Smart or fluffy, Ozy. You can't have both
Most of my earlier career has been in formal (usually graphical) notations surrounding systems engineering, trying to create strong, machine usable ontologies and tools that allow users to create structures that represent the systems they're designing.
Graphical tools that allow you too create these things are great fun to program.
But using them is like alphabetising your record collection in seven dimensions - there are product, process, validation, verification, stakeholder, structure, function, behaviour, purpose axes that each system has a position on.
And to a lesser extent, UML tools suffer the same problems, and as the UML gets more and more complicated it only gets worse.
Part of the problem is that the value of getting machine usable information about your systems is high cost, and the cost is incurred by users who don't feel the benefit.
And part of the problem is the tools and notations themselves, the lack direct manipulation, and of notation condensation.
Flow happens in tools when the user forgets they exist, and feels like they're directly manipulating the concepts. Part of that is in the UI design - you have to design for flow - and related to that is notation condensation.
In the UML graphical notation for class diagrams, there are conventions for warts to add to indicate access modifiers on attributes and methods, for example +foo : integer[0..7] means the public attribute named 'foo' being of a collection of zero to seven elements of the type named 'integer'. Of the four UML tools I can think of off hand (ARTiSAN RTS, Poseidon, Rose 2002, Enterprise Architect) only Poseidon lets you enter the text directly and updates the relevent properties in the abstract syntax; the others you either have to open a dialog to edit them (EA), edit each part in a separate in-place widget (Rose) or use their own, non-standard syntax (RTS). So you have to both learn the notation, and learn the tool's take on how to interact with the notation.
ArgoUML/Poseidon has a very good HCI, it's a pity it's currently broken by bugs in Apple's Java so I can't use it. Of course, OmniGraffle lets you enter whatever you like, as it's not based on the semantics of UML, but is a diagram editor.
But even with the condensation, to say something consistent in UML means there are many non-graphical links. Some of these are apparent in the source code, such as a method containing its sequence of operations - not possible to show in UML; others, such as embodiment or requirements trace are relations between models, and the UML doesn't show these well. There is notation in SysML for some of the interesting relations between models, but I've yet to see anything that implements it, and I'm not convinced it will get used - it's a notation that spans diagrams, and the unit of context in UML is the diagram, which is it's greatest flaw. It's a language that hasn't the means to say the interesting things about a project.
So you can go to 'fluffy' approaches. One of the things I'm playing with is getting an atom feed of changes to the codebase of one of my projects, and building a temporal web showing a measure of the time each part was updated. I suspect there is a nexus around a couple of classes, which moves over time as the work progresses. Though I'm not sure what such a measure would mean, other than I was updating these things about the same time.
The thing that we want is to use as many sources of ambient information, information that is zero user cost, to get the most useful view of the project. UML is high cost and fractured; what I want is something like the feeling I have of the shape of the system.
Smart and fluffy, I want both.