Jonathon's AI Wiki

gstack — Garry Tan's 23-Tool Claude Code Setup

8 min read

Source: raw/gh-star-garrytan-gstack.md (GitHub star stub, 2026-05-20) + README extraction from github.com/garrytan/gstack; operational-workflow detail from raw/How_Y_Combinator_s_CEO_ACTUALLY_Uses_Claude_Code.md (YouTube walkthrough, 2026-05-20).

gstack is the Claude Code skill pack Garry Tan (President & CEO of Y Combinator) maintains as his own daily-driver, open-sourced as a one-line install. 23+ slash commands map to roles a small startup team would otherwise need to hire — CEO/founder, design partner, eng manager, staff engineer, QA, SRE, technical writer, security officer. MIT license, single-branch clone into ~/.claude/skills/gstack. Tan claims a ~810× year-over-year LOC pace (with the caveat that “AI inflates raw LOC”), reporting 240× the entire 2013 year’s output by April 18, 2026.

Key Takeaways

  • 23+ slash commands grouped into 5 role buckets: planning/design, development/review, testing/deployment, documentation/security, and utilities. Each /command is a persona invocation, not a tool macro.
  • MIT license. No paid tier, no waitlist, no telemetry call-out on the repo surface.
  • 99.9K GitHub stars at the 2026-05-19 capture — but verified by the README’s identity claim that the author is Y Combinator’s CEO. The star count grew rapidly in May 2026 following the YC announcement.
  • One-line install: git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup. Requirements: Claude Code, Git, Bun v1.0+, Node.js (Windows only).
  • Team mode (optional): ./setup --team + gstack-team-init required adds a project-level CLAUDE.md commit that mandates gstack usage across contributors.
  • Productivity claim with caveat: Tan publishes a “logical code change” metric (not raw LOC) to defend against the “AI inflates LOC” critique. 11,417 vs 14 logical lines/day reported. ^[inferred — the methodology behind “logical code change” is not documented in the README]
  • Anti-pattern guard rails baked in: /careful warns on destructive commands, /freeze restricts edits to specific directories, /guard combines both. The pack treats safety as a first-class concern, not an afterthought.

The 23+ Commands by Role

Planning & Design

  • /office-hours (CEO/Founder) — Reframes product ideas with forcing questions.
  • /plan-ceo-review, /plan-eng-review, /plan-design-review, /plan-devex-review — Pre-implementation reviews from CEO, eng manager, designer, and DX-lead lenses.
  • /design-consultation (Design Partner) — Complete design system creation.
  • /design-shotgun (Design Explorer) — Multi-variant visual generation.
  • /design-html (Design Engineer) — Production-ready HTML conversion.

Development & Review

  • /review (Staff Engineer) — Production bug detection.
  • /investigate (Debugger) — Root-cause analysis.
  • /design-review (Designer/Coder) — Live design audits with fixes.
  • /devex-review (DX Tester) — Real onboarding testing.

Testing & Deployment

  • /qa (QA Lead) — Testing with automatic bug fixes.
  • /qa-only (QA Reporter) — Bug reporting without code changes.
  • /ship (Release Engineer) — PR preparation and CI handling.
  • /land-and-deploy (Release Engineer) — Merge-to-production automation.
  • /canary (SRE) — Post-deploy monitoring.
  • /benchmark (Performance Engineer) — Performance baseline tracking.

Documentation & Security

  • /document-release — Auto-updates docs after each ship.
  • /document-generate (Documentation Author) — Creates missing docs via the Diataxis framework.
  • /cso (Chief Security Officer) — OWASP Top 10 + STRIDE threat modeling.
  • /retro (Eng Manager) — Weekly team retrospectives.

Utilities

  • /browse (QA Engineer) — Real browser automation.
  • /codex (Second Opinion) — Independent OpenAI review.
  • /autoplan — Automated multi-review workflow.
  • /learn (Memory) — Cross-session knowledge management.
  • /pair-agent — Multi-AI browser coordination.
  • /careful, /freeze, /guard — Safety layers.
  • /open-gstack-browser — AI-controlled Chromium with sidebar.
  • /setup-deploy, /setup-gbrain, /sync-gbrain, /gstack-upgrade, /unfreeze — Setup and maintenance.

Implementation

Tool/Service: gstack (Claude Code skill pack) Setup:

git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && \
  cd ~/.claude/skills/gstack && ./setup

Cost: Free (MIT). Cost is whatever your Claude Code subscription + token usage runs. Integration notes: Lives under ~/.claude/skills/gstack; no system-wide install needed. Team mode commits a project-level CLAUDE.md gate. Pairs with /learn for cross-session memory; pairs with /freeze + /guard as a directory-scoped safety net.

Differentiation vs Adjacent Frameworks

  • vs Superpowers (Jesse Vincent) — Superpowers is broader (full software-development methodology); gstack is opinionated YC-CEO-shaped (CEO/design/eng-mgr/release/QA roles explicit).
  • vs oh-my-claudecode (OMC) — OMC is team-first multi-agent orchestration with HUD + smart routing; gstack is single-developer “I am the CEO and the team” workflow.
  • vs seven-skills-business — Similar “shipped real workflow” angle, but gstack ships as an installable skill pack while the seven-skills set is reference reading.
  • vs SuperClaude — SuperClaude is meta-programming + 30 /sc:* slash commands + MCP integrations; gstack is role-personas without explicit MCP coupling.

Open Questions

  • “Logical code change” methodology. Tan reports 11,417 vs 14 logical lines/day, but the README doesn’t define how “logical” is measured. AST diff? Reviewer-rated? Self-graded? Worth verifying before citing the 810× figure.
  • Star-count provenance. 99.9K stars at 2026-05-19 is exceptional for a 1-2 month old repo, even with YC CEO endorsement. Worth periodic check for whether the count survives normalization.
  • Persona prompt visibility. READMEs typically don’t reveal the system prompts each persona uses. Reading the actual skills/*.md files to evaluate prompt quality is the next step before recommending.
  • Diataxis framework references. /document-generate cites Diataxis — interesting because Diataxis is a documentation taxonomy (tutorial / how-to / reference / explanation) rather than an automation pattern. Worth comparing against the Anthropic skills doc structure.

Refresh — 2026-05-20 (operational workflow walkthrough)

[YouTube signal — 2026-05-20] Walkthrough video breaks the README’s 23-command list into Garry Tan’s actual four operating strategies, exposing operational detail not surfaced in the repo surface:

  • Strategy 1: Give Claude a role. Three custom-CLAUDE.md primitives within G-Stack make the role system load-bearing — skill routing (describe what’s needed; let the system pick the agent/skill, instead of explicitly invoking), search-before-build (Claude must search the existing code first before writing anything new — kills “reinvent the wheel” at codebase scale), and AI effort compression (not every task needs deep thinking — quick rename = fast response, architecture = deep analysis). Reportedly a major CTO texted Garry that gstack’s /cso role caught a security flaw their entire engineering team had missed: “This is like god mode.”
  • Strategy 2: Parallel sessions, but capped at 5 — not 10. Garry runs 10+ Claude Code sessions simultaneously across independent workspaces (one builds, one reviews, one runs security checks). Source-walker pushes back: “having 10 different Claude Code sessions running simultaneously is an anti-goal.” Recommended ceiling for normal operators is 5 concurrent sessions — sweet spot between parallelism and the cognitive cost of approving across many threads without blindly rubber-stamping.
  • Strategy 3: Build your own tools (massive ROI shift). Garry shipped 3 production services + 40+ features in 60 days, part-time while running YC full-time. Quote: “the cost to build it with AI is literally lower than the cost of explaining what you want to someone else.” The personal tool he calls out is G-Brain — a personal knowledge system that ingests meetings + emails + tweets + calendar into a searchable brain that Claude Code reads before every response. The 240×/810× LOC claim’s operational context: 10K lines/day across ~3 projects, alongside an 8-9-hour-meeting day. Personal-tool-building bar has collapsed.
  • Strategy 4: Plan before you build — /office-hours is the load-bearing command. Forces a structured pre-coding interview: What problem does it solve? Who is it for? What does success look like? What should it NOT do? Source-walker frames this as YC office hours productized — “you can get office hours with Garry without actually being in Y Combinator.” This is the strategy that makes the other three work; skip it and you build fast in the wrong direction.

Operational ranking the source-walker recommends:

  1. Strategy 1 (give Claude a role) — implement today; high-confidence win.
  2. Strategy 4 (plan before building / /office-hours) — the multiplier on everything else.
  3. Strategy 3 (build personal tools) — calculus has changed; stop overthinking ROI.
  4. Strategy 2 (parallel sessions) — value confirmed but cap at 5, not 10. Romanticization of “managing 10 agents like a CEO” produces incomplete output for everyone except Garry.

Counter-signal worth tracking. Source-walker flags founder pushback (“Garry should be embarrassed for tweeting this”) + slop critique on G-Stack itself + Garry’s claim that “over 90% of new repos from today forward will use G-Stack” as overreach. The article’s earlier Open Question about “logical code change” methodology stands — and the source-walker himself signals not being a full believer in the framework even while documenting it.

Try It

  1. Audit-first install. Clone into a sandbox profile (CLAUDE_HOME=~/sandbox-claude ./setup) before merging into your main ~/.claude/skills/ — gstack ships its own opinions on safety, model routing, and project structure.
  2. Read skills/cso.md and skills/review.md first. Security and code review are the highest-impact personas. Verify the prompt quality matches your own standards before using on real code.
  3. Pair /freeze with directory scoping before running /ship or /land-and-deploy on production code — lock the agent out of paths you don’t want touched.
  4. Cross-check claims with the WEO Marketly team if Tan’s “logical code change” metric is going to be cited externally — the methodology isn’t documented in the README.
  5. Test team mode on a side project, not the main repo. The gstack-team-init required flag commits to the shared CLAUDE.md and propagates to every contributor on next pull.

Graph View

Backlinks