It's not hard to get them to copy a style, you just have to provide examples and they will happily produce similar text including grammatical and spelling mistakes. The trouble is with the composition and novelty. Most of the big models have had all of the interesting parts hidden behind a wall of RLHF. Local models are better since you can use ones that are not indoctrinated as a "helpful assistant" and also control the system prompt, temperature and see the top K alternate tokens which let you steer them in interesting ways.