Many early-stage YC teams now bake gstack principles into their CLAUDE.md as the "Digital Employee Handbook."

πŸ’Ž Advanced Integration

1. Greptile PR Review

Native support for Greptile (AI PR Review).

  • Automated Triage: gstack reads Greptile comments and auto-categorizes them.
  • Closed-Loop Fix: Real bugs are patched; FPs are logged; responses are automated to prevent comment pile-up.

2. Design Risks (/design-consultation)

Go beyond "safe" designs. gstack suggests "Deliberate Design Risks" to boost brand identity and generates interactive HTML previews in DESIGN.md.

3. Silent Updates & Standard Parts

  • Auto-Upgrade: Set auto_upgrade: true in config.yaml for silent skill refreshes.
  • Unified Language: Forces a standardized CLAUDE.md, ensuring humans and AI speak the same "Engineering Language."

βš™οΈ Setup & Configuration

1. Basic Setup

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

# Codex / Gemini Installation (Global)
git clone https://github.com/garrytan/gstack.git ~/gstack
cd ~/gstack && ./setup --host codex

2. Project Integration (CLAUDE.md)

Add the full skill list to your project root to ensure the Agent knows its capabilities:

## gstack Rules
Always use gstack's /browse. 
Available Skills: /office-hours, /plan-ceo-review, /plan-eng-review, /plan-design-review, /design-consultation, /review, /ship, /land-and-deploy, /canary, /benchmark, /browse, /qa, /qa-only, /design-review, /setup-browser-cookies, /setup-deploy, /retro, /investigate, /document-release, /codex, /cso, /autoplan, /careful, /freeze, /guard, /unfreeze, /gstack-upgrade.

3. Advanced Guardrail & Utility Commands

For high-stakes environments, use these specialized guardrails:

  • /canary: Executes a Blue-Green deployment strategy manually or via CI.
  • /benchmark: Runs performance stress tests and compares against baselines.
  • /careful: Forces extra-rigorous auditing, ideal for legacy or critical paths.
  • /freeze / /guard: Locks specific files or directories to prevent AI from modifying them during large-scale refactors.
  • /land-and-deploy: Atomic "Pre-flight check -> Land -> Prod Deploy" sequence.
  • /setup-browser-cookies: Imports Chrome/Arc cookies for testing private/auth-gated sessions.

4. Windows Environment

On Windows, you must use a terminal (Git Bash or WSL recommended).

  • Node.js required: gstack falls back to Node.js for browser automation on Windows.
  • Troubleshooting: If skills fail, run cd ~/.claude/skills/gstack && bun install && bun run build.

πŸ†˜ Troubleshooting (FAQ)

  • Skills missing?: Run ./setup in the skill directory to rebuild.
  • Update failed?: Run /gstack-upgrade or pull the latest from GitHub.
  • Privacy?: Run gstack-config set telemetry off. Only anonymous stats are sent by default.

πŸ“– Case Study: Building an "AI Chief of Staff"

  1. Product Definition:
    • User: "I want a calendar summary app."
    • Execute /office-hours.
    • AI Insight: "What you're describing is actually an AI Chief of Staff. Start with the core engine; ship tomorrow."
  2. Strategic Review:
    • Execute /plan-ceo-review.
    • AI Challenge: "The sync is too slow. A 10-star experience must be millisecond-responsive."
  3. Autonomous Build: The AI writes 2400+ lines across 11 files with full logic coverage.
  4. Final Audit:
    • Execute /review: Finds and fixes a subtle race condition in the auth flow.
    • Execute /qa https://staging.myapp.com: Simulates clicks and verifies the event bus.
  5. Ship: Execute /ship to land the PR and deploy.

MIT License | Developed for the Virtual Engineering Era | 2026

About

Extracted Specialist Roles & Prompt Library for gstack (by Garry Tan). Transform Claude Code and GPT-5 into a Virtual AI Software Factory.

Topics

roles prompts ai-agent prompt-engineering claude-code gstack

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors