It's an interesting idea, but seems unfeasible. I assume the editor will be shipped in a binary, which contains the trained models if machine learning is used in it, but it would still take a fair bit of computation power and memory to run. Furthermore it seems like overkill. I can express a concept much clearer in code than I can in comments. Comments are there to help me get back into the code if I've stepped away from it.<p>I think that writing function skeleton and automatically generating test, and vice versa, however could be feasible and it should be possible to do. The decision of which skeleton/test to generate could be done with a very inexpensive statistical computation or a simple machine learning model. The outline of the project seems far too large to be feasible and I think some of it would hinder development rather than ease it, but the test/function generation seems useful.