illwrks
8 hours ago
I have been tinkering with this over the past few months. One story with a specific genre and demographic target, but as I've been reading each draft there are so many small issues that crop up, inconsistency across various parts of the story, items that are introduced, questions that go unanswered etc. Unless your goal is to become an editor of AI content, it's not worth the effort.
Perhaps in time it will be better but you will need a proper pipeline to write, assess, track changes etc... perhaps all that's needed is a GitHub of sorts for writing.
However... what is the chance of any book finding an audience if there are so many to compete with. It might make more sense in having something created just for you though.
sieve
7 hours ago
> inconsistency across various parts of the story, items that are introduced, questions that go unanswered etc.
This is a workflow problem.
What the LLM normally needs:
- very basic outline: what the story is about
- basic cast of characters: like fiction/plays of the 18-1900s
Then you go chapter-by-chapter. But the most critical thing is that you MUST provide the chapter beats. From start to end. Location, environment, events, any critical conversations. For a chapter that is 2-3K words, this chapter outline often becomes 2-400 words or more. Because you have the story in your head, you can handle the stuff that the LLM is likely to screw up.
There are ways to build a system where the LLM can keep track of many things, but then that becomes programming, which is just as boring.
shalmanese
4 hours ago
I had an amazing AI system that did this and worked well for me.
What you did was you started off with the high level vision of the story, you'd figure out the characters in detail, supply their motivation, think about how those things might conflict with each other etc. Then you'd think about how to add various plot points into it so different characters with motivation might find productive tension with each other. You map it out chapter by chapter, what beats go where, what things happen to push interesting conflict.
But the amazing thing about this AI system was it let you go deeper, you could go into individual conversations and actually place the words people said to each other, add color to the words by appending it with stuff like "she said, hesitatingly". Entire descriptions of scenes could be vibe generated by you imagining in your head exactly what it would look like and finding perfect words such that the image in your head would be transmitted over to the reader.
Finally, the AI system had a way to inspect the psychological state of people who might want to read your work by recruiting people who are friends of yours to read each work in progress and you can interrogate them just like a chatbot over what worked well, what didn't, did you understand what was going on over here? Did you feel invested in the characters? You can use this feedback to spin up a brand new version.
Anyway, all this stuff is a lot of work but don't believe the people who say you can just press a button and get a good story out. If you put in the effort, this AI program called notepad.exe can have you generating genuinely human sounding prose!
illwrks
7 hours ago
I won't go into too much detail, but I have a plot outline, development bible, story mountain, character outlines, style guide and more. I have four agents running on various tasks; writing, error checking, consistency and critique. They can generate a good story but there are holes and issues so it's far from perfect and I wouldn't give it away for free, never mind put a price on it.
To add to this, in my 9-5 I work as a designer. Colleagues in other parts of the business sometimes come to us with stuff they have created using AI tools. To a layman they look really good, but if you know what to look for you can spot the issues.
At this point in time my opinion is that any AI generated thing is great at making fools of the unaware, or those who are not looking at something critically. It can create something that looks like a duck, walks like a duck, quacks like a duck, but on closer inspection it's not a duck.
sieve
7 hours ago
I tried the programmer's method. Automating stuff. Then I realized I prefer the wood carver's method. So I take it slow.
I think and write the outline for a chapter. Then I have the LLM produce it. I read the draft. Tweak things. Reroll. Take the best from multiple rolls and merge them to produce a organic whole. Manually patch the minor differences in pacing and tone.
Rinse and repeat.
It takes 100-250 hours across 3-4 months to produce a story. But I drop 99.9% of the stuff within the first day or two because the plot or characters do not work.
I use the web front end directly. Let's me switch between any model including Gemma 4 running locally.
grey-area
6 hours ago
No, this is a problem with the nature of LLMs and the text they generate.
sieve
6 hours ago
I do not expect the LLMs to do everything. They are assistants. You still bear the responsibility of stitching everything together into a cohesive whole. Make changes as required.
That is where expectations differ I guess.
grey-area
6 hours ago
But they generate objectively bad text. Why do you need an assistant to do that?
Perhaps for some uses (say a summary of some business document, or generating code which is then discarded or edited) that's acceptable. For literature, do you really feel they are useful tools? So many changes would be required IMO that you're better off just writing the story you want to tell.
klibertp
3 hours ago
> But they generate objectively bad text.
By default, yes. That's exactly the same for code: by default, even with planning and patient nudging towards best practices, you get passable code at best. Not elegant, not performant, and not particularly readable, either. Basically, an uninspired salaryman type of code.
In both domains, you can get much better results in some specific circumstances. Prompt, skills, memory, and the task must align, but when they do, you can get good quality building blocks that you can then work with. It's crucial to recognize the instances where there's a chance of getting better-than-average results and ones where you could dump your whole week of tokens into and still end up with a mess. This is a skill of the LLM operator.
Once get that skill, the only issue is capitalizing on it: basically, how you fit handling of the generated building blocks into your workflow. If it's seamless, it can be a big win. If it's not - yeah, writing by hand from scratch is often better.
grey-area
an hour ago
I’ve certainly heard from LLM operators that there is a lot of skill in developing prompts and guiding their agents, but frankly it doesn’t seem like there is much benefit at that point unless you’re willing to accept mediocre results or spend lots of time rewriting and trying to find errors.
Given the examples of writing I’ve seen so far from them I’m skeptical that LLMs are useful for any writing where quality and truth are important.
sieve
4 hours ago
> But they generate objectively bad text.
They don't. I don't think people are taking full advantage of their capabilities. They produce perfectly functional prose. Sometimes, they produce something genuinely amazing to read. But that is quite rare.
People should stop expecting brilliant prose from LLMs without doing the preparatory work themselves first. Very few human authors can manage that in any case
illwrks
2 hours ago
I’d partially agree. They can generate very good descriptive text however the issue I’ve found is that when something is implied it’s not fully ‘understood’ in the context of everything else and then gets ignored by the agent.
sieve
2 hours ago
My solution to this, as the human-in-the-loop, is pretty simple. If the generated draft is missing something that was implicit in the chapter outline, I ask the LLM to incorporate it and rewrite the para/section. Then I fix the outline to make the point explicit so that a reroll does not make the same mistake again
grey-area
an hour ago
Show me some examples of good long form text please. Your hedging in reply (perfectly functional) does not inspire confidence.