Capturing writing style with ngram-based input and individual-character input are very, very different tasks. That's several ballparks higher in difficulty.

With ngrams, Markov models are perfectly sufficient. With individual characters, complex concepts need to be remembered across many, many characters of input.