There are obviously tons of AI powered editors and IDEs out there at this point: VSCode with GitHub Copilot (or SuperMaven); Cursor; Void; Pear. I personally find it better to just directly use the web apps for ChatGPT and Claude and use regular VSCode.

But I notice that I constantly find myself in a loop where I generate code (either a whole code file or just a section of code), paste it in, and then immediately see bugs/errors/warnings from the linter (Ruff for python, eslint for typescript, etc.) in VScode and copy all the "messages" about these issues from VScode and paste them into the LLM web app as a follow up message.

Or, I might paste the output of `bun run build` or `cargo build` or whatever if the warnings don't show up in the editor. Very often these are type errors in Typescript (if I haven't been careful enough specifying exact types), or they get the API/syntax wrong for an npm/pip library (at this point, the flagship commercial LLMs seem to be pretty good at getting the standard library code mostly right in popular languages like Python/Typescript).

That is, I personally am the one "closing the loop" between the LLM and the compiler/interpreter. And of course, that's not enough by itself to fix all bugs or to ensure that your code works, because you need to really run things end-to-end with real data and look at the outputs to see that they make sense; but you can certainly catch a huge number of problems like that.

I know that a few months ago there was a lot of buzz about the demo of Devin AI, which was more of an agent that closed the loop between code generation inference and checking if the code works (or at least compiles or doesn't trigger the linter), and then can iteratively refine/fix the code until it does work, and commit the changes to your git repo when it's working with a meaningful commit message.

And I know there are at least a few open source versions of something similar, such as https://aider.chat/

I guess my question is, are any of these good enough now to use seriously in day to day work? Do they actually save time? The examples I see on the Aider website for example don't seem that appealing, because I would rather just directly use the webapp from ChatGPT or Claude directly when starting a project from scratch.

Where these projects would be most helpful to me is when I have the basic structure of the project already set up, but there are a bunch of issues and problems and I don't want to have to sit there mechanically going through all the code files and iteratively fixing them myself in a manual process. Is anyone here using a project like that in the manner I described? Is it helpful to you? If so, can you describe your workflow?