I have thought about this a lot. I was going to do some stuff with gzip, but fundementally, I came to the conclusion that the most elegant solution (no hidden chars / extra encoded data coupled with export) might actually be training a neural net to do it. Also, this would mean being able to import diagrams that weren't even made on Cascii. Training data can be self-generated too.

Perhaps a good middle ground would just be offering the gzip dump.

v1.0.1 now has offline import/export

https://github.com/casparwylie/cascii-core