gstack Is a Costume Trunk, Not a Stack
gstack Is a Costume Trunk, Not a Stack
Garry Tan — CEO of Y Combinator — shipped an open-source Claude Code framework called gstack. It packages his personal development philosophy into slash commands: /office-hours challenges your idea like a skeptical product strategist, /plan-ceo-review runs "Brian Chesky mode" and demands the 10-star version of your ticket, and a roster of persona skills play CEO, Eng Manager, Staff Engineer, and Release Engineer. It hit 89,700 GitHub stars in under two months.
Let's be blunt about what it actually is: a costume trunk. A very well-distributed box of hats for one model to wear, one at a time.
We've written before that prompt frameworks solve the wrong layer. gstack is the most successful prompt framework ever shipped, which makes it the clearest demonstration of the problem.
The Org Chart Fallacy
gstack's pitch is that a solo developer gets "a full startup team" — each command activates a specialized cognitive mode mapped to a role in an engineering org.
Here's why real orgs work, though. It isn't the job titles. A staff engineer catches what the product lead misses because they hold different information (they read the code, not the deck), face different incentives (they're on call for it), and hold actual blocking power (their review gate is enforced by the repo, not by politeness).
A persona prompt has none of those three properties. "Eng Manager review" and "CEO review" in gstack are the same model, in the same session, with the same context window, wearing a different hat. It cannot know something the "CEO" doesn't know — it just saw the same tokens. It cannot want something different. And it cannot block anything; it can only emit more text into the same conversation that produced the mistake.
That's not a review. That's asking the author to squint at their own work in a mirror labeled "reviewer." Sometimes the squint helps — role prompts measurably shift output distribution, and we won't pretend otherwise. But the delta between "one model squinting" and "a second process with its own state, tools, and a gate that can fail your build" is the entire difference between vibes and engineering.
We run role separation as architecture, not as costume: reviewer agents are separate processes with their own context assembly, security review runs a distinct checklist against a distinct threat model, and the gates are hooks that block the commit — not a paragraph suggesting you reconsider. The role only means something when the system enforces it.
Its Own Users Are the Benchmark
The most honest data on gstack comes from the people who actually adopted it. One detailed month-long field review found:
- 6 of 35 commands survived into stable daily use. The rest "rang hollow" or needed heavy adaptation.
- Roughly a quarter of the commands become dead weight unless you adopt the whole ecosystem —
/open-gstack-browserand the GBrain memory layer assume wholesale buy-in. - The steep on-ramp wasn't Claude Code; it was learning which of dozens of overlapping commands to run in which order.
- The reviewer's endpoint was a fork keeping the handful of functional commands and deleting the role-play planning layer entirely.
When your power users' steady state is a fork containing 17% of your product, the product is the 17%. Everything else is packaging.
And look at which 17% survives: /qa, /review, /ship, /investigate, /learn. Not one persona in the list.
The Tell: The Good Part Isn't a Prompt
Ask anyone who likes gstack what the best part is and you get the same answer: /qa. It drives a real Chromium browser, takes screenshots, verifies the UI actually renders, fixes bugs with atomic commits, and generates regression tests.
Notice what /qa is: a tool with an execution loop and a check that can fail. It touches reality. The browser either renders the page or it doesn't; the screenshot either shows the button or it doesn't. It's the one component in the box that doesn't work by pretending to be a person — and it's the one component everyone keeps.
gstack accidentally ran the experiment and published the result. Personas get forked out; tools survive. The framework's own adoption data refutes the framework's thesis.
This matches everything we've measured on our side. After reading every prompt ever typed into this system, the pattern was unambiguous: the median prompt is 56 characters, because the leverage was never in the prompt. It's in what surrounds the model — context assembly, memory it can query, tools that touch reality, and gates that can say no. Prompts are the output of that system, not the system.
89.7K Stars Is a Distribution Number
Two months to 89,700 stars would be a historic product signal — if stars measured product. They measure reach, and Garry Tan has one of the largest founder audiences on earth attached to the most powerful brand in startups.
This is worth saying plainly because it's the actual business model on display. Y Combinator's core product has always been distribution and a stamp: the batch, Demo Day, the network, the implied endorsement. That machine is genuinely powerful. gstack is that machine pointed at a repo — office hours as a slash command, founder-mode-as-a-service, the accelerator's pattern language shipped as markdown. The stars tell you the machine works. They tell you nothing about whether persona prompts improve software, and the field reports above suggest mostly not.
There's a quiet irony in the mantra here. "Make something people want" — the YC koan — has a control group now: people starred what someone famous made, then quietly forked out the part they actually wanted. The market spoke twice.
What gstack Gets Right
Credit where due, because the diagnosis inside gstack is correct even though the treatment is markdown:
Different phases of development need different modes of thinking. True. We build this as explicit processing strategies — but the mode selection changes what the system does (context depth, effort budget, which agents and gates engage), not just which hat the text wears.
Opinionated structure beats a blank page. Also true. A newcomer to Claude Code gets real value from someone else's guardrails on day one, and gstack's Think → Plan → Build → Review → Test → Ship loop is a sane default shape.
Verification belongs in the loop. /qa is genuinely good, precisely because it's the least gstack-like thing in gstack.
If gstack is your first week with an agent, install it, use /qa, and learn the loop. Just watch what you're still running in month two — and notice that the answer, empirically, is the tools.
The Takeaway
The costume trunk is fun. The hats are well-made. Some of them are modeled on genuinely great operators.
But a solo developer plus six hats is still a solo developer. What actually multiplies you is boring by comparison: separate processes with separate state, memory you can query instead of re-explain, tools that touch reality, and review gates wired into the machinery so "no" is enforced rather than performed.
An org chart you can /slash into a chat window is content. The system that makes the roles real is architecture. Build — or buy — the second thing.
Sources: YC Startup Library — Inside Garry Tan's AI coding setup · The Claude Codex — gstack: a critical review after one month of use · Augment Code — gstack hits 89.7K stars · MindStudio — What is gstack?