On March 12, 2026, Y Combinator CEO Garry Tan dropped gstack on GitHub under an MIT license and immediately dragged half of developer Twitter into an argument about whether a stack of Claude Code skill files counts as infrastructure or cosplay. The pack is opinionated. It is a set of skill.md prompts that simulate roles inside an engineering organization so Claude Code can behave less like a single chat window and more like a tiny org chart with opinions. Tan tweeted that a CTO friend called it god mode, said the friend found an XSS with it, and predicted that 90 percent of new repos would eventually use something like it. That prediction is what lit the fuse. By March 17 TechCrunch was writing the love-hate postmortem while Product Hunt traffic piled up for the simple reason that the YC CEO had shipped it.
What gstack actually is on disk
I cloned it the way I clone anything that arrives with a CEO attached: with curiosity and a slight flinch. Gstack is not a compiler, not a runtime, and not a hosted agent platform. It is a collection of Claude Code skills, meaning markdown prompt packages that tell the model how to act in named roles across a development workflow. Think reviewer, planner, shipper, the hats a small eng org wears in sequence. The point is opinionated process. Instead of asking Claude Code one vague question at a time, you load a skill that already encodes how Tan thinks that role should behave.
That design choice matters more than the branding. Plenty of people already keep private prompt libraries. Shipping them as MIT-licensed skill.md files makes the workflow forkable and easy to argue about in public. Supporters on Product Hunt said the value is exactly that opinionation. They want a biased workflow that already knows when to write a plan, when to open a diff, and when to ship. If you already run Claude Code daily, a skills pack is a product-shaped object even when the bytes are mostly English.
Tan's own framing leaned hard into leverage. The CTO friend anecdote, god mode, an XSS find, the 90 percent forecast, was not a quiet README note. It was a distribution strategy. When the head of YC says the default future of repositories runs through his prompt pack, people listen for reasons that have nothing to do with token quality. Product Hunt attention followed the title.
I am not going to pretend I reproduced the XSS find in a controlled lab. The public story is the anecdote Tan told, not a CVE write-up. What I can say is that the claim fits a pattern I see weekly: coding agents are decent at grinding through boring security checklist work when you give them a role and a target. Whether that is god mode or a competent intern with infinite patience depends on how much you romanticize interns.
Open source licensing is doing real work here. MIT means companies can absorb the files, rewrite the roles, and never send a pull request. That is how opinionated workflows spread. It is also how they get blamed when someone pastes a skill into a regulated environment without reading it. A prompt is code-adjacent enough to fork and social enough to meme.
The backlash arrived on schedule
The pushback was as predictable as the launch. One founder line making the rounds said Tan should be embarrassed, and that if a CTO truly believed 90 percent of new repos would run on this, that CTO should be fired. Founders are exhausted by AI theater. A YC CEO predicting near-universal adoption of his own GitHub repo trips every allergy at once: status, distribution, and the fear that fundraising optics now outrank engineering taste.
Mo Bitar's video, titled along the lines of AI making CEOs delusional, treated gstack as exhibit A in a broader rant about prompt piles dressed up as platforms. His cut was blunt. A bunch of prompts is a bunch of prompts. Calling it an eng org in a folder does not make it an eng org. That critique lands for anyone who has watched a README grow wings on launch day. It lands less cleanly for people who already live inside Claude Code and care more about reusable role prompts than packaging purity.
TechCrunch's March 17 piece held both objects in the frame: the viral CEO launch and the substance of a skills pack people could actually try. I read it less as a verdict and more as a weather report. The storm was never really about markdown syntax. It was about who gets to declare the default toolchain for the agent era, and whether YC's megaphone is a substitute for adoption curves.
There is a fair version of the criticism. Predicting 90 percent market share for a two-day-old prompt pack is marketing, not measurement. Founders who ship quietly for years hear that number and taste blood. At the same time, dismissing every opinionated workflow tool as delusion ignores how software culture moves. Lint rules, issue templates, and Cursor rules all started as someone else's taste. Some became infrastructure. Most died. Gstack has not earned either fate yet.
I also keep the XSS anecdote in the maybe column. Security finds make great launch garnish. Without a public disclosure trail, they are stories, not evidence. If Tan's CTO friend filed something real, great. If the line was vibes-based salesmanship, that is also data about how AI tools are sold in 2026. Believe the repo. Discount the prophecy.
How I am scoring it after the noise
My working test for tools like gstack is rude and simple. Does it change what a competent engineer does on Tuesday, or does it only change what people argue about on Monday? Opinionated Claude Code skills can clear that bar if they shorten the distance between task and merged change. They fail if they become costume for executives who want to look technical without reading diffs.
Supporters are not imaginary. I have watched teams get mileage from role-shaped prompts that force planning before code and review before merge. If gstack encodes a strong version of that loop, it is useful the way a good playbook is useful. You can disagree with the author and still steal the structure. That is the open source bargain. The YC logo makes more people open it.
The Product Hunt spike will fade. What remains is whether skill.md packs become a normal unit of developer tooling the way editor extensions did. Claude Code already treats skills as a first-class surface. Tan just shoved a loud, branded bundle onto that surface and dared the industry to look away. Looking away was never an option once the 90 percent line hit the timeline.
If you run a shop evaluating this, do the boring things. Read the skills. Strip the brand voice. Run them on a repo you own. Measure whether review quality or cycle time moves. Ignore the god mode language until a stranger reproduces a find you care about. And please do not fire a CTO over a tweet about a prompt pack.
As for Tan, embarrassment is the wrong frame. Overreach is the right one. CEOs on the AI tooling frontier will keep overselling the folder they shipped last night. Critics will keep calling folders delusional. Somewhere between those poles, a lot of us will quietly keep the two skills that help and delete the rest. That is not god mode. It is how programming culture metabolizes hype when the license is MIT and the install cost is a clone.
- LLMs
- Open Source
