I've taken a throwaway approach to diagramming, where I'll produce them more or less on demand for a meeting or presentation, but not think of them as an enduring artifact. PlantUML is my friend here because I can knock out ugly but gets-the-point-across diagrams in 30 minutes before a meeting and check them into source control, so I can then take the bones of older diagrams and rework them for a fresh meeting.
I used to whiteboard for this, but that hasn't carried over well in the remote world. What I miss about whiteboarding though is that you can tell a story as you draw, so whoever's viewing can watch something unfold from a blank slate while I'm walking them through the history of whatever system we're describing. That said, I can make a PlantUML diagram much more correct than a whiteboard.
All that said I too would love for more of that to be automated increasingly through AI. And I suppose it should make sense conceptually, because for me the value of a crafted diagram over an automated one is that no one really wants to look at the insane ERD of an OLTP database or a production object model. They want the digestible high level vision of the important bits, or the bits that are relevant to the conversation taking place. So it's a summarization problem. How to get the right data to produce a correct summary is interesting--I'm sure if I look there's a dozen papers to read on a similar subject :).
I used to whiteboard for this, but that hasn't carried over well in the remote world. What I miss about whiteboarding though is that you can tell a story as you draw, so whoever's viewing can watch something unfold from a blank slate while I'm walking them through the history of whatever system we're describing. That said, I can make a PlantUML diagram much more correct than a whiteboard.
All that said I too would love for more of that to be automated increasingly through AI. And I suppose it should make sense conceptually, because for me the value of a crafted diagram over an automated one is that no one really wants to look at the insane ERD of an OLTP database or a production object model. They want the digestible high level vision of the important bits, or the bits that are relevant to the conversation taking place. So it's a summarization problem. How to get the right data to produce a correct summary is interesting--I'm sure if I look there's a dozen papers to read on a similar subject :).