I mean LLM can go through all the files in a source code and find repetitions that can be abstracted. Reorganize files into more appropriate structures etc. It just needs an optimal algorithm to provide optimal context for it.