7 Gears, One Founder: What Garry Tan's gstack Reveals About Building Solo With Claude Code in 2026

# 7 Gears, One Founder: What Garry Tan's gstack Reveals About Building Solo With Claude Code in 2026

Garry Tan runs Y Combinator. He also, apparently, ships production code alone on weekends using Claude Code. When someone with his vantage point publishes their personal AI stack — what he calls the gstack — solo founders should pay attention. Not because it's aspirational. Because it's operational.

This article breaks down what the gstack actually contains, why it maps so cleanly onto the solo founder problem, and what it implies for anyone trying to build a one-person startup with AI right now.

The four-role problem every solo founder faces

A solo founder isn't just a founder. They're a product manager, an engineer, a QA tester, and a customer. These roles have different rhythms, different incentive structures, and different failure modes. Trying to context-switch between them manually is where most solo founders bleed time and make bad decisions.

The gstack is structured around this reality. Tan's workflow doesn't try to make one person faster at doing four jobs. It tries to give each role its own agent — a dedicated mode of operation with the right tool, the right prompt discipline, and the right output format. The founder coordinates. The agents execute.

This is the key distinction: the founder is the orchestrator, not the executor. When a QA agent is running browser tests, it isn't the founder context-switching into tester mode. It's a separate process. The founder stays in strategy.

What the gstack actually contains

The gstack centers on Claude Code — Anthropic's agentic coding environment — as the primary execution layer. But the architecture around it matters as much as the model itself.

The seven "gears" map roughly to:

1. Ideation and spec writing — Claude as a thinking partner, not a code generator. Prompts are structured to produce product specs, not vibes. 2. Architecture decisions — Using Claude Code to reason about system design before writing a line. 3. Implementation — Agentic code generation with explicit context windows, not one-shot prompting. 4. Testing — Autonomous browser agents running QA cycles. The founder reviews results, doesn't run the tests. 5. Debugging — Claude Code with full repo context, not pasted snippets. 6. Documentation — Auto-generated, kept current as code changes. 7. Deployment and monitoring — Lightweight infra decisions made with AI input, not tribal knowledge.

None of these are magic. What makes them work together is workflow discipline — consistent prompt structure, clean handoffs between gears, and a founder who knows when to override the agent and when to trust it.

Why this only works if you stop being the bottleneck

The most important insight from the gstack isn't technical. It's organizational.

Most solo founders use AI tools reactively — they hit a wall, they ask Claude, they get an answer, they move on. That's using AI as a search engine with better syntax. The gstack is different. It's a pre-planned division of cognitive labor, decided before the work starts.

That requires one discipline above all others: knowing which gear you're in before you start a session. Are you the PM right now? The engineer? The QA lead? Each role gets a different Claude Code context, a different prompt template, and a different definition of done.

Founders who skip this step end up with AI slop — output that looks finished but hasn't been stress-tested by the role that would actually catch the problem.

Julien de Waal, who runs Sprinkal, Sonscape, and Nova Labs under the Waalhalla holding structure, operates across multiple AI-native ventures simultaneously using a similar division of labor — agentic systems handle execution loops while he retains decision authority at the coordination layer. It's the same architecture Tan describes, applied outside pure software.

Claude Code specifically: why it's different from ChatGPT for builders

Claude Code isn't a chat interface with a code block. It's a terminal-native agent with file system access, the ability to run commands, read full repos, and iterate autonomously. The difference in practice:

• You don't paste code. Claude reads the repo.
• You don't explain context. Claude holds it across a session.
• You don't manually apply suggestions. Claude writes the diff.

For a solo founder, this compresses a 3-hour debugging session into 20 minutes — not because the AI is smarter, but because the feedback loop is tighter and the context loss between steps is eliminated.

This is why revenue per employee metrics for AI startups are starting to look structurally different from anything we've seen before. The productivity ceiling for a single technical founder has moved. Not slightly — by an order of magnitude.

The gstack isn't for everyone, and that's the point

The gstack requires a founder who can think in systems, write clear specs, and evaluate AI output critically. It doesn't remove the need for taste, judgment, or domain expertise. It removes the need for a team to execute on those things.

That's a meaningful distinction. The one-person unicorn model isn't about replacing competence with AI. It's about a competent founder operating at a scale that previously required 10 people.

Tan's gstack is evidence that this isn't theoretical anymore. A YC president is using it on weekends to ship. The tools are production-grade. The workflow is documented. The only remaining variable is the founder's willingness to adopt the discipline.

What to actually take from this

If you're building solo in 2026, three things from the gstack are worth implementing immediately:

1. Assign roles before sessions, not during. Decide if you're in PM mode or engineer mode before you open Claude Code. Don't mix them.

2. Use Claude Code for full-repo context, not snippets. One-shot prompting with pasted code is a fraction of what the tool can do. Give it the repo. Let it read.

3. Let agents own closed loops. QA, documentation, monitoring — these don't need your direct attention. They need your review. Design the workflow so the agent runs the loop and you approve the output.

The gstack isn't a product. It's a posture. And it's the clearest public signal yet that the AI-native company landscape in 2026 will be defined by founders who build this way — not companies that add AI to an existing headcount model.

---

Is your company eligible? Submit to the leaderboard → onepersonunicorn.co/submit