The Builder Returns
After 13 years away from coding, Gary Tan did something remarkable: he built multiple production-ready applications—including a full-featured blogging platform with AI-powered journalism—while running Y Combinator full-time. And he did it all in a matter of weeks.
"I'm relatively shocked myself," Gary admits. "It was 13 years of not coding and then suddenly boom, I'm doing about 400x the amount of work that I was doing that year."
This isn't just another AI hype story. It's a blueprint for how software development is fundamentally changing—and why your productivity as a developer is about to 10x.
Gary's List: When Your Blog Writes Investigative Journalism
It started with a simple need: Gary wanted to bring together people who cared about improving California's education system. So he built Gary's List—a blogging platform that became something much more interesting.
The progression:
- 2008 (Posterous v1): $4M, 6-7 people, 18 months
- 2013 (Post Haven v2): $100K, 2 people, 3 months
- 2024 (Gary's List v3): $200 (Claude Pro Max), 1 person, 5 days
But here's where it gets wild: Gary's List doesn't just let you publish articles—it writes investigative journalism for you.
"For the equivalent of $5 or $10 of Opus calls, I would estimate that it does the work of a real human being that would have to painstakingly through dozens of articles, read entire books about certain subjects, annotate them," Gary explains.
Token Maxing: The New Paradigm
Gary discovered what he calls "token maxing"—the idea that you should use as much AI compute as necessary to get complete, thorough results:
"You don't have to settle for what we did when we were humans writing the code. What if you absolutely boiled the ocean? If you were a human this would take you about a month to do this research—you can just zap the rocks harder."
The Gary's List system:
- Deep research: Crawls entire web, reads all relevant sources
- Cross-references: 13 sources say this, 7 disagree—it knows
- Full citations: Every claim sourced and verified
- RAG integration: Recursive crawling, agentic retrieval
All of this for less than $10 in API calls. Per article.
GStack: Accidentally Building the Perfect AI Development Stack
Gary didn't set out to build GStack. He just got tired of typing the same prompts over and over.
"All I did was I realized that I was doing the same things over and over again and then I got sick of typing the same thing. So I went into my Apple notes, typed in all the things that I found myself writing over and over again into Claude Code."
The Core Skills
1. Plan-Review The proto-skill that started it all. Key insight: Make Claude draw ASCII diagrams first.
"Once I started saying 'actually before you start your work make an ASCII diagram of all the data flows, all the inputs and outputs,' it loaded all of the context in and then it just did the work more completely."
2. CEO Plan (10-Star Experience) Inspired by Brian Chesky's famous question: "What would a 10-star experience look like?"
Uses metaprompting to push beyond the obvious solution:
- What's the 10x check?
- What delivers 10x more value for only 2x the effort?
- What's more ambitious?
3. Developer Review + Testing Gary learned this the hard way: "When I was writing code myself, I would always do the minimum amount of testing because it was just not very fun. That's when I realized, oh, I can get to 100% test coverage."
The machine doesn't care—it'll write comprehensive tests without complaining.
4. QA with Playwright Got tired of manually testing? GStack's QA skill:
- Spins up a browser using Microsoft Playwright
- Reads what changed in your branch
- Tests the UI automatically
- Reports back
"It's like having a blackbox browser. It blew my mind when it first worked. Mini AGI is already here."
The Workflow
Here's Gary's actual daily workflow (he dropped 13 PRs in 48 hours):
- Come up with idea
- Run
/ceoskill in plan mode - Run design review (if UI)
- Run developer review
- Run
/reviewfor comprehensive testing - Approve → Claude implements
- Run
/qato automatically test in browser - Run
/codexfor final bug sweep - Ship
"I have bug fixes right next to giant features. The cool thing about GStack is when I run it through this program—office hours, CEO review, design—it asks me questions along the way."
GBrain: Your Personal Knowledge Base
After building Gary's List and GStack, Gary wanted to bring all his coding patterns into his AI assistant. The result: GBrain, a RAG system for OpenClaw.
The insight came from Andrej Karpathy's post about "knowledge LLM wikis" and realizing:
- GPT in OpenClaw was loading too much into context
- He already had learned patterns from building Gary's List
- Why not extract those patterns and make them reusable?
"I could just open this project in conductor and then the first thing I write is like, go look at Gary's List, look at how we do chunking, embedding, hybrid RRF, RAG—extract it and use it for my OpenClaw."
The system uses:
- Vector embedding
- Hybrid RRF (Reciprocal Rank Fusion)
- Smart chunking
- PostgreSQL with pgvector
Now when Gary builds something new, the patterns from previous projects are automatically available.
Thin Hardness & Fat Skills
Gary crystallized his philosophy in a post about "thin hardness and fat skills":
Fat Skills (Markdown):
- LLM instructions
- Handles generic cases
- Understands context and motivation
- Written in plain English
- Like a wedding planner's checklist
Thin Hardness (Code):
- Deterministic logic
- Real actions (API calls, database queries)
- Brittle but reliable
- Unit tested
- Traditional programming
"All of the difficulty in agentic engineering today is when people try to do things that should be in markdown in code, and it fails because code is brittle—it doesn't understand special cases. Code literally doesn't understand what you want or who you are."
Update — July 16, 2026: Claude Code engineer Thariq (@trq212) replied to a thread citing this framing with a refinement — thin prompts, thick artifacts, thin skills: most public SKILL.md files are too big; depth belongs in skill folders and artifacts, not the entry point.
The magic is knowing which is which:
- Markdown: "Call 20 venues and find one that's available"
- Code: The actual Twilio API call
The Lines of Code Controversy
Gary got roasted on Twitter for saying he coded at "100x" his previous rate. Then he ran the numbers properly.
The reality:
- Stripped down to logical lines of code (removed comments, formatting)
- Found that his own 2013 code dropped 70% after normalization
- His Claude Code output stayed roughly the same
- Actual multiplier: 400x
But here's what people miss: "People get very upset because it's easy to pad lines of code if you're a human writing code. Whereas Claude Code doesn't do that—it'll maybe build the wrong thing, but it's not trying to optimize for lines of code the way a human working a job would."
The industry standard for a professional software engineer: 30-50 logical lines of production-ready code per day.
For Gary in 2013: 14 lines per day (part-time coding).
With Claude Code: Thousands of lines per day, fully tested.
Ferrari vs. Mechanic: The Current State of AI Coding
Gary has the perfect analogy:
"Using OpenClaw these days is like driving a Ferrari. It's exhilarating, insane—it figures things out you would never think a machine could figure out, so quickly. But it's also like a Ferrari in that you better be a mechanic. It's a Ferrari that will break down on the side of the road when you most need it, and you need to get out with your wrench and pop the hood and fix it yourself."
This is where we are: Homebrew Computer Club phase.
The Apple I was a breadboard in a wooden case. If you wanted a personal computer in 1976, that's what you had to do.
If you want AI-powered development in 2026, you need to:
- Spend 2-3 hours learning
- Invest $500-1,000 in tokens and compute
- Be willing to fix things when they break
"But once you get it, it's like we're in the kit car Ferrari phase. You can drive anywhere."
The Personal AI Revolution
Gary sees where this is going:
"Every single person on the planet will have their own personal AI. We could either live in a world where we have our own AI, where we have our own data, our own integrations—or it's corporate controlled, like your Facebook feed, where you don't know who wrote that algorithm."
The key question: Will you have control over your own tools, or will your tools have control over you?
"Unless you have your own prompts and you can write it for yourself, you are below the API guideline for some PM or developer that is not you, who will not understand your needs, will not understand what you uniquely care about."
Time Billionaire: Buying Millions of Years of Consciousness
Here's Gary's most profound insight:
"I am in a crazy rush in my brain. I could probably live 10 billion lifetimes to live in this body right now, and I need every single moment to count. But if you can token max, it's like—I can be a time billionaire. It's not my own time. It's the time of a machine doing work for me."
You can buy infinite time by borrowing time from the machines.
This reframes everything:
- San Francisco rent is expensive → But more expensive not to be there
- Token maxing is expensive → But more expensive not to do it
- Time is scarce → But machine time is abundant
"As long as I'm building something that's actually of great value to me, I'm going to spend $500 in a single day on tokens."
Practical Lessons for Developers
1. Start with Real Pain
Gary built Gary's List because he cared about education policy. The motivation kept him pushing through the rough edges.
2. Use ASCII Diagrams First
Before Claude starts coding, make it draw:
- Data flow diagrams
- State machines
- Dependency graphs
- User flows
This loads the context properly.
3. Get to 80-90% Test Coverage
"If it's not tested and you're just throwing users in there, it's slop—10x worse than human-written code."
Let the machine write the tests. It doesn't get bored.
4. Token Max Everything
Don't settle for the 80% solution when you can have 100%:
- Read all 20 sources, not just 3
- Generate all edge case tests
- Build comprehensive documentation
- Cross-reference everything
The machine doesn't care. Zap the rocks harder.
5. Build Your Skills Library
Every time you type the same prompt twice, make it a skill:
- Review patterns
- Testing approaches
- Architecture decisions
- Design principles
6. Use Multiple Agents
Gary uses both Claude Code and Codex:
- Claude Code: The ADHD CEO (fast, creative, broad)
- Codex: The 200 IQ CTO (deep, analytical, debugger)
"/codex" finds all the problems Claude missed.
7. Keep Humans in the Loop
"It would surprise me very much if someone really truly did manage to make a thing that could just make software without the human in the loop. I never want to be entirely out of the loop. I just want the machine to do the stuff that I don't want to do."
Use ask_user_question liberally. Supply the agency.
The Disconnect: Why People Don't Get It
Gary identifies three barriers:
1. Model Access "A lot of these capabilities, you have to be on the latest and greatest models. It's actually quite expensive to use them and burn all the tokens—for now."
People try the free tier and bounce.
2. Time Scarcity as an Advantage "The fact that I tried to build all this while being CEO of Y Combinator actually helped me because my time is so scarce. Those minutes were insanely scarce, so I was constantly pushing myself to figure out how to automate everything."
Full-time engineers don't feel the same pressure.
3. Belief "All someone has to do is believe, right? Stop fighting, just open Claude Code and try it."
Until you push through the initial friction, you won't understand.
The Stack Overflow → ChatGPT → Claude Code Evolution
Jared nails the progression:
Stack Overflow (2008): Amazing! I can look up answers.
ChatGPT (2022): Better! I can ask questions and get answers.
Claude Code (2024): Holy shit! It just... does the work.
"Things have moved so quickly. With Stack Overflow, you'd ask questions, copy code, paste it, run it, see what happens, paste it back. With Claude Code, you don't need to do the copy-pasting. It executes and runs the code."
The mentality shift: "It doesn't actually matter if it's brittle and requires fixing because you can actually just have another agent sat there fixing it all the time."
Who Is This For?
Gary's message to developers:
"You and I are not different. We're the same. We started in the same place. I don't think of myself as in the sky yet, even though people seem to talk like I am. I'm just a person trying to do a thing. If I sit next to Boris [Churney], we have the same prompt. We have the same MacBook Pro. There's nothing that stands between you or me from drawing on millions of years of tokens to serve humanity."
If you're technical and have taste, you're the person who will benefit most.
Getting Started: Your Next Steps
-
Get Claude Pro Max ($200/month)
- It's like rent: expensive not to have it
- Token maxing requires access to best models
-
Install Claude Code
- Start with something you care about
- Real pain = real motivation
-
Build Your First Skill
- What prompt do you type twice?
- Make it a reusable skill
-
Set Up Testing
- Aim for 80-90% coverage
- Let Claude write the tests
-
Try OpenClaw + GBrain
- Your own AI assistant
- Your own data and prompts
- Your own control
-
Push Through the Friction
- First few hours are rough
- It will break
- You'll need to fix it
- That's the point
-
Token Max
- Don't optimize for cost
- Optimize for completeness
- Zap the rocks harder
The Golden Age of Open Source
"This is why we're entering the golden age of open source. I could just open any project in conductor and extract patterns from it."
When you can build in days what used to take months, when you can ship 400x more code, when you can be a time billionaire borrowing consciousness from machines—
We're all builders now.
The question isn't whether this is possible. Gary proved it is. The question is: Will you pick up the wrench?
Links & Resources
- GStack: Gary's collection of AI development skills
- GBrain: RAG system for OpenClaw
- Gary's List: AI-powered journalistic blogging platform (garyslist.org)
- Claude Code: Get started at claude.com/code
- Y Combinator Startup School: Apply for July 25-26, 2026
Update — July 4, 2026: Gary Tan revisited GStack and OpenClaw cost economics alongside two other founders on the Naval podcast — see Naval podcast roundtable: AI compute, jobs, and ASI.
"What a time to be alive."