>>533290750
My strategy with deepseek lately has been to directly handhold the model through the early parts of the story. I give it direct feedback explaining what I liked and didn't like. If a particular output keeps having a particular problem, I sometimes edit the instructions and retry.
I keep that back-and-forth in context. Once I build a few paragraphs of something decent it gets better. And then I just have to nudge it a bit or remind it not to do certain things. I edit out any garbage as I go (like unnecessary or nonsensical statements, or "somewhere, a car door slammed" shit).