As the sibling comment talks about, I think DJT's tweeting pattern helps. He often uses short sentences, so with Markov order 2 (every three words generated is

As the sibling comment talks about, I think DJT's tweeting pattern helps. He often uses short sentences, so with Markov order 2 (every three words generated is a quote, more or less), there number of seams per sentence is lower. (By seams, I mean places where the generated text seems disjointed because the Markov chain consists of two non-semantic or semantically ambiguous tokens.)