RepoDaily · 2026-06-24 · Dataset / Public directory

gstack: Garry Tan's Open-Source AI Engineering Toolkit for Claude Code and Beyond

#8 Dataset / Public directory TypeScript +1,012 garrytan/gstack Open repository

Twenty-three slash-command skills that turn Claude Code into a virtual engineering org — CEO, reviewer, QA, security officer — MIT-licensed and multi-agent ready.

Repo typeDataset / Public directory
Best forSolo founders, technical CEOs, and staff engineers who want structured AI-assisted workflows across planning, review, QA, and release.
Risk levelMedium
Time to evaluate30 minutes to install and run through /office-hours and /review

Primary question: Do these opinionated skill prompts meaningfully improve your shipping cadence compared to ad-hoc prompting?

86/100

RepoDaily adoption score

RepoDaily rates this as 86/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.

Directional score from RepoDaily sources and adoption notes, not a benchmark.Risk: Medium
81Evidence quality

4 source(s) across 3 source category/categories, plus a RepoDaily-specific evidence module when available.

98Installability

8 workflow step(s), 5 next-action step(s), and 1 command/install signal(s) were detected.

69Maintenance confidence

Trending momentum is +1,012 stars, with maintenance/release/issue signals counted when present.

93Production readiness

Risk is marked medium, with 5 security note(s) and 4 explicit skip condition(s).

91Differentiation

3 opportunity lens item(s), 5 alternative(s), and 0 type-specific section(s) support differentiation.

82License clarity

License source or license wording is present.

84Agent / AI fit

6 AI/agent-related signal(s) were detected in the article text and metadata.

Project overview

gstack is Garry Tan's personal collection of 23 specialist Claude Code skills, each exposed as a Markdown-defined slash command. Instead of writing freeform prompts, you invoke roles like /plan-ceo-review, /review, /qa, /cso, and /ship — each backed by a detailed prompt that codifies a specific engineering discipline.

The repo is written in TypeScript with Bun, ships a bundled headless browser binary, and works across at least ten AI coding agents including Claude Code, Cursor, Codex CLI, Factory Droid, and others. Setup auto-detects installed agents and deploys skills to the correct directories.

Tan frames gstack as the tooling behind his own productivity acceleration — 3 production services and 40+ features in 60 days while running YC full-time. Whether or not you accept the LOC-based productivity framing, the skills themselves are well-structured prompts that anyone can fork and adapt.

Problem it solves

  • Unstructured prompting produces inconsistent quality across planning, code review, QA, and release
  • Solo builders and small teams lack the specialist roles (designer, security officer, QA lead) that larger orgs take for granted
  • AI-assisted code review often misses production-readiness concerns because prompts are too vague
  • Browser-based QA testing is hard to automate from within an AI coding session

How it works

  1. Clone the repo into ~/.claude/skills/gstack and run ./setup — Claude Code discovers the skills automatically
  2. Add a gstack section to your CLAUDE.md referencing the available skills
  3. Use /office-hours to describe what you're building; the skill interrogates your idea with structured questions
  4. Run /plan-ceo-review or /plan-eng-review on feature ideas to lock architecture and scope
  5. Run /review on any branch with changes for structured code review
  6. Run /qa on your staging URL — the bundled headless browser opens a real browser session for visual and interaction testing
  7. Run /ship to prepare the PR, or /land-and-deploy for the full release flow
  8. Optionally enable team mode with ./setup --team so teammates get gstack automatically on every clone

The 23 Specialist Skills

  • /office-hours — product interrogation with forcing questions
  • /plan-ceo-review, /plan-eng-review, /plan-design-review, /plan-devex-review — multi-perspective planning
  • /design-consultation, /design-shotgun, /design-html, /design-review — design-focused workflows
  • /review — structured production-readiness code review
  • /qa, /qa-only — browser-based quality assurance on real URLs
  • /cso — security officer running OWASP + STRIDE audits
  • /ship, /land-and-deploy, /canary — release engineering
  • /investigate — root cause debugging methodology
  • /retro — weekly engineering retrospective
  • /document-release, /document-generate — documentation engineering
  • /autoplan — end-to-end feature build orchestration
  • /browse — web browsing via the bundled headless browser

Agent Compatibility

gstack detects which AI coding agents you have installed and deploys skills to the appropriate directories. Explicitly supported agents include Claude Code, OpenAI Codex CLI, OpenCode, Cursor, Factory Droid, Slate, Kiro, Hermes, and GBrain. Adding a new host requires only a TypeScript config file with zero code changes, per the docs.

Operational Learnings

At the end of every skill session, the agent reflects on failures — CLI errors, wrong approaches, project quirks — and logs them to ~/.gstack/projects/{slug}/learnings.jsonl. Future sessions surface these automatically, so gstack adapts to your codebase over time. View accumulated learnings with /learn.

Testing Infrastructure

  • Tier 1 — Static: bun test, free, validates SKILL.md correctness and command references
  • Tier 2 — E2E: spawns claude -p subprocess, ~$3.85/run, tests full skill execution
  • Tier 3 — LLM eval: LLM-as-judge scoring of generated docs, ~$0.15/run standalone
  • All E2E runners are hermetic with API connectivity pre-checks

Who should pay attention?

Good fit if

  • You use Claude Code or another supported agent daily and want structured, repeatable workflows
  • You're a solo founder or small team that needs specialist roles without hiring
  • You want to study well-crafted skill prompts as a starting point for your own
  • You value browser-based QA and security audits integrated into your AI workflow

Skip for now if

  • You don't use any of the 10 supported AI coding agents
  • You work in a regulated environment where external tool auto-updates are prohibited
  • You prefer minimal dependencies and don't want a bundled browser binary
  • You need skills in a language other than English (prompts are English)

Risks and cautions

Medium

MIT-licensed and straightforward to install, but the skills are opinionated prompts that embed Tan's workflow philosophy. Teams should review each skill's prompt before adopting in production.

  • Skills are Markdown prompt files — teams should audit them for alignment with internal practices
  • Team mode auto-updates from the upstream repo, which could introduce unexpected prompt changes
  • E2E and eval test tiers incur API costs (~$3.85/run for E2E)
  • The bundled browser binary adds a non-trivial dependency to your environment
  • Requires Bun v1.0+ which may not be standard in all dev environments
  • Includes a dedicated /cso skill that runs OWASP + STRIDE security audits
  • Team mode uses a throttled, network-failure-safe auto-update check (once per hour)
  • No vendored files are committed to your repo in team mode — skills are symlinked
  • Setup creates symlinks rather than copying files, reducing repo pollution
  • MIT-licensed with source fully auditable in the public repository

Alternatives to compare

ApproachWhen to useTrade-off
Claude Code built-in prompts
You want zero setup and are happy with your own ad-hoc promptingFree
Cursor Rules (.cursorrules)
You primarily use Cursor and want lightweight project-scoped instructionsFree
Aider
You want a terminal-based AI pair programmer without skill-based structureFree
Continue.dev
You want an open-source AI coding assistant for VS Code/JetBrains with custom commandsFree
Custom internal prompt library
Your team has mature workflows and wants full control over prompt designEngineering time

What this trend reveals

Prompt engineering reference

Each SKILL.md is a well-structured, battle-tested prompt. Studying them teaches effective AI coding agent prompt patterns — role definition, step decomposition, output formatting.

Read review/SKILL.md and qa/SKILL.md in the repo to evaluate prompt quality directly.

Internal skill forks

Teams can fork gstack and customize skills for their own stack, conventions, and compliance requirements while keeping the multi-agent deployment infrastructure.

Clone the repo, read CONTRIBUTING.md, and test editing a SKILL.md with bin/dev-setup.

New host support

Adding support for a new AI coding agent is documented as requiring only one TypeScript config file. This is a low-friction contribution path for agent builders.

Read docs/ADDING_A_HOST.md to confirm the scope of work.

Best next action

Install gstack and run two skills

The fastest way to evaluate gstack is to install it in 30 seconds, then run /office-hours and /review on a real project. The README explicitly says to stop there — you'll know if it's for you.

  1. Ensure Claude Code, Git, and Bun v1.0+ are installed
  2. Open Claude Code and paste the install command from the README Quick Start section
  3. Run /office-hours on a feature you're planning
  4. Run /review on a branch with uncommitted changes
  5. Decide whether to enable team mode for your repo

RepoDaily verdict

gstack is a well-crafted, MIT-licensed collection of opinionated AI coding skills backed by Garry Tan's credibility and real daily use. It's most valuable as both a ready-to-use workflow toolkit and a learning resource for how to structure effective AI agent prompts. The multi-agent support and self-improving learnings system set it apart from simple prompt snippets.

Sources