Mitchell Hashimoto demonstrates his approach to AI-assisted development while building Ghostty terminal emulator. He treats AI tools like junior engineers, providing clear architectural guidance and well-scoped problems rather than open-ended tasks. His workflow involves maintaining architectural control, working in parallel with AI agents for tasks like refactoring and cleanup, and using multiple AI models competitively. While AI excels at refactoring and simple bug fixes, it struggles with architectural decisions, high-performance data structures, and languages like Zig.
Table of contents
Maintaining Architectural ControlPrompting Like Coaching Junior EngineersWorking in Parallel with AgentsFrom Working Code to Maintainable CodeCurrent Limitations with AgentsModel Selection and TestingSort: