I beg to disagree, Salvatore... Have a go at VS Code with Agent mode turned on (you'll need a paid plan to use Claude and/or Gemini, I think). It gets me out of vim, so yeah, it's that good. :)
Tip: Write a SPEC.md file first. Ask the LLM to ask _you_ about what might be missing from the spec, and once you "agree" ask it to update the SPEC.md or create a TODO.md.
Then iterate on the code. Ask it to implement one of the features, hand-tune and ask it to check things against the SPEC.md and update both files as needed.
Works great for me, especially when refactoring--the SPEC.md grounds Claude enough for it not to go off-road. Gemini is a bit more random...
Interacting with the LLM directly is more work. What I mean is that a wrapper in the best conditions will not damage too much the quality of the LLM itself. In the chat, you continuously avoid suboptimal choices executed by the model, if you are a very experienced code, and a fix here, a fix there, you continuously avoid local minima. After a few iterations you find that your whole project is a lot better designed than otherwise.
Oh, sure. I don't rely on the LLM to write all the code. I do "trust" it to refactor things to a shape I want, and to automate chores like moving chunks of code around or build a quick wrapper for a library.
One thing I particularly don't like about LLMs is that their Python code feels like Java cosplay (full of pointless classes and methods), so my SPEC.md usually has a section on code style right at the start :)
Tip: Write a SPEC.md file first. Ask the LLM to ask _you_ about what might be missing from the spec, and once you "agree" ask it to update the SPEC.md or create a TODO.md.
Then iterate on the code. Ask it to implement one of the features, hand-tune and ask it to check things against the SPEC.md and update both files as needed.
Works great for me, especially when refactoring--the SPEC.md grounds Claude enough for it not to go off-road. Gemini is a bit more random...