Of course, what exactly a level of abstraction depends on how you model, and personal opinion. But I find it a more useful starting point then "line count".
What you say about specifications certainly rings true. In those cases the reader's preferred abstraction level is pretty much given.
In principle I agree. In practice, if your function needs to do more than say five things one after another then something has gone wrong. If you were sending a human a set of instructions on how to do something, that's the point where you'd start grouping them into logical "meta-steps" (e.g. prepare the ingredients, make the broth, prepare the meat, finish cooking the dish).
Chopping a 12-step function into 3 4-step functions without regard for what each step actually does will make it less readable, just like doing the same to a recipe, or putting paragraph breaks and full stops into a passage at random. But a single function that does a long sequence of things is a code smell in the same way as an overly long sentence; it's well worth looking for a way to logically group these steps into a structure that will help the reader understand.
> Modern IDEs help, but there's only so much they can do.
Exactly. In my experience (YMMV) 100% of the time the complaints about using "too many functions" (or similar) come from people who use poor tools or don't really know how to use them.
If your code does n operations in sequence, just list those operations. Don't hide them off somewhere.
Code that reads linearly is by far the easiest to read. At least when you are familiar with the language and idioms. It's like the difference between a novel and a choose-your-adventure book.
Modern IDEs help, but there's only so much they can do.