I think that's like the defining question: like, will you have control over your own tools, or will your tools have control over you? Using OpenClaw these days is like driving a Ferrari, and it's like exhilarating. It's insane. Like, you get to do things like it figures things out you would never think a machine could figure out, and it does it so quickly. But then it's also like a Ferrari in that you better be a mechanic. Like, it's a Ferrari that will break down on the side of the road, you know, when you most need it, and you need to get out with your wrench and pop the hood and like fix it. You know, you're going to have to fix it yourself. And so this is a very exciting time in computer science and technology.
Welcome back to a special episode of the Light Cone. In this episode, we're going to talk about how Garry Tan got back to building. If you follow us on Twitter, you'll know that after a multi-year hiatus to become an investor, Garry Tan is back to being a builder. And in the last couple months, he shipped hundreds of thousands of lines of code and built popular open-source projects that have gone from nothing to more than 100,000 stars on GitHub. And he did all of this while having a very demanding job running YC full-time. A lot of people on the internet don't even think that this is possible and are somewhat like in disbelief, but it actually happened. We know because we were here to see the whole thing. And so today we're going to talk about how he did it.
Well, I'm relatively shocked myself. So I'm amazed as well. It was 13 years of not coding, and then suddenly, boom, I'm doing about 400x the amount of work that I was that year. The last time I was even sort of like two-thirds of the time writing code. Maybe to start things off, how about we go back to the project that started it all off, which was Garry's List. Oh, yeah. And just like talk about a few months ago how you powered up Claude Code and like started to get back to coding.
It was right after one of the Light Cone episodes, right?
Oh yeah, definitely. I realized that I wanted to bring together all the people who believed what I believed, particularly for California. And so I started a 501c4, and now it's a C3 and a PAC, which is sort of what a lot of political groups do. It's a very common way to bring people together. You know, everyone focuses on the money, but we're trying to bring together smart people. You know, what I learned in the years of working in San Francisco politics is that bringing together people is so powerful, and that's what a mass social movement is. And I said, 'Okay, well, why don't I just make a website where we start doing that?' And it would just start with, why don't I start writing about the issues that I'm worried about? It's like I want children in school. You know, people watching this from all around the world might find it very, very strange. Like I find it strange that it was not possible and still very, very hard for a seventh grader or eighth grader in middle school in San Francisco public schools to be able to take algebra. And that was, you know, a math education thing. Like, you know, if I didn't get to do that when I was in public schools in the East Bay of the Bay Area, there's no way I would have studied engineering at Stanford. I never would have written code. I never would have been able to do any of these things. So it was close to my heart, and I realized like, hey, it's time to write code. And I ended up building Posterous, my first YC startup from 2008.
What was Posterous for people who don't remember it?
Yeah, Posterous was dead simple blogs by email. It grew to be a top 200 website on the internet, and then Twitter ended up buying it for about $20 million. So that was sort of like my first bag, really. I actually built it again as Post Haven when Twitter, you know, bought it for the amazing people that we had hired, and they shut down the startup. It would have cost a couple million dollars to buy it back from Twitter, and at the time I had no money in the world. So the next best thing was why don't I write it again? And then in January of this year, I ended up writing it a third time. Only, you know, the first time it took about, you know, $4 million and, you know, six or seven people and about a year and a half. And then the second time it, you know, took about, I don't know, a hundred grand and two people, me and my co-founder Brett Gibson, who now runs Initialized, and maybe like three months or so. And then in this case, it took about $200, which was my Claude Code Max account, and probably five days. Full-featured blog platform, does everything you want, and then on top of that, like full RAG, full agentic retrieval, like be able to, you know, sort of go out and read all of the internet, like every tweet I've ever done, recursive crawl, deep research of any topic. The algebra thing is just one of a whole lot of different issues that we really, really care about. And to be able to go ingest the internet, you know, see all the arguments for and against, and then to craft incredibly detailed reports on the back end about what are all the quotables. Like I think people who are big followers of the Light Cone might remember one of our first episodes about agentic systems with Jake Heller, actually. So Jake created CaseText, and he described exactly what I ended up building for basically journalistic long-form articles about any, you know, sort of issue or, you know, piece of news that was happening. And so, you know, anyone can go to garrys.org work today, and, you know, we do about two or three relatively, you know, researched, all fully sourced articles about what's going on in California and San Francisco and LA and like how do we build a better government.
This is the thing I feel like people miss about Garry's List, don't fully get, is that it's like the classic thing we've been talking about here, which is like software was you build software to let people use it. So it's like you build a blogging platform and people like write blogs and maybe like they start their own Substacks eventually or they write articles. But Garry's List is both a blogging platform, but it actually does the work of a high-quality investigative journalist. It's not just something that a journalist uses to publish their articles.
Yeah. I mean, basically for the equivalent of like $5 or $10 of Opus calls. I mean, I would estimate that it does the work of like, you know, a real human being that would have to like go painstakingly through dozens of articles, read entire books about certain subjects, annotate them. I mean, going back to the CaseText example, like the thing that Jake taught me was that you need to think about what a human would do with the context given. Like, what would it retrieve? Like, does it go to the library? What kind of book would it look for? What does it search on for search, you know, on the web? I mean, the great thing now is like you don't have to just do that. Like you can get Perplexity's API and you can do deep research there. You have X's API, you can do deep research there. You know, Grok's API, if you need to like do research on X using the Grok API is actually very, very good, and you can just grab all of the context. This is sort of going back to the philosophy of 'boil the ocean,' which is one of my essays. It's like particularly when building agentic software now, you don't have to settle for what we did when we were humans writing the code, like, and that goes for research as well. What if you absolutely boiled the ocean? Like what is, you know, the total completionist? Like if you were a human, this would take you about a month to do this research. You can just, you know, zap the rocks harder. You know, you pay more money and you might be token maxing, but you should token max. Like basically, if there is incremental work that makes something more complete, more awesome, more, you know, in the case of this type of writing, like we want it to be more representative of reality. Like, you know, we don't just settle for one source when we can get 20 sources and we can cross-reference them. We can figure out like, well, these 13 sources say this and the seven sources disagree with that. And then, you know, you want to feed all of that context into like your core prompt, and then you can basically make a better decision than what you would like just, you know, a human being clicking on a link, reading a headline, and that's all you understand. And I think if you token max, like that's actually the coolest thing you can do now. And it's not just in, you know, generating articles. It's not, you know, it's clearly in writing code, right? I think now it's going to permeate every part of society. Like every thing that we would call knowledge work could be token maxed. And I don't think that it means that we're going to get rid of people. I think it means that people need to still supply the agency. Like I need this. Like I'm the one who's sitting here caring about algebra. Like I want kids like me who couldn't afford private school. You know, San Francisco is the one city in the world that has the highest rate of private school attendance, probably in the entire country, actually. And that's not okay. Like you shouldn't have to be rich to have a good education. And, you know, I don't know why that's controversial. And so for me, it's like this, you know, mass sort of shift in technology was happening, and then I had a need and a want and a desire, and it was a burning desire. Like it hurts me and pains me to think about 10, 12, 13-year-old kids who don't know algebra and like could have, but some bureaucrat or, you know, some virtue-signaling person in power says like, actually, I don't want that kid who wants to learn algebra to learn it.
So I think in this process of basically solving your own pain and need from the young Garry and building Garry's List, you sort of discover a lot of patterns on token maxing and this new way of building that led you to the next project, which was GStack.
Like I actually did not plan to make GStack. All I did was like 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. I typed in all the things that I found myself writing over and over again into Claude Code, and it was pretty simple stuff. It's like here's the plan review. One of the things I started doing is I really love asking Claude to make ASCII art diagrams. One of the things I discovered is sometimes Claude would just get confused and like write bugs or not be complete. But once I started saying, actually, before you start your work, make an ASCII diagram of all the data flows, all the inputs and outputs, what are the user flows, what are the error messages, and you can see this. It's like data flow, state machines, dependency graphs, processing pipelines, decision trees. Once it did that, it loaded all of the context in, and then it just did the work more completely. Like it boiled the ocean better, and it broke down into a bunch of different sections. Like here's architecture review, code quality, test. I mean, one of the things I learned building Garry's List was that when I was writing the code myself, I would always do the minimum amount of testing because it was just like not very fun. I knew I needed to have it, but I'm here to write, you know, fun new code. I, you know, did not like write to write tests. And then honestly, like I hit all the things that everyone else hits when they start vibe coding, which is like this is slop. It's not working that well. Like it works fine for the 80% case, but if any users actually touch it, it starts falling over. And then that's when I realized, oh, I can get to 100% test coverage. I've since learned that 100% is probably too much. Like hitting 80 to 90% is usually the best practice at this point. But yeah, this is basically the first version of plan-review. I know everyone knows the office hours skill, which is, you know, what people can use and I still use when I'm trying to make a brand new product or a brand new feature. It simulates what we do when we're working with a company. It's like, how do you know that people want this? You know, who's it for? What does it do? And what's the impact, right? But this is like the proto skill. Like this is I didn't even know skills existed. And I posted this and it went viral. Like, you know, 200,000 people saw that. And then I made another version of it that was a much more expansive version. I called it the mega plan. And then I ended up renaming it to the CEO plan. We've probably talked about meta-prompting before. I used meta-prompting here. I took the other review plan that we had, and then I said, 'Okay, well, let's do a version of this, but like imagine Brian Chesky sitting with you, right?' Like Brian Chesky has this great line about what is a 10-star experience. So, and you know, the point of it is everyone thinks about hotels in terms of like three, this is a two, three-star experience, this is a four-star experience, and he like goes, you know, through the list like five stars. It's like everyone, you know, yeah, cool. Like, but he's like, 'What's a six-star and what's a seven-star and what's an eight-star?' And like he goes all through that entire list. And that's one of my favorite like product and design exercises to go through like as a mental exercise. And then the cool thing is like you can do that every single time now. And so that's what this is. You know, this prompt basically tries to figure out what is the platonic ideal of what this is. These are sort of like the two things that are pretty awesome. One is what is the 10x check? What is more ambitious and delivers 10x more value for only 2x the effort, right? And so for whatever reason, coming out of latent space helps the model like really visualize. Like so I'm plan CEO skill I actually really enjoy because I'm an ADHD CEO and I love potential, like pure potential. And so this is like the one like I can't believe this is just literally two little sentences, but like this unlocks an incredible amount. And so that's how GStack started, actually. Not as, you know, I didn't want it to be anything other than like, well, I need to make some skills, and I had heard that people were making like skill repos. But then the third thing I did was I started using these two skills so much that my Conductor instance was getting very backed up. So this is how I use Conductor. This is actually my real setup, like...
So this is your like daily workflow. This is how you've been shipping hundreds of thousands of lines of code a month. It's all, it's all in here.
Yeah, that's right. So I dropped like 13 PRs in the last 48 hours, and then, you know, I just queue them up. Like anytime I come up with a new idea, I come in and here it is. You know, I love using the CEO skill. I loved using the skill to like really make it super well tested. I did that all in plan mode. And then I'd click approve here, and then, you know, Claude would go and do all the stuff. And then I did that so much that I ended up having like 15 different features that were all queued up waiting for me to manually test it. Like it passed it, you know, it passed end-to-end testing, it passed integration, it passed unit tests. But like at the end of the day, I still need to, you know, for Garry's List, it's like pop open the Rails server and like, you know, load that user and like make it into that configuration for that particular user and like manually just make sure it works. And I got sick of doing that, and I was trying to use Claude in Chrome MCP, and it was very, very slow, two to three seconds for every turn.
I was like, this is not usable for QA, but I had heard that Microsoft had released Playwright, which is sort of an alternative testing framework. In retrospect, it's like actually there was like agent harness and like all these other like tools that I could have used. But the upside and downside of Claude Code is it's so easy to just start something that I just popped open. Like I literally went in here, and this is probably what I did. It's like I'm so sick of using Claude in Chrome MCP. It's too slow. Let's go ahead and wrap Microsoft's Playwright. Can we do that? And then I just pressed enter. And then, you know, one of the things that emerged with GStack is that like this is how I create new features. Now, of course, you know, what it's going to do now is like, 'Hey, dude, you already did that,' which is hilarious. You know, I have bug fixes right next to giant features. And then the way GStack works, there's a CEO, there's a designer, there's actually a developer experience person in there. There's a number of design tools, and then PlanG is the last one. And then I actually usually run SLCEX. And I recently added a slash-claude in Codex.
So one of the cool things that I actually learned from YC alums, I came to an event and brain totally frazzled, but, you know, went to one of our batch events, and we were just shooting the shit about what was going on with Claude Code versus Codex. And at the time I was a total Claude Code-only guy, and I realized, oh, a lot of people actually prefer Codex. Why is that? And I discovered that Claude Code is ideal for the ADHD CEO, but once in a while there's a, you know, Claude Code will just BS a bunch of stuff. Like Claude models are very, very good, but like they are not the smartest, it turns out. And so a lot of people, you know, explained to me that if you have a problem that's much crazier, you need the 200 IQ nearly nonverbal CTO. So you can just call in a friend, and then that's what like /codex is. It's a, you know, GStack skill that takes whatever plan your plan is, or if you're out of plan mode and you already implemented it, it'll take your repo and it'll run Codex in a command line prompt with the prompt that says find all the problems and all the bugs, and it reports it back to Claude Code, and then you and Claude Code can work through those, feed that feedback. And then I have since added, if you use Codex as your main coding agent, you can actually go and type slash-claude and have Claude come and be the CEO briefly if you want as well. The cool thing about GStack is when I run it through this program, like I always, I do, I start with office hours, CEO review, like I do design if there's UI, if I know a developer needs to use it, which is like practically all of GStack and GBrain stuff, I run the developer review, and then I do review, and then Codex. Once that plan is done, I've worked through all of the issues. The GStack relies very heavily on ask user question, so because, you know, and that's that to me is like really important. That's where the human, you know, vibe coder, operator, agentic engineer needs to supply their understanding of what's going on, what are we building. There's not really a substitute to that. 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 like that. You know, it's a controversial take, I think, but 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. And so, you know, basically QA is a good example. And, you know, I mean, that's hilarious. Coming back to the demo, it's like I type something into the modern version of GStack, and it's like, dude, what are you doing? Like, we already built that. We have Browse. Browse is a long-lived HP daemon with 70 commands as a CLI. And then QA is just Browse. But in the prompt for QA, it says look in your context. What did we do on this branch? If there's UI or any mutation of data, go and use the browser to test that thing, which is cool. It's like having a black-box browser. It blew my mind when it first worked. It's like mini-AGI is already here. You know, I realize this is not true AGI. True AGI would be like I'm not even here. And actually that's fine in this respect. Like as a builder, you know, selfishly, I hope that we never have to stop. I hope that the machines never figure it out because that would be really cool. Like then, you know, humans are really important, and like engineers who know how to do this, who have taste and design and product feedback and, you know, the real customer in mind, like we're going to be like we basically have wings for as long as we do.
YC Startup School is back. We're hand-selecting the most promising builders in the world and flying them out to San Francisco for July 25th and 26th to discuss the cutting edge of tech. Apply now for a spot. Okay, back to the video. I think you crystallized a lot of this thinking in this post on X about thin harnesses and fat skills.
which actually encompasses all of this philosophy on how to token max.
Yeah. I mean, some of it came out of being trolled on the internet relentlessly about markdown, and like I, you know, I'm just like peddling a markdown instead of markdown, and it's like, you know, I guess my lived experience at this point is that markdown is actually code. It's just like this compiled in a different way, but like you can get the computer to do really astonishing things. Like I mean even this, it's like...
Could we have imagined that I would be talking to something that has replaced Visual Studio? Like I don't use Visual Studio at all. Like there's no reason to, like when I can talk to my agent and my agent can do this, right?
The article actually, the name actually came from our partner Pete Kumin. We have had to build an internal agent, and, you know, we call that the harness over and over again. And then at some point using Claude Code all day, we realized like, you know, why should we rewrite a version of that over and over again? Like, you know, we should just use the things that are really awesome as, you know, harnesses. Like a harness is the core loop that takes the user input, gives it to the LLM, runs what the LLM does, like it can do tool calls and things like that. I mean, why would we build that? Like what we should be spending all our time doing is thinking about what markdown should there be. And the way to think about markdown is if you were an event planner and throwing a wedding and you were trying to write down a checklist of how to throw a wedding again, like what would you write in plain English to teach the next person who had to do it what to do? All of that should be in the markdown. Whereas all the things that should, you know, be deterministic, like I mean, or is a real action, like a wedding planner might have to call like 20 venues, right? But you wouldn't use markdown for that. Like you would make a, you know, a call to Twilio, for instance, right? There's like a, you sort of 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. It actually, you know, code literally doesn't understand what you want or who you are. It is like, you know, executing deterministic zeros and ones in a Turing-complete loop, right? Like it doesn't know. But then now we have LLMs that have latent space, and they know who you are, and it knows what your motivations are, and it can handle generic cases. And then, you know, a lot of the magic right now as an engineer is like figuring out, okay, how much of it is over here in LLM land and how much of it is over there in code land. And then, you know, if you combine that with the other thing I learned, which is like get to 80 to 90% tests, like if it's not tested and you're just throwing users in there, like it's slop, you know, 10x worse than like human-written code because like you just have no idea what's going to happen. And so that's like one of the things that people have to do. It's like, all right, not only do you need to figure out what's going on in latent space and deterministic space, you also have to make sure that like it's, you know, unit individually tested and then the integration is tested. And then going back to boil the ocean, like the machine doesn't care, it'll just do it. It's amazing. Like just zap the rocks more, and you can get to 90% test coverage, and then you can have a system that, you know, is not quite perfect. Like, you know, OpenClaw right now, there are lots of like failure cases, but it's 95% there. You know, it's, I feel like using OpenClaw these days is like driving a Ferrari, and it's like exhilarating. It's insane. Like you get to do things like it figures things out you would never think a machine could figure out, and it does it so quickly. But then it's also like a Ferrari in that you better be a mechanic. Like it's a Ferrari that will break down on the side of the road, you know, when you most need it, and you need to get out with your wrench and pop the hood and like fix it. You know, you're going to have to fix it yourself. And so this is a very exciting time in computer science and technology because it's like this is Homebrew Computer Club. You know, the moment when the Apple I came out, like the Apple I created by Steve Jobs and Steve Wozniak was a breadboard inside like literally a wooden case hammered together with like nails and duct tape, you know, and if you wanted a personal computer, that's what you had to do. And that's where we're at right now. Like you have relatively, you know, smart technical and, you know, people who had to study computer science have to spend like two or three hours and like maybe like $500 or $1,000 in both tokens and cloud to actually get something like that running. But like once you get it, it's like we're sort of in the kit car Ferrari phase. It's like then you can drive and you can go anywhere, and you know, you want to shout to the hills like, 'Hey, I got a Ferrari.' Even the part about fixing yourself, I feel people, it's just like one of those things until you've like pushed through, you just don't quite get.
If I really zoom out, it's almost like things have moved so quickly. Like if you think way back, just having Stack Overflow as a website that you could consult when you got stuck on a programming problem felt like amazing. And then it's like ChatGPT launches, like, oh, now I've got this like interactive thing that's way better than Stack Overflow. But you're still sort of doing the same thing. And you're like asking questions and you're copy and pasting code and you're running the code and seeing what happens and copy and pasting it back. And then you sort of with Claude Code, you sort of push through and you realize you don't need to do the copy and pasting anymore. It just like actually like executes and runs the code. And even with OpenClaw, I found out when I set it up, yeah, it's annoying because it can like effectively brick itself and it does a bunch of annoying things. But if you actually have like Claude Code, like it'll fix it.
Yeah. I just have Claude Code running, it will just like fix it, and it's clearly not the way things will be long term. But there's this like mentality shift of it doesn't actually matter if it's brittle and requires fixing because you can actually just have another agent like sat there like fixing it all the time.
Yeah, I feel like this evolution, I was like completely Claude Code pill and still am, but like probably only like 50% or 60% of my time like building product or agentic engineering is in Claude Code now at some point basically.
Almost half of it is through OpenClaw now.
Yeah. Which is very interesting. I mean, then again, I'm also spending a lot, most of my time working on GBrain itself. So GBrain came about because I met you, obviously, we had Peter on the show. And then I finally got around to it. It was like one weekend I said I got to check this out, like what's going on with OpenClaw. Let's get it going. And this was about the time Karpathy wrote his ex-post about knowledge LLM wikis. And so I was like, okay, well, I have a repo full of markdown. All my, you know, I should put all of my context into that markdown. And then at some point I realized, oh shoot, it's just using Grep. And Grep is not that good. Like it's, you know, wasting context. It's loading a lot more into context than it needs to. And then I sort of fell into a rabbit hole. I just went into Conductor, clicked quick start, and then I had GStack built into Conductor already. And, you know, basically this was how I started. I, you know, it was actually much more interesting than that. So I didn't start off from nothing. One of the things I've learned as you write like a larger and larger corpus of code is like you have it loaded in your brain. You're like, oh, well, in order to build an agentic newsroom for Garry's List, I actually had to learn about vector embedding and hybrid RRF and chunking. Like when you're in there trying to make it work, you're just like very applied. It's like I have an output that I want. I want the article to look like this. It needs to be of this quality. It needs to have these citations. Like you start building up your, you know, your tests and integration tests, and like you end up with like a product that's like battle-tested from like the output that you want. And so I sort of put two and two together. And I, you know, and this is something that, you know, anyone can do actually. It's like this. This is why I think we're entering the golden age of open source. I could just open, you know, this project in Conductor, and then the first thing I write is like, you know, go look at, you know, tilda/garrys-list, like look at how we do chunking, embedding, you know, hybrid RRF, RAG, like all of this, and then just like extract it. And then I want to use Postgres with pgvector, and like I want a, you know, full RAG system for my OpenClaw. And then sort of like one thing led to another. It's like then I have, you know, 10 windows in GBrain, and I'm just like at it. What's cool about OpenClaw, I mean, maybe this is a good example. This is actually my OpenClaw. I did go ahead and ask it, um, how, you know, how did I actually get into it? January 23rd.
I had a tweet that was like, Claude Code this week has awakened my 25-year-old self, the one that chugged Red Bulls and stayed up till dawn coding. We're so back.
