Codebase is 250% AI generated
2 days ago
- #AI-generated code
- #LLMs
- #software development
- The author built an AI-powered tutor for advanced chemistry with a codebase that is 250% AI-generated, highlighting the significant human intervention involved.
- Information Architecture (IA) should be handwritten as it forms the nucleus of the software, dictating object ownership, identification, relationships, and lifecycle.
- Useless features should be aggressively removed in AI-generated code to prevent them from propagating and bloating the system.
- Consistent naming is crucial for greppability, aiding AI in understanding and modifying the codebase effectively.
- Setting up frameworks and patterns in the codebase helps AI generate code that aligns with the desired structure and style.
- AI excels at handling third-party integrations and one-off setup tasks, reducing the need for deep knowledge of specific tools or APIs.
- Debug tooling can be enhanced by leveraging AI to interact with databases and other systems, though supervision is necessary to prevent errors.
- LLMs are currently blind to visual elements, relying on text-based interactions and tools like Playwright for limited visual understanding.