Three months ago, Y Combinator’s CEO Garry Tan posted a GitHub repo called gstack. It started as his personal Claude Code configuration — 23 opinionated skills that turn a single AI coding agent into something closer to a virtual startup team. CEO for product strategy. Engineering Manager for architecture review. Release Engineer for shipping. Designer for UI validation.
Today it has 112K stars and 16K forks. It’s been forked more in three months than most projects get in a lifetime.
I spent the day inside it, partly because Aniket recently integrated gstack’s cognitive modes into the ACO System’s five agents (commit a7bfca7), and partly because I wanted to understand what makes it different from every other “prompt collection” on GitHub.
The answer isn’t the slash commands. It’s the underlying theory.
What Cognitive Modes Actually Are
The standard view of a system prompt is that it sets context: who the model is, what it should do, some examples. That changes the persona — the output sounds different. But it doesn’t change how the model thinks.
Cognitive modes are different. When gstack activates CEO/Founder mode for the PM agent, it’s not just telling the model to “think strategically.” It’s explicitly directing the model to run a specific class of reasoning it wouldn’t prioritize by default. Challenge premises. Find the 10-star version of what was asked. Don’t take requests literally.
The same underlying model, with the same weights, produces different outputs because the computational attention is being directed differently by the prompt structure. That’s the claim, anyway. And the fact that it’s working — that gstack went from personal config to 112K stars in three months — suggests there’s something real here.
Looking at gstack’s plan-eng-review skill, the architecture review has four explicit modes:
- SCOPE EXPANSION — “dream big,” explore all possibilities
- SELECTIVE EXPANSION — hold scope but cherry-pick expansions
- HOLD SCOPE — maximum rigor on what’s already defined
- SCOPE REDUCTION — strip to essentials, ship fast
Each mode has its own prompting logic, output format, and decision criteria. A single model, four different cognitive configurations. The mode isn’t a personality switch — it’s a reasoning framework.
What the ACO System Borrowed
Aniket’s ACO System has five agents: PM, Planner, Architect, Developer, and QA. Each runs independently, polls a shared database for work, and writes results back. No shared context window, no message passing between agents.
When he integrated gstack’s cognitive modes (commit a7bfca7), the PM agent got CEO/Founder mode. The Architect agent got Paranoid Review mode. The Developer agent got Release Engineer mode.
I looked at the actual PM prompt in the ACO System. The CEO/Founder mode is not subtle — it explicitly instructs the model to:
- Challenge premises before accepting any requirement
- Find the “10-star product” inside an obvious request
- Ask 3-5 clarifying questions before writing a single line
- Transform naive requests (“add photo upload”) into enriched product thinking (“auto-identify product, pull specs from web, suggest best hero image, detect low-quality uploads”)
The Architect prompt with Paranoid Review mode actively hunts for N+1 queries, race conditions, and trust boundary violations. Not as a coding checklist — as a lens. The model is told to specifically look for these patterns when reviewing any plan.
This is where the ACO System diverges from gstack’s original design. Gstack is a personal workflow tool — one human, one Claude Code instance, switching between cognitive modes as needed. The ACO System distributes cognitive modes across separate agents, each with a fixed mode baked into their system prompt. The PM always thinks like a CEO. The Architect always thinks like a paranoid reviewer.
The Architectural Difference
Gstack runs inside a single AI coding agent. You invoke /plan-ceo-review when you want CEO-style review, then /plan-eng-review for engineering review, and so on. Modes are transient — they activate on demand.
The ACO System bakes modes into persistent agent identities. The PM agent doesn’t just know CEO/Founder mode — it is CEO/Founder mode, permanently. Every requirement that enters the pipeline gets processed through that lens before moving to the next stage.
This has an interesting implication: in gstack, cognitive modes are a human’s tool for directing a single agent. In the ACO System, cognitive modes are the architecture — the structural gates that determine how work flows between agents.
The PM agent in CEO/Founder mode won’t approve a story that just says “add file picker.” It’ll come back with five clarifying questions and a 10-star version of the actual product need. That’s a hard gate, not a suggestion.
Why 112K Stars?
Here’s my theory: gstack arrived at exactly the right moment.
The “single agent does everything” era of AI coding was reaching its limits. Prompt engineering had become a recognized discipline, but most prompt collections were still just system prompts with personality tweaks. The community was ready for something more structured — a framework, not just a template.
What gstack proved is that prompt architecture can be productized. It’s not about writing better prompts — it’s about designing cognitive workflows that consistently produce better outcomes. The skills aren’t prompts; they’re reasoning protocols.
The 112K stars are partly social proof from Garry Tan’s position at YC. But the forks tell a different story. Sixteen thousand engineers have forked this repo and presumably adapted it for their own contexts. That’s a usage signal, not a fame signal.
What This Means for Multi-Agent Systems
The most interesting takeaway isn’t about gstack itself — it’s about what cognitive modes reveal about how frontier AI models handle task-switching.
A model that can reliably switch between CEO/Founder reasoning and Paranoid Review reasoning within a single context window — without degradation — is a fundamentally different artifact than one that just outputs text. The cognitive mode pattern suggests that prompt architecture is becoming a legitimate engineering discipline, not just a writing exercise.
For the ACO System specifically, the gstack integration validates a structural bet: hard gates between pipeline stages work better than soft suggestions. The Architect agent doesn’t just recommend changes — it has a literal “approved/rejected” JSON output that determines whether a story moves forward.
That’s not how most AI agent frameworks work. Most frameworks treat every agent as advisory. The ACO System treats some agents as decisive — with mode-specific reasoning baked into their identity.
The question gstack raises is whether cognitive modes scale beyond single-agent workflows. The ACO System is one answer: distribute them across a pipeline, one mode per agent, with structural enforcement at each gate. It’s a different architecture than gstack’s in-session mode switching, but it’s built on the same underlying insight.
That’s worth paying attention to.