How Garry Tan Ships 10,000 Lines of Code Per Day While Running YC (Steal His Workflow)

Garry Tan, president and CEO of Y Combinator, shipped 600,000+ lines of production code in 60 days. That's 10,000-20,000 lines per day. Part-time. While running the most important startup accelerator on the planet.

His last weekly retro across 3 projects: 140,751 lines added, 362 commits, ~115,000 net lines of code. In one week.

This isn't a theoretical framework. It's a documented, open-source workflow called gstack — and you can use it today.

The Mental Model That Changes Everything

Most people use AI as a copilot — an autocomplete that finishes your sentences. Garry uses AI as a team. The difference is not incremental. It's 100x.

His system, gstack, assigns Claude Code specialized roles:

| Role | What It Does | |------|-------------| | CEO / Founder | Rethinks the product — finds the 10-star version hiding in your feature request | | Eng Manager | Locks architecture, data flow, edge cases, and test plans | | Senior Designer | Rates design quality 0-10, explains what a 10 looks like, then fixes it | | Staff Engineer | Reviews code for production bugs that pass CI but blow up at 3am | | QA Lead | Opens a real browser, clicks through flows, finds and fixes bugs, writes regression tests | | Security Officer | OWASP Top 10 + STRIDE threat models with concrete exploit scenarios | | Release Engineer | Syncs main, runs tests, pushes, opens PR | | Technical Writer | Updates all docs to match what just shipped |

One person. Eight specialists. All running in parallel.

The Sprint Loop: Think → Plan → Build → Review → Test → Ship

This is the sequence that makes it work. Each step feeds into the next:

1. Think — /office-hours Start by describing what you want to build. The AI pushes back on your framing — "you said daily briefing app, but what you actually described is a personal chief of staff AI." It asks 6 forcing questions that reframe the problem before you write any code.

2. Plan — /plan-ceo-review → /plan-eng-review The CEO agent challenges scope (should you expand? reduce? hold?). The eng manager agent locks architecture with ASCII diagrams for data flow, state machines, and error paths. You get a test matrix and failure modes before writing a single line.

3. Build — Implementation With a reviewed plan, Claude Code implements across multiple files. The plan context means the AI writes better code because it understands the architecture, not just the current file.

4. Review — /review A staff engineer reviews the diff. Auto-fixes obvious issues. Flags production risks and completeness gaps. This catches the bugs that would ship to users.

5. Test — /qa The QA agent opens a real Chromium browser, navigates your app, clicks through user flows, finds bugs, fixes them with atomic commits, and writes regression tests for every fix. This is where Garry said it "let me go from 6 to 12 parallel workers."

6. Ship — /ship Run tests, audit coverage, push, open PR. If you don't have a test framework, it bootstraps one from scratch.

The Parallel Sprint Breakthrough

Here's where it gets wild: Garry runs 10-15 of these sprints simultaneously.

One session doing office hours on a new idea. Another reviewing a PR. A third implementing a feature. A fourth running QA on staging. Six more on other branches. All at the same time.

Without the process structure, 10 agents would be 10 sources of chaos. With the sprint loop — think, plan, build, review, test, ship — each agent knows what to do and when to stop. You manage them the way a CEO manages a team: check in on the decisions that matter, let the rest run autonomously.

How to Apply This Even If You're Not a Developer

The mental model scales beyond code:

For operators and managers:

• Think: Use AI to challenge your project brief before execution
• Plan: Have AI create a detailed implementation plan with risks and dependencies
• Build: Delegate the execution tasks to AI (research, writing, analysis)
• Review: Have AI audit the output for completeness and quality
• Ship: Use AI to format, distribute, and track results

For founders:

• Run competitive analysis, product strategy, financial modeling, and content creation as parallel workstreams
• Each "agent" (Claude project, custom GPT, or automated workflow) handles one workstream
• You make the strategic decisions; AI does the execution

Getting Started with the gstack Approach

You don't need to install gstack to use the mental model (though you can — it's free and MIT-licensed):

Step 1: For any project, start by writing a CLAUDE.md or system prompt that defines your stack, conventions, and constraints. This is the single highest-leverage thing you can do.

Step 2: Before building, ask AI to challenge your plan. "Act as a product advisor. Here's what I want to build. Push back on my framing. What am I not seeing?"

Step 3: After building, always review. "Review this [code/document/plan] as a senior [engineer/consultant/editor]. Find the issues that would be obvious in production but aren't obvious right now."

Step 4: Always test with fresh eyes. Don't let the same AI that built something review its own work without a new context window.

Why This Matters for Your Career

Andrej Karpathy said he hasn't typed a line of code since December 2025. Garry Tan ships more code now than at any point in his 20-year career. The Karpathy quote that started the gstack README:

"I don't think I've typed like a line of code probably since December, basically, which is an extremely large change."

The people who figure out how to work with AI agents as a team — not just as an autocomplete — will operate at a fundamentally different scale than everyone else. This isn't a marginal improvement. It's a category shift.

Join AI Builder Club to learn these workflows and build alongside people who are pushing the frontier.