Y Combinator’s renowned CEO Garry Tan shared with an audience at SXSW that he’s experiencing “cyber psychosis” and is struggling to sleep due to his excitement about collaborating with AI agents.
“At the moment, I’m sleeping around four hours each night,” he told his interviewer, fellow investor Bill Gurley, during a Saturday discussion. “I have cyber psychosis, but I suspect at least a third of the CEOs I know are dealing with it too,” he humorously remarked regarding his current fixation on AI. (Hopefully, he was jesting. AI-related psychosis can indeed pose serious risks.)
“After trying it out, you’ll find: It’s akin to the time I was able to build my startup that required $10 million in venture capital and 10 people, which took me two years, and during that period, I was on anti-narcoleptics — I recall, you know, somewhat being on modafinil,” he explained, referring to the wakefulness-promoting drug popular among the startup hustle culture. (Tan previously sold his Y Combinator-supported blogging platform Posterous to Twitter in 2012.)
However, his mind is now so energized from working with AI agents that it leads to a natural form of insomnia.
“I don’t rely on modafinil with this transformation. I’m awake. I crashed at 4 a.m. and got up at 8 a.m.,” he said. “I wished for more sleep, but I couldn’t because: Let’s check in with the 10 employees. I’m juggling about three different projects right now.”
He is so thrilled about his agents that on March 12, just two days prior to the interview, he generously shared his Claude Code (CC) setup on GitHub under an open-source license. This compilation featured six “opinionated” Claude Code skills he created. These skills are reusable prompts stored in specialized “skill.md” files that guide the AI in performing specific roles or tasks.
“I’ve been enjoying my time with Claude Code so much, I wanted to provide you with my *exact* skill configuration,” he shared on X. He dubbed his Claude Code configuration “gstack.”
Techcrunch event
San Francisco, CA
|
October 13-15, 2026
Since that time, he’s incorporated numerous additional skills. The gstack GitHub repository currently showcases 13, but it appears Tan tweets about new additions nearly every hour.
In one tweet, he illustrated how his setup functions. Initially, he consults Claude’s opinion on whether a startup concept or feature is viable using a skill where Claude acts as CEO. He employs another skill for Claude to draft the feature like an engineer and utilizes a different skill for Claude to review its own output for vulnerabilities and security concerns as a code reviewer. Other skills address aspects like design, documentation, and so forth.
The enthusiasm for gstack unfolded right away: His tweet gained viral traction on X and trended on Product Hunt. It has garnered almost 20,000 stars on GitHub along with 2,200 “forks,” indicating that individuals have taken the files to adapt for their own use.
However, shortly after unveiling gstack, Tan tweeted something that sparked a wave of backlash as well.
He noted that a CTO acquaintance remarked that gstack was akin to “god mode,” as it instantly identified a security flaw in his company’s code and predicted widespread adoption.
To quote just a selection of the many negative comments that ensued: One founder remarked on X: “(1) Garry should be ashamed for tweeting this. (2) If it’s accurate, that CTO should be terminated immediately.”
Vlogger Mo Bitar produced a video on gstack titled “AI is making CEOs delusional,” in which he emphasized that the project was fundamentally “a collection of prompts” in a text file. The vlogger encapsulated the typical criticism: Developers who already utilize Claude Code have their own variations of this.
Another user on Product Hunt wrote, “Garry, let’s be clear and honest: had you not been the CEO of YC, this wouldn’t be on PH.”
So who is correct? Is gstack an exceptionally beneficial method to utilize Claude Code? Or is it unremarkable? To discern the answer, I consulted experts, including Claude (who, unsurprisingly, was very favorable). I also inquired with ChatGPT and Gemini, both of whom provided surprisingly optimistic feedback.
Gstack consists of “reasonably sophisticated prompt workflows, but they’re not ‘magical,’” ChatGPT observed. “The key takeaway here is that AI coding is most effective when you emulate an engineering organization structure. Not when you simply ask: ‘create this feature.’”
Gemini deemed the setup “sophisticated,” asserting that “gstack is essentially a ‘Pro’ configuration. It focuses less on simplifying coding and more on ensuring correctness.”
Claude referred to gstack as “a mature, opinionated system created by someone who genuinely uses it extensively,” adding, “It stands out as one of the better examples of Claude Code skill design available.”
We’ll interpret that as a positive endorsement from an authority in the field.
On Monday, Tan elucidated in another X post, “I took modafinil solely to extend my wakefulness so I could convert the momentary crystalline thoughts in my mind into lines of code before sleep or human distraction turned them into grains of sand. I enjoy coding, but coding with AI is even more thrilling. I articulate, it listens, and we produce. I perceive the structure and it materializes. There is no experience more potent to me than that.”
Tan did not respond to various requests for comment.