Inside Garry Tan's Claude Code Setup
In a Nutshell
Garry Tan, Y Combinator CEO, shares his Claude Code setup via GStack—an open-source tool (70k+ GitHub stars) that turns AI into a specialized engineering team with skills like Office Hours (YC-style product reframing), adversarial reviews, and design generation. Demo builds a tax app extracting 1099s from Gmail and banks via AI browser automation, refining a half-baked idea into a feasible MVP through structured reviews that fix bugs, handle edge cases, and score designs. In the agent era, thin scaffolding unlocks models' potential for 10-50 daily PRs; build fast—what will you make?
These notes were generated by AI and may contain inaccuracies.
Garry Tan is president and CEO of Y Combinator. He's an engineer who spent the first decade of his career building software full-time. He studied computer systems engineering at Stanford, was employee number 10 at Palantir as engineer, designer, and product manager. He co-founded Posterous, a microblogging platform sold to Twitter, and built the first version of Bookface, YC's internal social platform and knowledge base. He's written a lot of code in his career.
We're in a completely new era of building software: the agent era. The way to get agents to do real work is like humans: as a team with roles, process, and review. He built GStack to encode this 3 weeks ago, now with more GitHub stars than Ruby on Rails. GStack turns Claude Code into an AI engineering team. He's coded more in the past two months than all of 2013, the last time he worked hard as an engineer.
Started playing with Claude Code in January after hearing Andrej Karpathy and Boris Cherny say they weren't manually writing code anymore. Got hooked. Built all of Posterous (originally took 2 years, $10M, 10 engineers) using it.
Out of the box, the model wanders, doesn't know your data, guesses—leading to plausible-looking code that silently breaks. Bottleneck is not model's intelligence; once set upright, models are smart enough for extraordinary work. Scaffolding should be thin harness, fat skills. GStack is his open-source repo implementing this: skills acting like a team of specialists.
Office Hours skill modeled after YC partners' office hours with startups. Starts with six forcing questions to reframe product before building.
Demo in Conductor (GStack built-in): quick start → GStack → build tax app to fish 1099s from Gmail (tax day). Uses Office Hours first—distilled 10% strength version of thousands of hours from 16 YC partners.
User idea: app to extract 1099s from Gmail and financial institutions (banks email tax docs, some don't; search inbox + accept URLs to download 1099-INT PDFs).
Gary mode shows reasoning traces. Model thinks: fresh project, initial commit, no prior docs, startup mode. Asks: What's the strongest evidence someone wants this?
Garry's response: Personal pain yesterday hunting 1099s from >5 bank accounts; accountant sends annoyed emails/phone calls.
Model notes pain is friction/annoyance, not penalties. Points out TurboTax/HR Block have 1099 import; Plaid connects banks—why not solving?
Garry expands to funnel: hook is find 1099s; expansion to tax prep matchmaking/lead gen for CPAs (wedge strategy). Charge % of transaction (10x more than $2-5/mo/year for aggregation).
Revised plan: AI browser automation—user logs in, AI navigates/downloads PDFs in visible browser (no Plaid/stored creds).
Likes browser automation on user's actual browser (not cloud). GStack uses Codex to fix bugs.
Claude (Opus) is ADHD CEO (billion ideas, fun); Codex is autistic CTO for tough work.
Office Hours yields plan with 3 approaches:
- A: Gmail search for tax doc notifications → checklist of banks issuing 1099s (small effort/risk).
- B: Full-stack Gmail + AI browser automation + CPA marketplace (what Garry wants).
- C: CPA-first go-to-market.
Garry picks B, adds: use browser to skip Google OAuth, search Gmail automatically; ask for other banks; detect CPA from emails.
End of Office Hours: from half-baked idea to browser automation for inbox search/downloads + bank logins + email to CPA. Feasibility check often kills bad ideas (1/3 times).
Multi-step adversarial review: catches/fixes issues (no failure handling, privacy, 2FA). Doc survives 2 rounds, auto-fixes 16 issues, score 6/10 → 8/10 (3 issues remain). Approve design doc.
Skip Plan CEO Review → Design Shotgun: generates multiple AI design versions (farms to OpenAI Codex + image gen). 3 options for main checklist dashboard (~60s):
- A: Command Center (4/5 stars, Linux hacker style).
- B: Friendly card-based with progress ring (5/5, pick).
- C: Split View (too complicated).
Select B, lock in.
2 of 28 commands shown. >70,000 GitHub stars. Users spend 80-90% time in Office Hours, Plan CEO Review, Auto Plan.
Sprint process: Office Hours → CEO/Engineering/Design/DX Review → code → Review (staff-level bug catch) → CLI around Playwright/Chromium (headed/headless browser for QA: screenshots, interactions, downloads, regression tests, CSS/JS bug assessment) → Ship tool (PR readiness).
Runs multiple Conductor windows (3-4 per project): parallel PRs/branches/features. Automates QA (least fun part).
Prior tools like Claude-in-Chrome MCP slow (2-3s/action, context bloat). GStack's SLQA and SL Browse wrap Playwright CLI: fast browser use for agents.
Runs 10-15 parallel Claude Code sessions: Office Hours on new ideas, community PRs (400 pending across open-source projects with tens of thousands stars). Waves of evaluation. Paranoid about supply chain attacks—GStack protects.
No to-do list: + icon in Conductor → new work tree → Office Hours/CEO/Adversarial/Review → land. Does 10-50 PRs/day based on meetings.
GStack: github.com/gritan/GStack (likely github.com/garrytan/GStack). Office Hours gives YC-style thinking/pushback/reframing.
This is the most incredible time in history to build software. The barrier to building just collapsed. The only question left is what are you going to build? It's time to let it rip. Go make something people want.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
