I’ve been running the Qwen 3.6 MoE model on my gaming laptop and have been using it to critique screenplays lately. It can’t write worth a shit, but it sure can tell me what’s wrong with my own writing.
It pointed out things like this dialog is too wordy and it ruins the joke, this story beat feels “unearned”, this character feels like a “vehicle for jokes”, this line is an “exposition dump”, etc. When I ask it for suggestions on how to fix the problems it points out, its suggestions are all stupid, but when I fix the problems myself, it no longer flags them in follow-up critiques in a new chat session.
Qwen has obviously been trained on a lot of screenplays and writing how-to books. For example, I changed the character names in a classic sitcom script, removed the series and episode title, and it recognized the writing style of the series and then even told me what episode it was. It also gives me the same advice that screenwriting books I’ve read preach, except it can point out specific cases where said principles are not being applied.
It would be better to have a human critic, of course, but finding a human who is skilled at writing and who will take the time to critique your work can be difficult. Your friends also may not give you honest feedback. Qwen will, though. It’s not sycophantic at all from what I’ve seen, and in fact, it ripped my passion project to shreds. After I fixed all the problems it found, though, it ended up being a much better piece of writing, IMO.
I think using a LLM as a tool to improve your own work instead of as a slop generator is the right way to go. I also feel better about running them on my own hardware and using about the same amount of power I’d use if playing a game instead while also retaining control of my data.