Garry Tan: 'I am 90x of myself' — G Stack hits 60K GitHub stars, 30K daily users as Tan evangelizes AI-native software development

Summary

Garry Tan is building G Stack — an open-source, MIT-licensed software development framework — as a live experiment in what he calls AI-native software development. The project has hit 60,000 GitHub stars and 30,000 daily users, and Tan claims his personal output is now 90x what it was in 2013, the year he built both Post Haven and Bookface.

G Stack started from a specific frustration. Tan was building Garry's List with six or seven Conductor windows open and found himself doing manual QA between agents — the bottleneck he'd worked so hard to eliminate. Claude's Chrome MCP was too slow and bloated, so he built a thin CLI wrapper over Playwright that let agents run browser-based QA autonomously. That was version one. The framework now has 30 skills, including a plan-engineering review that targets 100% test coverage and a "plan CEO review" modeled on Brian Chesky's 10-star experience framework. A design-shotgun skill can generate 100 visual variants in a single pass. Tan is currently managing 200 open pull requests against the repo.

Garry's List as the test bed

Garry's List, the newsletter product Tan built on top of G Stack, is less a product launch than a live infrastructure experiment. Every email is personalized and sent from Tan's own address. Replies go to his inbox, where AI helps him read and route them into a personalization system. He describes one user — an engineer in Pleasanton who had never engaged in politics before — now running roughly 100 conversations a month with the agent and sending letters to elected representatives based on issues in his life. Tan frames this as exploring personalized communication at scale, something packaged newsletter tools can't replicate.

Where moats come from when software is fungible

Tan's answer to the question of competitive advantage in an era of cheap software is three-part: agency (the ability to prompt effectively), taste (the ability to evaluate outputs), and trust. Of the three, he argues trust matters most for enterprise companies. Once a customer has gone through onboarding and integrated a product into their workflow — once someone gets promoted or avoids being fired because of it — switching costs are real and durable. The volume-of-software flex, exemplified by Tim Draper's stunt of rebuilding the entire YC batch, impresses Tan technically but doesn't resolve the trust question.

He's more generous about Cluely, acknowledging the company has real revenue and a sound core idea, even while distancing himself from the troll-marketing approach.

The 90x claim

At a JPMorgan CEO conference the prior week, Tan told a room of former-technical-now-full-manager CEOs that they should start coding again. His pitch was direct: "I am 90 of myself." He's producing 36,000 lines of code per day, all machine-written, oriented entirely toward solving real problems rather than hitting a metric. His view is that the mid-transition moment — where agents still estimate timelines in human terms unless explicitly asked otherwise — is precisely when technically fluent founders and executives have the largest edge. The agent, when asked how long a task will take, defaults to "a human would need a month"; pushed further, it says fifteen to twenty minutes.

Tan is also considering opening Bookface, YC's internal social network with 16,000–17,000 alumni and roughly 40% daily active usage, to a broader audience of builders and techno-optimists — not the general public, but beyond the current alumni-only access that has already started leaking through shared logins.