gstack Part 1: Not a Command List but an AI‑Powered Software Delivery Pipeline

Why the Author and Garry Tan Matter

Garry Tan, the president and CEO of Y Combinator, is a prominent venture capitalist, designer, and engineer who publicly treats AI not as a occasional helper but as a full‑blown productivity system, breaking development into distinct phases where agents intervene.

What gstack Actually Is

gstack is a repository built around SKILL.md that defines a set of specialist skills —each responsible for a specific part of the software delivery chain. Its core claim is to split software development into reusable, coordinated specialists rather than a loose collection of commands.

Skill for redefining problems

Skill for product and design review

Skill for engineering review

Skill for debugging

Skill for real‑browser QA

Skill for release, documentation, and retrospection

Skill for safety guardrails

Thus gstack is not a simple toolbox; it is a methodology that makes AI a stable participant in the delivery pipeline.

The Core Workflow Line

The essential sequence is:

Think → Plan → Build → Review → Test → Ship → Reflect

Each stage maps to a concrete command, e.g., /office-hours for turning vague requirements into a design doc, /plan-eng-review for pre‑implementation engineering review, /qa for browser‑based testing, and /ship for automated final delivery.

How gstack Differs from Typical AI Development Tools

Most AI tools can perform individual tasks, but their workflows are unorganized, leading to instability as prompts grow. gstack externalizes the workflow, providing explicit, repeatable steps and guardrails, which keeps the process reliable over time.

What gstack Enables

Clarify the problem : Commands like /office-hours turn informal ideas into a design document.

Solidify the plan : Commands such as /plan-eng-review and /plan-design-review capture failure modes, test coverage, and responsive behavior before coding.

Validate in a real browser : Skills like /browse, /setup-browser-cookies, /qa, and /design-review let the agent interact with the UI, not just static code.

Automate the final mile : /ship, /document-release, and /retro automate merging, documentation, and retrospection.

Add safety guardrails : Commands /careful, /freeze, and /guard prevent accidental code damage in high‑risk scenarios.

Why Study This Repo

It presents a clear AI‑native engineering mindset, moving beyond “AI writes code” to “AI collaborates across the whole pipeline.”

It provides concrete, actionable workflow designs rather than vague promises.

It helps developers transition from merely using agents to orchestrating them effectively.

Common First‑Time Pitfalls

Documentation may not match generated skill names (e.g., /debug vs. /investigate).

Skill files under .agents/skills/.../SKILL.md are generated, not hand‑written; changes should be made to templates.

Safety skills behave differently across hosts, requiring inspection of the actual generated skill rather than relying solely on README descriptions.

Bottom‑Line Value Statement

gstack is not about making AI do a little more; it attempts to reshape software development into a pipeline that is inherently suited for AI collaboration, turning AI from a code‑writing assistant into a full‑cycle development partner.