How Garry Tan (YC CEO) Uses Claude Code: Inside the 23-Tool gstack Setup

"I just open-sourced my entire Claude Code setup I used to average 10K LOC and 100 PRs per week in the last 50 days."

That was Garry Tan, President and CEO of Y Combinator, on March 12, 2026. He dropped a repo called gstack. Six weeks later it has 97k stars, 14.5k forks, and is still pulling 915 stars in the last 24 hours.

Half of X is calling it god mode. The other half is calling it a folder of prompts in a text file. Both sides are partially right. Neither side is showing you the actual table of what is in there.

This post does.

What gstack Actually Is

gstack is an open-source pack of opinionated Claude Code skills. MIT license. One install line. Each slash command embodies a specialist persona with its own priorities, constraints, and outputs. The work moves through a fixed loop: Think, Plan, Build, Review, Test, Ship, Reflect.

As of 2026-05-15, the README ships 23 core skills. The count grew from 6 at launch, to 13 by the TechCrunch coverage on March 17, to 23 today.

That progression matters. The repo is moving fast. Pin the commit you cloned if you want a stable reference.

Why The Source Matters

YC has funded 4,000+ companies. The CEO running that engine is publishing the exact skill setup he uses to ship code. That is the rarest possible signal for a Claude Code workflow.

You can argue with the design. You cannot argue with who shipped it.

The 10K LOC and 100 PRs per week number is self-reported. Frame it that way. The shape of the workflow is the part you can copy.

The Role-Based Architecture

gstack groups every skill under a job title. A role owns priorities, constraints, and outputs. Free-form prompting asks Claude to wear many hats inside one message. Roles split the hats and pass work between them.

Six core roles plus a Chief Security Officer:

• CEO / Founder: Scope, product framing, what to build
• Designer: Visual system, mockups, AI slop detection
• Eng Manager: Architecture, data flow, edge cases
• Release Manager: Sync, test, push, PR, deploy
• Doc Engineer: Docs in sync with shipped code
• QA Lead: Real browser testing, bug fixes
• Chief Security Officer: OWASP Top 10 plus STRIDE threat modeling

The loop reads as one sentence: Think (Office Hours plus CEO review), Plan (Eng plus Design plus DevEx review), Build, Review, Test (QA plus CSO), Ship, Reflect (retro).

The 23 Tools

Sourced verbatim from the gstack README and docs/skills.md, accessed 2026-05-15.

Think and Plan

1. /office-hours (YC Office Hours): Six forcing questions that reframe the product before any code
2. /plan-ceo-review (CEO / Founder): Scope review across four modes: expansion, selective, hold, reduction
3. /plan-eng-review (Eng Manager): Architecture, data flow, diagrams, edge cases, test matrix
4. /plan-design-review (Senior Designer): 0 to 10 ratings per design dimension, flags AI slop
5. /plan-devex-review (DX Lead): Interactive developer-experience audit, three modes

Design

1. /design-consultation (Design Partner): End-to-end design system, research plus mockups
2. /design-shotgun (Design Explorer): 4 to 6 AI mockup variants with taste-memory learning
3. /design-html (Design Engineer): Mockup converted to production HTML, ~30KB, zero deps

Review and Investigate

1. /review (Staff Engineer): Production bug detection with auto-fixes and coverage audit
2. /investigate (Debugger): Systematic root-cause analysis, traces data flow, three-fix limit
3. /design-review (Designer Who Codes): Post-ship design audit and atomic-commit auto-fixes
4. /devex-review (DX Tester): Live onboarding audit with timing and error screenshots

Test

1. /qa (QA Lead): Browser testing, fixes bugs, generates regression tests
2. /qa-only (QA Reporter): Bug reports only, no code edits
3. /cso (Chief Security Officer): OWASP Top 10 plus STRIDE, 17 false-positive exclusions

Ship

1. /ship (Release Engineer): Sync, test, audit, push, open PR, bootstraps frameworks
2. /land-and-deploy (Release Engineer): Merge, run CI, deploy, verify production health
3. /canary (SRE): Post-deploy monitoring, console errors, perf, failures
4. /benchmark (Performance Engineer): Core Web Vitals, resource sizes, before-and-after diffs

Document and Reflect

1. /document-release (Technical Writer): Auto-updates docs, Diataxis coverage map
2. /document-generate (Doc Author): Generates missing reference, how-to, and tutorial docs
3. /retro (Eng Manager): Weekly retro, per-person breakdowns, streak tracking
4. /browse (QA Engineer): Real Chromium browser, ~100ms per command

That is the 23-tool count locked to today. The README is moving fast.

A Real Workflow End To End

A clean session moves through six commands in this order: think, plan, build, review, test, ship.

/office-hours runs first. Six forcing questions reframe the feature. You answer in plain English. Output is a sharper scope.

/plan-ceo-review checks scope across four modes: expand, selective, hold, reduce. Often kills work before it starts.

/plan-eng-review drafts architecture, data flow, edge cases, and a test matrix.

You implement against the plan as Claude normally would.

/review scans for production bugs and applies atomic auto-fixes. Coverage audit runs in the same pass.

/qa opens a real Chromium browser, runs the flow, fixes bugs it finds, and writes regression tests.

/ship syncs, tests, audits, pushes, and opens a PR in one chain.

Each command keeps Claude in one role for the duration of that step. That is the design choice the rest of the system is built on.

The Power Tools Most Posts Skip

Beyond the 23 there is a second tier of utility commands: /codex, /careful, /freeze, /guard, /unfreeze, /open-gstack-browser, /setup-deploy, /gstack-upgrade, /setup-browser-cookies, /setup-gbrain, /sync-gbrain, /autoplan, /pair-agent, /context-restore, /learn. Two CLI binaries also ship with the repo: gstack-model-benchmark and gstack-taste-update.

The big one for teams is /pair-agent. It coordinates Claude Code, Codex, and Hermes against the same task. Garry sits on top of this layer when he runs many sessions at once.

His own framing: "gstack is powerful with one sprint. It is transformative with ten running at once."

What Garry Says, In His Own Words

Three quotes worth keeping near the workflow.

On the philosophy: "A single builder with the right tooling can move faster than a traditional team."

On the cost: "I sleep, like, four hours a night right now. I have cyber psychosis." (SXSW with Bill Gurley, March 2026)

That last quote matters. Throughput at this rate is not the average user's normal day. Treat the headline numbers as a ceiling, not a floor.

The Reception Was Split

The launch trended on Product Hunt and pulled 33k stars in week one. Now 97k. Garry's tweet hit 849k views. A CTO friend called it god mode.

The pushback was loud too. Mo Bitar shipped a critique calling gstack "a bunch of prompts in a text file." Sherveen Mashayekhi argued the visibility came from Tan's YC role, not from the artifact's merit.

The steelman of the critique is fair. gstack does not run novel infrastructure. Each skill is markdown plus a system prompt. You could write the same files yourself in a weekend.

The counter is also fair. Most people will not write the same files in a weekend. They will write three, ship none, and revert to free-form prompting. A finished, opinionated set, vetted by someone shipping 10K lines a week, is a different artifact than your own half-written prompts.

Both readings are correct. Pick the one your team needs to hear.

Five Lessons For Founders

Roles beat prompts. Defining a persona with priorities and constraints produces more consistent output than asking one Claude to wear seven hats.

Forcing functions before code. /office-hours exists because reframing the product is cheaper than rewriting it.

AI slop is a real failure mode. A dedicated Senior Designer skill exists specifically to catch it.

Ship is a workflow, not a button. /ship and /land-and-deploy chain sync, test, audit, push, deploy, verify into one command.

Parallelism multiplies leverage. Garry runs 10 to 15 sprints at once. Solo builders get a structured team. Team leads get a fleet.

Clone gstack In 5 Minutes

You need Claude Code, Git, Bun v1.0+, and Node.js if you are on Windows. A Claude Pro subscription ($20/mo) is required for Claude Code itself.

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

Then follow the loop. Start with /office-hours to reframe the product. Move to /plan-ceo-review and /plan-eng-review. Build. Run /review. Run /qa. Ship with /ship. That is the loop Garry is using to move at his stated rate.

The Honest Bottom Line

The CEO of YC published his exact Claude Code stack. The 23 commands are listed. The criticism is real. Pick the parts that fit your team and leave the rest.

Full breakdown with the complete tool table, comparison to default Claude Code, and where gstack stops: https://www.buildthisnow.com/blog/guide/agents/garry-tan-gstack-claude-code

Tags: #ClaudeCode #YCombinator #AI #DeveloperTools #BuildInPublic