🪴 Brain

gstack

4 min read

gstack

garrytan/gstack — 23+ opinionated skills for Claude Code from Garry Tan (CEO of Y Combinator). Turns Claude Code into a virtual team: a CEO who thinks the problem through, an eng manager who freezes the architecture, a designer who catches AI slop, a reviewer hunting production bugs, a QA opening a real browser, a security officer (OWASP+STRIDE), a release engineer opening the PR. Everything as slash commands in Markdown, MIT licensed.

Author’s reference point: in 60 days he shipped 600k+ lines of production code (35% tests), 10-20k lines per day part-time, while writing for YC full-time.

🔗 Links

Description

Download or use

# Paste the prompt into Claude Code (installs globally + appends a section to CLAUDE.md):
# "Install gstack: run git clone --single-branch --depth 1
#  https://github.com/garrytan/gstack.git ~/.claude/skills/gstack
#  && cd ~/.claude/skills/gstack && ./setup ..."

Also works on Codex / Gemini CLI / Cursor / Factory Droid via --host codex|auto|factory.

🗒️ Description

🧩 7-step sprint (Think → Plan → Build → Review → Test → Ship → Reflect)

Key idea: the skills feed each other. /office-hours writes a design doc → /plan-ceo-review reads it → /plan-eng-review produces a test plan → /qa runs it → /review catches bugs → /ship verifies the fix.

SkillRoleWhat it does
/office-hoursYC Office Hours6 forcing questions before code; pushback on framing, alternatives, design doc for downstream
/plan-ceo-reviewCEO/FounderFind the 10-star product within the request; 4 modes (Expansion/Selective/Hold/Reduction)
/plan-eng-reviewEng ManagerArchitecture, data flow, ASCII diagrams, edge cases, test matrix
/plan-design-reviewSenior Designer0-10 score per design dimension; AI Slop detection; AskUserQuestion per choice
/design-consultationDesign PartnerFull design system from scratch (research + creative risks + mockups)
/design-shotgunDesign ExplorerMany variants, comparison board in the browser, taste memory
/design-htmlDesign EngineerProduction HTML with Pretext (text reflow, dynamic heights), framework detection
/reviewStaff EngineerBugs that will pass CI and blow up in prod; auto-fix + completeness gaps
/investigateDebuggerIron Law: no fixes without investigation; 3-fail stop rule
/csoChief Security OfficerOWASP Top 10 + STRIDE; 17 false-positive exclusions, 8/10+ confidence gate
/qaQA LeadReal browser, clicks the flows, fixes them, regression test per fix
/shipRelease EngineerSync main + tests + coverage audit + push + PR; bootstraps a test framework if missing
/land-and-deployRelease EngineerMerge → wait for CI → deploy → verify production health
/canarySREPost-deploy monitoring (console errors, perf, page failures)
/codexSecond OpinionIndependent review with OpenAI Codex CLI; 3 modes (review/adversarial/consultation)
/retroEng ManagerWeekly retro per-person; /retro global cross-projects + cross-AI (CC/Codex/Gemini)
/learnMemoryCross-session learnings: review/search/prune/export project patterns

🧩 Power tools

/careful (warning before rm -rf/DROP TABLE/force-push), /freeze (lock edits to a single directory), /guard (= careful + freeze), /connect-chrome (Side Panel extension, watch live), /setup-deploy, /gstack-upgrade.

🧩 Real browser mode

$B connect launches your Chrome as a headed window controlled by Playwright — you see every click the agent makes live. The side panel = chat sidebar for directly steering Claude. $B handoff for CAPTCHA/MFA: open a visible Chrome with cookies, solve it, $B resume.

✍️ Reasoning for

The strongest pieces for me: /qa with a real browser + /codex as a cross-model second opinion + /retro global. These are exactly the spots in my workflow on Qamera AI / PLSoft where I lose the most time on manual checking. /document-release also sounds like something I need for Brain (auto-updating README/CLAUDE.md/CONTRIBUTING).

Risks and limitations:

  • 23+ skills = a large surface to learn. Plan: start with /office-hours + /review + /ship + /qa, the rest as needed.
  • Opinionated stack — some bits (Bun-only, Pretext) may clash with existing conventions in a repo.
  • Lots of overlap with Superpowers (TDD/brainstorm/plan/review) — I will need to pick one methodology, not both.

Alternatives considered

  • Superpowers — more methodology-first (TDD, subagent-driven), gstack is more rolling-team
  • Karpathy Skills — minimal 1-CLAUDE.md answer, gstack is maximalism in the other direction
  • Archon — YAML workflow engine + worktree isolation; gstack stays in native Claude Code skills
  • Awesome Claude Code — curated list, gstack is a ready-made stack

🔗 Resources

Template: tool

Graph View

Backlinks