My thought on #2 is to continually evaluate and store off a 'best' sample of the author's writing taken from everything produced that hits on a number of key metrics (ie. general wordiness, adjective usage, humor, unusual quirks that are statistically significant), then compile that down to where it might seem to be complete gibberish but an effective and performant prompt.

Many of us have 20, 30, or more years of written electronic communication.

Main issue preventing just* chucking this into one of the LangChains is (a) unrolling email threads mixing many people's writing, (b) keeping partitioned tones (workplaces, topics, networks, cultures) distinct.

* No such thing as "just".