Coding again, after thirteen years.
For most of the last decade, Gary Tan was an investor. He raised funds, ran initialized, then took the top job at Y Combinator. Then, quietly, sometime after a Lightcone taping in early 2026, he opened a terminal. What followed was the kind of run that engineers half his age would consider obscene: hundreds of thousands of lines shipped, an open-source project past 100,000 stars, and a small empire of agentic side projects — all while running YC full time.
"I'm relatively shocked myself," he says, on this episode of Lightcone. "It was thirteen years of not coding, and then suddenly, boom — I'm doing about 400× the work I was the year I last sat in front of a code editor."
The internet, predictably, did not believe him. The Lightcone hosts — Diana Hu, Jared Friedman, and Harj Taggar — did. They watched it happen.
How a school-board fight built a startup.
The first project, Gary's List, did not start as a software pitch. It started as a grievance. Tan grew up in the East Bay, took algebra in seventh grade, and parlayed that into engineering at Stanford, a company called Posterous, and a $20 million exit to Twitter. Then he learned that today's San Francisco eighth-graders mostly can't take algebra at all — and that fixing it would require a movement.
So he did what he always does: he built the website first.
Posterous took $4 million, six people, and eighteen months. Post Haven, the second time around, took $100,000, two people, and three months. Gary's List — full blogging platform, full retrieval-augmented research engine, full "agentic newsroom" — took five days and a $200 Claude Code Max subscription.
For the equivalent of five or ten dollars of Opus calls, it does the work of a real human being who would have to painstakingly read entire books on a subject, annotate them, and cross-reference twenty sources. — Gary Tan
The trick, Tan says, came from an old conversation with Jake Heller of Casetext: think about what a human would do with the context, then have the agent do all of it. Don't pick one source — pick twenty. Don't skim — boil the ocean. "If there is incremental work that makes something more complete, more representative of reality — you should token-max it."
The rise of tokenmaxxing.
The verb arrived almost as a joke. Then it stuck.
To tokenmax is to spend tokens like a founder spends rent in San Francisco — generously, on the understanding that the cheap option is the actively expensive one. You don't ask the model to do the minimum; you ask it to do everything a meticulous human would, and you let it run.
Tan's analogy: rent. "Early in a YC batch, founders go, this apartment is thousands a month, should I really pay it? And the answer is: no, you should pay more — to be in Dogpatch, to be where the serendipity is. Token-maxing is the same. Sure, economize on the desk. Don't economize on the model."
How GStack happened by accident.
GStack — the open-source skill library Tan didn't mean to publish — emerged because he kept retyping the same prompts into Claude Code. So he opened Apple Notes, pasted in the recurring scaffolding, and saved them as skills.
The first one was Plan Review. Before any feature, Claude was asked to draw an ASCII diagram of every data flow, every state machine, every error path. Once Claude saw the architecture in latent space, the work came back complete instead of half-baked.
Then came Plan CEO, born from a metaprompt. Tan fed in Brian Chesky's 11-star experience exercise and asked the model to imagine the Airbnb founder reviewing the spec. "What is the platonic ideal of this feature? What would deliver 10× more value for 2× the effort?" Two sentences of prompt; an enormous unlock.
The skills compounded. Then the agents stacked. By March he was running fifteen Conductor windows in parallel, each working a different ticket from his queue.
Thin harness, fat skills.
Halfway through the episode, the conversation turns from tactics to taxonomy. Tan and partner Pete Koomen have been writing about a split that Tan now considers the central question of agentic engineering: what belongs in code, and what belongs in markdown?
The harness — the loop that takes user input, hands it to the model, runs the tools — is generic. Stop rewriting it. Use a good one. Make it thin.
The skills — the markdown — are where everything specific to your problem lives. They're not configuration. They're plain-English instructions to the next person who has to do the job, except the next person is an LLM with latent space and a sense of who you are.
Code is brittle. It doesn't understand who you are or what you want. The magic is figuring out how much of your system belongs in markdown — and how much in deterministic zeros and ones. — Gary Tan
Most agentic-engineering failures, in Tan's reading, come from putting markdown logic into code, or code logic into markdown. A wedding planner writes a checklist in English; she doesn't write English to call twenty venues. She uses Twilio for that.
It's a Ferrari. You'd better be a mechanic.
The episode opens — and closes — on the same metaphor. Using OpenClaw and Claude Code today, Tan says, is like driving a Ferrari. The car will figure out things you would not have thought a machine could figure out, and it will do them faster than you can read.
The car will also break down on the side of the road, at the moment of maximum inconvenience, and you will need to pop the hood. There is no AAA.
This is, he insists, the Homebrew Computer Club moment. The Apple I was a breadboard in a wooden case held together with duct tape. If you wanted a personal computer in 1976, you had to be the kind of person who could solder. We are there again — except the breadboard is a $500 token spend, a stack of markdown, and the taste to know which agent is telling you the truth.
The future is personal AI.
The episode's strongest claim is also its most political. By this time next year, Tan predicts, every person on the planet will have their own personal AI — not a hosted product feed, not a corporate algorithm, but their own prompts, their own data, their own integrations, running in their own loop.
Will you have control over your own tools, or will your tools have control over you? That is the defining question. — Gary Tan
The alternative, he says, is the Facebook feed: an algorithm written by someone else, optimising for someone else's business model, and you don't know who or what.
The cost of admission is real — a few hundred dollars in tokens, a willingness to write your own markdown, the mental shift from asking the assistant to directing the agent. But the payoff, Tan argues, is the closest thing yet to time travel: a way for one person who cares about a problem to throw a million machine-years at it.
Buying back time with tokens.
Asked whether running YC helped him ship more code — by leaving him no time to manually click around — Tan laughs and says: probably, yes.
"My kids are time billionaires," he says. "I'm not. But if you can token-max, you can buy millions of years of machine consciousness. I'm not a time billionaire in my own life. I can be one in someone else's."
Diana Hu calls it a beautiful quote. Jared Friedman says it should be on X immediately. The episode ends.
- Tokenmaxxing
- Deliberately overspending on model tokens so the agent does the complete, exhaustive version of a task — the version a meticulous human would have done if they'd had a month.
- Thin harness, fat skills
- Use a generic agent loop. Put the specifics — your taste, your invariants, your edge cases — in markdown skills, not in code.
- GStack
- Tan's open-source bundle of Claude Code skills (CEO review, plan review, codex bug-hunt, Browse QA). Started as personal scratch, now a public repo.
- Personal AI
- An agent running on your hardware, your prompts, your data — the post-Facebook-feed version of "your" software.