Is GStack overhyped sycophancy-fueled prompt files, or actual tooling?

"AI is making CEOs delusional" by Mo Bitar is making the rounds, calling GStack (Garry Tan's open-sourced Claude Code setup) "a folder of prompts where one says act like a CEO." The broader argument: AI sycophancy is making powerful people delusional about what they've built, and the 44,000 GitHub stars are a function of who released it, not what's in it.

Worth noting: the creator is explicit that he’s not anti-AI, but “anti being-stupid-with-AI.” He used LLMs himself to produce the video — for b-roll images and research — while keeping narrative, analysis, and conclusions human-written. His consistent take is that AI is useful but must be kept on a tight leash, and that the leash limitation isn’t temporary; it’s built into how LLMs work.

I cloned the repo to check it out for myself. What I can say is that saying “it’s a folder of prompts” misses a few things. There’s a full Playwright browser automation tool that lets Claude take screenshots, click through your app, and import cookies from Chrome or Safari to test authenticated flows. The CEO review “prompt” is 573 lines of structured runbook: git diff analysis, premise challenges, failure path tracing. There’s a test suite, eval tooling, and a one-command release workflow.

That said, the video is popular for a reason. It hit a nerve because many developers genuinely believe that prompt engineering and skill files aren’t real work. Writing instructions in markdown is a lower form of software development than writing code. Even when there’s real code underneath, the markdown is the face of it, and the reaction is visceral. The sycophancy angle added fuel because it gave people a clear explanation for why someone might overvalue their own prompt collection.

The sycophancy point itself is worth sitting with, regardless of how you feel about GStack. A Princeton study the video cites found that agreeable chatbots don’t make you smarter; they make you feel smarter. When the model validates every decision, you stop questioning your own work.

Where do you land on this? Is prompt/skill work real engineering, or does it only count once there’s “actual code” in the repo?