gstack: Claude Code's 23-Skill AI Agent Workflow
/gstack
gstack packages 23 role-based Claude Code skills — CEO, eng manager, QA lead, security officer, release engineer — into slash commands that walk a feature from idea to a shipped PR. Reach for it if you already run Claude Code solo and want forced checkpoints like /review and /cso before every merge. Skip it if you want a two-minute setup instead of an opinionated, multi-step process.
Understanding gstack: An AI Software Factory
gstack is Garry Tan's open-source pack of 23 specialist Claude Code skills plus 8 power tools, each a Markdown prompt triggered by a slash command like /office-hours, /review, or /ship. It threads a project through a fixed sprint order — plan, build, review, test, ship, retro — instead of leaving one long chat to freelance the whole thing.
The Challenge: Shipping Software with AI Agents
A developer running Claude Code alone in one long chat skips the checkpoints a real engineering org has by default: nobody separately reviews the UI for AI slop, nobody runs a security pass before merge, nobody re-tests in a live browser before shipping. Tan built gstack after a 60-day run of his own — 3 production services and 40+ features shipped part-time, alongside running Y Combinator — that convinced him the missing piece wasn't the model. It was structure around it.
Key Capabilities: gstack's Virtual Engineering Team
- ✓23 role skills cover CEO, eng manager, senior designer, DX lead, staff-engineer reviewer, QA lead, and chief security officer, each a Markdown file invoked as a slash command.
- ✓/cso runs an OWASP Top 10 + STRIDE threat model with 17 named false-positive exclusions and an 8/10+ confidence gate, so findings come with a concrete exploit scenario instead of a raw scanner dump.
- ✓/office-hours opens a sprint with six forcing questions and writes a design doc that /plan-ceo-review, /plan-eng-review, and /plan-design-review all read from downstream.
- ✓/review finds bugs that pass CI but break in production, auto-fixes the obvious ones, and flags the rest — a race condition, say — for you to approve.
- ✓/qa drives a real Chromium browser against a staging URL, fixes what it finds, and auto-generates a regression test for each fix.
- ✓/ship syncs main, runs the test suite, audits coverage, pushes, and opens the PR; /land-and-deploy carries that through CI and a production health check.
- ✓Team mode (./setup --team) keeps the skill files out of the repo — no vendoring — and checks for updates once an hour, silently, without failing on a bad connection.
Who Benefits from gstack's AI Workflow?
- •Solo founders who want a /plan-ceo-review pushback pass before writing code, so scope doesn't quietly balloon mid-sprint.
- •Tech leads who want every PR — theirs or an agent's — to pass through the same /review and /cso sequence before merge.
- •First-time Claude Code users who'd rather type /office-hours than stare at a blank prompt and guess what to ask for.
- •Teams sharing a repo who want gstack required for everyone via team mode instead of each person installing it separately.
Getting Started: Installing gstack
Installation is documented and takes about 30 seconds for Claude Code. Requirements: Claude Code, Git, Bun v1.0+, and Node.js (Windows only). Paste one line into Claude Code — git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup — and Claude runs it, then adds a gstack section to CLAUDE.md itself. For a shared repo, ./setup --team plus gstack-team-init required (or optional) commits a team-mode config, so nobody vendors the skill files and everyone gets hourly, silent auto-updates. For other agents — OpenAI Codex CLI, OpenCode, Cursor, Factory Droid, Slate, Kiro, Hermes, GBrain — clone to ~/gstack and run ./setup --host <name>; setup auto-detects which agents are already installed. OpenClaw needs no separate install: it spawns Claude Code sessions via ACP, so any gstack skill just works once Claude Code has it. Four OpenClaw-native skills install separately through clawhub install.
Using gstack: A Full Development Sprint Example
The README walks through one sprint end to end. Describe a feature — a daily briefing app for a calendar, in its example — then run /office-hours; it pushes back on the framing, pulls out capabilities you didn't name, and writes a design doc. /plan-ceo-review and /plan-eng-review read that doc and add a 10-section strategic review plus ASCII data-flow diagrams and a test matrix. Approve the plan, and Claude writes it — 2,400 lines across 11 files in about 8 minutes, in the README's example. /review auto-fixes two issues and asks approval on a race condition. /qa opens a real browser against the staging URL, finds a bug, and fixes it. /ship runs the tests (42 to 51, nine new ones) and opens the PR. Eight commands, no manual handoffs in between.
Strengths
- ✓Every skill is a plain Markdown file behind a slash command — inspectable and forkable, nothing compiled or hidden.
- ✓The pipeline enforces its own order: /office-hours writes the doc /plan-ceo-review reads, /plan-eng-review writes the test plan /qa executes, so a step can't get skipped by accident.
- ✓/cso filters security noise on purpose — 17 named false-positive exclusions and a confidence gate — instead of dumping a raw scan.
- ✓MIT-licensed and free; Tan documents his productivity methodology and links a reproduction script rather than just stating a multiplier.
- ✓The same setup script auto-detects and installs for Cursor, OpenAI Codex CLI, OpenCode, Factory Droid, and five more agents beyond Claude Code.
Dependencies and Prerequisites for gstack
- △The productivity numbers — a claimed 810x jump in logical lines shipped per day versus 2013, 240x the 2013 year's output year-to-date — come from Tan's own repos and his published methodology doc, not an independent benchmark.
- △Team mode and the Windows path add steps past the base 30-second install: Bun v1.0+, Node.js on Windows, and a commit to CLAUDE.md and .claude/ in your repo.
- △23 skills plus 8 power tools is a lot of surface. Most people won't touch /make-pdf or /diagram in a normal week, and picking between /plan-eng-review, /plan-devex-review, and /autoplan takes some trial.
- △The whole thing is opinionated by design: a fixed review order, a mandated security pass, a specific sprint shape. It fights an existing process more than it slots under one.
- △Every skill runs on whatever model backs your agent, so output quality and API cost are bounded by that provider — gstack doesn't change what Claude Code or Codex can and can't do.
gstack Alternatives and Integrations
Frequently Asked Questions about gstack
Per the README's install section, gstack requires Claude Code, Git, Bun v1.0+, and — on Windows only — Node.js.
gstack works with Claude Code plus nine more agents: OpenClaw, OpenAI Codex CLI, OpenCode, Cursor, Factory Droid, Slate, Kiro, Hermes, and GBrain, each reachable with a --host flag during setup.
gstack works with OpenClaw directly: OpenClaw spawns Claude Code sessions through ACP, so every gstack skill installed for Claude Code runs automatically inside it. Four skills — office-hours, ceo-review, investigate, retro — also install natively via ClawHub, no Claude Code session required.
gstack is released under the MIT license, free to use, fork, and modify — the README explicitly invites forking it.
gstack replaces one long unstructured chat with a fixed sequence — plan, review, test, ship — so a security audit or browser QA pass happens by default instead of getting skipped under deadline pressure. Tan's own methodology doc claims roughly a 810x jump in logical lines shipped per day versus his 2013 pace.
Not really. gstack assumes you're already running Claude Code or a similar coding agent from a terminal, cloning a Git repo, and reading slash-command output; the README targets founders, first-time Claude Code users, and tech leads — all of whom write code.
Who should try it — and who should skip
gstack fits a solo founder or small team already living in Claude Code who wants forced checkpoints — a CEO-style scope challenge, a security pass, a browser QA run — before code ships. It's a worse fit if you want a two-minute setup, a light single-command assistant, or you're not yet comfortable cloning a repo and running a shell script; the whole workflow assumes a terminal, not a GUI.
Related repositories
Curious whether gstack is right for you?
Let ChatGPT, Claude, or Perplexity look into it — click below and see what AI actually says about gstack.