Specifically within a single file, I find LLMs can easily extend and maintain large sets of slightly varying concrete code, yet they struggle when that code is

Specifically within a single file, I find LLMs can easily extend and maintain large sets of slightly varying concrete code, yet they struggle when that code is compressed into abstractions. The concrete examples form a dataset. Attention generalizes from that dataset. I'll agree that rule of three continues to apply for patterns across files, where there is less guarantee that all patterns will be read or written together in any given AI action.