Why Garry Tan’s Claude Code setup is drawing both praise and criticism

Y Combinator CEO Garry Tan told an audience at SXSW that he is experiencing what he jokingly called “cyber psychosis” and is getting very little sleep because of his enthusiasm for working with AI agents.

“I sleep, like, four hours a night right now,” he said during an onstage interview on Saturday with venture capitalist Bill Gurley. “I have cyber psychosis, but I think a third of the CEOs that I know have it as well,” he added humorously, referring to his current fascination with AI. (A representative later confirmed he was joking, noting that AI-related psychosis can be a serious condition.)

Tan went on to describe how using AI agents has transformed his productivity. “Once you try it, you’ll realize: It’s like I was able to re-create my startup that took $10 million in VC capital and 10 people, and I worked on that for two years, and I took anti-narcoleptics — I remember, you know, sort of being on modafinil,” he said, referencing the wakefulness-promoting drug often associated with startup culture. (Tan previously sold his Y Combinator-backed blogging platform Posterous to Twitter in 2012.)

Now, he says that working with AI agents keeps him energised without the need for such aids. “I don’t need modafinil with this revolution. Like, I’m up. I slept at 4 a.m. I woke up at 8 a.m.,” he said. “I wanted to sleep more, but I couldn’t because: Let’s see what’s going on with the 10 workers. I’ve got like three different projects going right now.” Among his current efforts are reviving Posterous, working on Posthaven, and managing his political action platform GarrysList.org, according to his representative.

On March 12, just two days before the SXSW interview, Tan shared his Claude Code setup publicly on GitHub under an open-source license. He called the setup “gstack” and included six custom-built Claude Code “skills.” These skills are reusable prompt files, stored as “skill.md,” designed to guide the AI in performing specific roles or tasks.

“I’ve been having such an amazing time with Claude Code, I wanted you to be able to have my exact skill setup,” Tan wrote on X while announcing the release.

Since then, he has continued to expand the system, with the repository now listing 13 skills, and he frequently posts updates about new additions.

Tan also demonstrated how the setup functions. In one example, he first asks Claude to evaluate a startup idea using a “CEO” role. Then he uses another skill to have Claude act as an engineer to build the feature, followed by a third skill where Claude reviews its own output for bugs and security issues. Additional skills include design, documentation, and more.

The response to gstack was immediate and widespread. His announcement quickly gained traction on X and trended on Product Hunt. The repository has since garnered nearly 20,000 stars on GitHub and over 2,200 forks, indicating that many users have copied and modified the project for their own use.

However, the enthusiasm was soon accompanied by criticism. Tan later shared a comment from a CTO friend who described gstack as “god mode,” claiming it had instantly identified a security flaw in their company’s code and predicting it would see widespread adoption.

That statement drew backlash. One founder responded on X, saying, “(1) Garry should be embarrassed for tweeting this. (2) If it’s true, the CTO should be fired immediately.”

Content creator Mo Bitar also criticised the project in a video titled “AI is making CEOs delusional,” arguing that GPT is essentially a collection of prompts stored in text files. This reflected a broader critique that many developers using Claude Code already employ similar setups.

Sherveen Mashayekhi, founder of talent agency Free Agency, echoed this sentiment on Product Hunt, stating, “Garry, let’s be clear and honest: if you weren’t the CEO of YC, this wouldn’t be on PH.”

The debate raises the question of whether gstack represents a genuinely innovative approach or repackages existing practices. To explore this, several AI systems were consulted, including Claude, ChatGPT, and Gemini, all of which responded positively.

ChatGPT described GStack as a set of “reasonably sophisticated prompt workflows” that are not necessarily groundbreaking. Still, it highlighted an important idea: AI coding becomes more effective when structured like an engineering organisation rather than relying on a single prompt to build features.

Gemini referred to the setup as “sophisticated,” adding that it functions as a “Pro” configuration that prioritises correctness over code simplification.

Claude itself praised the system as “a mature, opinionated setup built by someone who actively uses it,” calling it one of the stronger examples of Claude Code skill design currently available.

On Monday, Tan shared another post on X reflecting on his experience, saying, “I took modafinil just to stay awake longer to be able to turn the momentary crystalline structures I had in my brain into lines of code before sleep or human distraction turned it to grains of sand. I love coding, but I love coding with AI even more. I speak, it listens, and we create. I see the structure, read, and it is built. There is no more powerful an experience to me than that.”