Quality of Prose, Style and Voice

Writing
1

A little time has passed and there have been some changes in the platform (new agents, specifically), so I thought I’d revive this topic that we discussed privately.

A lot of the draft prose I’ve generated on this platform (for treatments and chapters, as opposed to scripts) sounds awfully on-the-nose about dramatica theory (‘we’ll only show what we can prove”) and is in general pretty bad, strangely clipped writing, with the strange sensation that every sentence is ‘folded back’ into the subtext, and makes distincly unpleasant reading, like having your eyeballs pulled out of your head unpleasant. When I’ve tried modifying the style by giving it examples of my writing and saying, write this in my voice and style, I haven’t had a lot of luck. I wasted several days arguing with the AI and eventually gave up.

I found that switching to Claude*, I was able to take some of the stuff I’d created on this platform and got much better stylistic results with the same instructions. Now, there is a stage of the process where it’s not bad at all to have the on the nose stuff because, dang, its focus on structure is just relentless.

We discussed that it’s possible that writing it with the storyform agent active will get different results than having one of the other agents active. I’m very interested in trying that out, and wondering if you have specific guidance on how to approach this problem.

I will note, it got one chapter right, and when I read that chapter at a reading here in NYC people loved it, so mixed results, not totally negative. And the plotting is pretty good. It’s the actual writing that’s problematic.

Tbh, I rewrite every word anyway, and I only expect a draft, but I expect a higher quality draft than the one I’m getting.

*I also got pretty bad results from my ChatGPT subscription.

3

The thing with the actual “writing” of a story is that its so subjective - and where the artist really comes into play. I totally agree with you that its not the best, the Story Reception agent would be the best one for this use case, but I agree - it’s not amazing, and I’m sure that other models write better.

I have a feeling OpenAI’s next writing model will be insanely good, so while I would suggest or encourage you to help us improve the current Agent, I’d rather just wait until that new model they keep hinting at is released (the one that’s supposed to be great at creative writing).

Our focus is more towards the author side of things (the Storyforming and Story Encoding), but would like to be able to accommodate the full spectrum eventually.

The latest I’ve heard is within the next 6 months - which I’m sure will race by before you even know it. :smiling_face:

4

I’ve found that if you upload a DETAILED outline document (a true outline, not just “they went here, they did this”) under Storyforming, you get solid scene drafts - including more dialogue (they’re not magnificent - that’s your job) if you include some dialogue with your outline.

The bots learn your characters quickly if you tell them what the character feels and give a few lines of dialogue to train them. I’ve been surprised what they’ve come up with, and I’m writing a book that’s set 560 years in the past. It’s far superior in terms of research details too, to any other chat bot I’ve tried to use. Yes, it does occasionally hallucinate on real-life details, but every bot does.

You can tell the bot, “This character’s arrogant, but he’s also scared of humiliatiion,” offer the bare bones of the beats you want in a scene, and ask the bot to frame one scene that ends in mild humiliation, another with moderate humiliation, and a third that absolutely destroys him, and the bot does it beautifully. If you don’t like what you get from the first storyform answer, tell the bot what tone or outcome or emotional kick you want and try again.

You can fine tune from there, rewrite from there, ask for other emotional options, whatever. But I’ve learned the bot responds exponentially to whatever details you feed into it - especially to emotional details. The more questions you ask it, the more ideas you give it too.

I write character-driven fiction, not action-driven. So your mileage may vary if you’re writing a noir screenplay or something that’s action-driven.

You can also ask the bot, “What do I need to ask you to do to get more X, Y, Z, or make you give me less A & B.” You can also tell it to keep a specific tone in your scene, and it will do that do. But you have to tell it what you want. It can’t read your mind. At least, not until you’ve treated it like a collaborator for, say, 40 hours or so on one project.

1 Like