How to Make Claude Code Your AI Engineering Team

Key Concepts

• Agent Era: A new paradigm in software development where AI agents perform complex tasks by mimicking human team structures (roles, processes, and reviews).
• GStack: An open-source framework designed to turn AI coding tools (like Claude Code) into a specialized engineering team.
• Thin Harness, Fat Skills: A methodology where the scaffolding (the AI wrapper) is kept minimal, while the "skills" (specialized modules) are robust and expert-driven.
• Adversarial Review: A process where the AI critiques its own design docs and code to identify flaws, security risks, and missing features.
• Software Factory (Level 7): A state of development where a single engineer manages multiple parallel AI sessions (PRs, branches, and features) simultaneously.

---

1. The "Agent Era" and GStack Methodology

Gary Tan, CEO of Y Combinator, argues that the bottleneck in AI-assisted coding is not model intelligence, but the lack of structured process. He introduces GStack as a solution to encode human-like team dynamics into AI agents.

• The Problem: Out-of-the-box models often "wander" or guess, leading to plausible-looking but broken code.
• The Solution: GStack provides a "thin harness" that forces the AI to follow a rigorous, multi-step workflow—from ideation to deployment—modeled after YC’s internal processes.

2. Core Skills and Frameworks

GStack operates through a series of specialized "skills" that act as team members:

• Office Hours: A distilled version of YC partner sessions. It forces the user to answer six critical questions to reframe the product, identify the "wedge" strategy, and validate the business model before writing a single line of code.
• Adversarial Review: An automated critique phase where the AI attempts to break its own design. It identifies gaps in privacy, 2FA, and error handling, often improving design scores (e.g., from 6/10 to 8/10).
• Design Shotgun: A visual brainstorming tool that uses image generation (via OpenAI Codex) to provide multiple UI/UX variations for the user to evaluate.
• Browser Automation (GStack Browser): A wrapper around Playwright and Chromium that allows the AI to interact with the web, download files, and perform regression tests in a visible browser environment.

3. Step-by-Step Development Process

The video demonstrates a workflow for building a tax document aggregator:

1. Ideation: Use office hours to define the problem (e.g., finding 1099s in Gmail).
2. Refinement: The AI challenges the user's assumptions, suggesting a "wedge" strategy (e.g., moving from document aggregation to a CPA marketplace).
3. Planning: The AI generates multiple technical approaches (e.g., OAuth vs. Browser Automation).
4. Adversarial Review: The AI identifies 16+ issues in the design doc and auto-fixes them.
5. Design: The AI generates UI options; the user selects the best one.
6. Execution: The AI writes the code, followed by a review phase that performs staff-level bug catching.

4. Key Arguments and Perspectives

• The "Autistic CTO" vs. "ADHD CEO": Tan suggests using different models for different roles. He characterizes Claude Opus as the "ADHD CEO" (great for ideation) and Codex as the "autistic CTO" (great for execution and bug fixing).
• Parallelization: Tan emphasizes that the future of engineering is managing 10–15 parallel AI sessions. He no longer maintains a traditional to-do list; instead, every bug report or feature request becomes a new "work tree" in Conductor.
• Supply Chain Security: By using GStack to review incoming open-source fixes, Tan mitigates the risks of supply chain attacks, treating the AI as a security gatekeeper.

5. Notable Quotes

• "The bottleneck here is not the model's intelligence. As long as you set the models upright, they are already smart enough to do extraordinary work on your codebase."
• "The scaffolding should be trivially thin. GStack is my implementation of the thin harness fat skills approach."
• "This is the most incredible time in history to build software. The barrier to building just collapsed."

6. Synthesis and Conclusion

GStack represents a shift from "AI as a chatbot" to "AI as a software factory." By automating the tedious parts of development—QA, design brainstorming, and adversarial review—Tan demonstrates that a single developer can achieve the output of a 10-person team. The framework is available at github.com/gritan/GStack, and the primary takeaway is that success in the agent era depends on the process you wrap around the model, not just the model itself.