A simple, interesting, and too-often-unspoken answer is that LLMs, especially with 100k+ context heaps, seem to be better at tasks that have contextual signific

A simple, interesting, and too-often-unspoken answer is that LLMs, especially with 100k+ context heaps, seem to be better at tasks that have contextual significance and scenery than they are at following one-off instructions. Or at least, that's very much the experience of my team. Like everyone else, I'm working on an MCP server right now. Mine though, is designed to populate a new context window with memories of Billy Strings shows and other bluegrass phenomena. Why? Well, in addition to that content being interesting to me and my wanting to see what LLMs have to say about it, I also notice that having this context seems to make them much better at solving bluegrass-adjacent problems in software, which is what I'm working on.