I've been experimenting with GPT to try to write code in an existing code base. This consists mainly of copy and pasting existing code snippets and getting edits or adding new functions. Or using tools like smol-dev [0] to generate entire projects me for
j-dev originally started as a refactor of smol-dev but I quickly realized it was unrealistic to nuke your entire project and rebuild from scratch. Plus it was overly expensive in terms of tokens used.
j-dev is meant to be used in existing projects and allows the LLM have discretion over what it need to know about your codebase, while giving the user full control. j-dev also handles invalid responses, and updates your code with permission, all while providing full transparency and a relatively simple implementation.
From experience using it, it's somewhat useful although still often has a shallow understanding of the code base despite being able to request more. But hopefully this gets better as LLMs improve or I get better at prompt engineering.
[0] https://github.com/smol-ai/developer
