How GStack turns Claude Code into an AI engineering team

Executive overview

Out of the box, AI coding models wander — they guess at context, producing plausible-looking code that silently breaks. The bottleneck is not model intelligence; it is scaffolding and process.

GStack encodes team roles and review loops into Claude Code, replicating how humans do real engineering work.

GStack is a thin harness with fat skills: open-source, built by Garry Tan, now with more GitHub stars than Ruby on Rails. It adds specialist roles — office hours, design, code review, QA, and browser testing — on top of Claude Code.

The office hours skill

• Asks six forcing questions before any code is written: problem, user, pain, business model, feasibility, alternatives.
• Mirrors the reframing YC partners apply in real founder office hours.
• Surfaces business model angles the founder hadn't considered — e.g. document aggregation as a wedge into CPA matchmaking.
• Runs a multi-step adversarial review of the resulting design doc, auto-fixing issues where possible.
• Example: design doc score moved from 6/10 to 8/10; 16 issues caught and fixed automatically.

Planning and design tools

• Auto plan compresses CEO, engineering, design, and developer experience review into one step using Garry's default recommendations — for when you don't want to iterate manually.
• Design shotgun generates multiple visual directions in ~60 seconds by farming image generation to OpenAI Codex, then asks you to rate and select.
• Approved design locks in before code is written, preventing drift.

Code review and QA

• Review runs a staff-level bug-catching pass after code is written — finds issues that plan mode didn't anticipate.
• Slash QA and slash browse wrap Playwright and Chromium at the CLI level, giving Claude Code a real headed/headless browser.
• Browser tool can take screenshots, click, fill forms, download media, run regression tests, and assess JavaScript and CSS bugs.
• Built because Claude-in-Chrome MCP had unbearable context bloat and 2–3 second latency per action.

The parallel workflow

• Garry runs 10–15 parallel Claude Code sessions simultaneously — separate work trees per feature or bug.
• One session: office hours on a new idea. Others: evaluating and merging open-source PRs.
• Can ship 10–50 PRs in a day depending on meeting load.
• Ship skill is the final pre-merge gate to verify a PR is ready to land on main.
• GStack reaches what Garry calls level 7 of the software factory; full level 8 remains out of reach.

What GStack does not do

• Does not replace judgment — office hours rejects roughly one in three ideas as not worth pursuing.
• Does not eliminate supply chain risk — Garry remains manually paranoid about dependency attacks.
• Does not write the code for you unprompted — it structures the process so the model has the right context before it starts.