My current theory is that this reflects a weakness in Claude’s ability to see the “big picture”.
When writing code, I have to explicitly tell it how to structure things at a high level, or the result is sort of a flattened spaghetti. Similarly, when it’s explaining things, it’s not good at pulling out unifying concepts and explaining top-down as a smart human would do. It groups little things together but often doesn’t generalize or synthesize explanatory connections from them.
I’ve been experimenting with explicitly working through a sequence of outputs at different levels of detail, but I haven’t found a consistently successful method.
I use both Claude and Codex for a lot of code- and prose-adjacent tasks, and I'm increasingly convinced that language models are quite terrible at, well, modeling. Not sure if that's ironic or not.
With the right blend of context and prompting, I can often get them to "lock into" an existing model, but using them to generate a novel outline or sketch, whether it's for an essay or a module, usually results in garbage.