I've also not had great experiences with giving it tasks that involve understanding how multiple pieces of a medium-large existing code base work together.

If that's most of what you do, I can see how you'd not be that impressed.

I'd say though that even in such an environment, you'll probably still be able to extract tasks that are relatively self contained, to use the LLM as a search engine ("where is the code that does X") or to have it assist with writing tests and docs.

Your conclusion is spot on. Fuzz generators excel at fuzzy tasks.

"Convert the comments in this DOCX file into a markdown table" was an example task that came up with a colleague of mine yesterday. And with that table as a baseline, they wrote a tool to automate the task. It's a perfect example of a tool that isn't fun to write and it isn't a fun problem to solve, but it has an important business function (in the domain of contract negotiation).

I am under the impression that the people you are arguing with see themselves as artisans who meticulously control every bit of minutiae for the good of the business. When a manager does that, it's pessimistically called micromanagement. But when a programmer does that, it's craftsmanship worthy of great praise.