Tan says some tiny startups reach $15 million with a few hundred AI files

Garry Tan, the president of startup accelerator Y Combinator, was explaining how he now works with artificial intelligence helpers he built for himself. He had made a personal research system called G-Brain, an automated news site called Dairyless.org, and a retrieval tool called Soda to feed the right information to an AI for each task.

Then he put his lesson in one line. “A markdown file is an employee,” Tan said. “And it’s an employee that will do the job perfectly every single time.”

By markdown file, he meant a plain text file with instructions, paired with a little code and tests. His method is to spend freely to run AI models at full strength, loading as many as 800,000 to one million tokens, the small chunks of text an AI reads, in any given request, which he calls token maxing. Do some hard task once, then “skillify” it, he said, by saving exactly how it was done so it can run again automatically on a schedule.

The first attempt is usually bad and expensive, Tan said, and takes many tries. But once the instructions are right, telling the agent what it got wrong becomes a one time bug fix. After that, the file handles the work in sales, marketing, or customer support, and any future mistake is just fixed once and stays fixed. With his engineering system, called G-Stack, he said he automated project management and checks for software tests until all he had left to do was final verification.

Tan said he was seeing companies that went from zero to $15 million in annual subscription revenue in about four months with two or three people and a few hundred skill files. He described it as recursive self improvement and a steady, step by step climb, with each solved task making the business stronger.

He added one limit. When the files pile up, he said, you need to track where facts came from, decide which fact wins when two conflict, and run automatic sweeps for accuracy. That favors founders in their 30s and 40s who have built companies before, he said, pointing to founder Peter Steinberger as someone who has been around the block and knows what to build. Take that experienced person and multiply him by 400, Tan said, and “you can outperform an entire department.”