gstack

What this is

gstack is Garry Tan's open-source setup for shipping with AI agents: opinionated slash commands that behave like specialists—product interrogation, CEO scope review, eng architecture lock-in, design critique, staff-level code review, real-browser QA, security audits, and release automation. Everything is Markdown skills, MIT-licensed, meant to be installed once and used daily—not a hosted product.

The pitch is familiar from the solo-builder era: one person with the right process can move like a small team. gstack encodes that process as repeatable commands instead of reinventing prompts every session.

Insight into the repo

Process, not a bag of tricks. Skills are ordered like a sprint: think → plan → build → review → test → ship → reflect. /office-hours produces a design doc that /plan-ceo-review and /plan-eng-review consume; /review and /qa feed /ship. Downstream steps read upstream artifacts so context does not reset every time you switch hats.

Roles as commands. Representative skills include /office-hours (reframe the product before code), /plan-ceo-review, /plan-eng-review, /plan-design-review, /autoplan (chained reviews), /review, /qa (browser-backed), /cso (security), /ship, /land-and-deploy, /retro, and /browse for agent-visible browsing. Full catalog and walkthrough live in the repo README.

Install model. Clone to a skills path (e.g. ~/.claude/skills/gstack), run ./setup, wire CLAUDE.md (or project AGENTS.md for OpenClaw). Team mode (./setup --team + gstack-team-init) nudges shared repos toward a common baseline with auto-update from upstream—no vendored copy of the whole stack in every project.

Not Claude-only. ./setup auto-detects hosts; ./setup --host cursor installs into ~/.cursor/skills/gstack-*/ alongside Claude Code, Codex, OpenCode, and others. Same skill shape, different agent surface.

How this relates to Nimbus Lab themes. gstack optimizes understanding and gates before execution—office hours and plan reviews before implementation, QA and review before ship. That rhymes with Execution is not the bottleneck and Optimize for alignment, not autonomy. It is complementary, not identical, to our Cursor Agentic Toolkit setup: gstack is slash-command roles and sprint rhythm; the toolkit is repo runtime, context index, and traceable workflow artifacts in git. Many teams could use both—gstack for session discipline, a committed .agentic/-style layer for durable ground truth.

Benefits

• Structured entry — /office-hours and plan reviews replace a blank prompt when you need intent before code.
• End-to-end loop — Plan, implement, review, browser QA, and PR ship as one vocabulary, not ad-hoc chat.
• Real verification — /qa and /browse push agents toward observable behavior, not “looks done in the diff.”
• Multi-host — Same factory on Claude Code, Cursor, Codex, and more via ./setup --host ….
• Team-friendly — Team init + auto-update reduce “works on my machine” skill drift across contributors.
• Forkable — Open source; adapt skills to your governance without waiting on a vendor roadmap.

Try it: install from the README quick start—/office-hours, /plan-ceo-review, /review, /qa on a staging URL is enough to feel whether the rhythm fits your project.