site banner
PaperclipPerfector  3d ago (text post)   2577 thread views

Small-Scale Question Sunday for August 16, 2026

Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?

This is your opportunity to ask questions. No question too simple or too silly.

Culture war topics are accepted, and proposals for a better intro post are appreciated.

2
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

I want to experiment with LLM assisted story writing and given how many folks here are LLM advocates, maybe someone can help me there. My main problem is that I absolutely suck at writing, so I need something that takes character, setting and (detailed) plot point descriptions for a scene and can expand that to text using my chosen LLM model (via any OpenAI api compatible provider).

More specifically, I want a free local app (preferably native Windows but a local webui will do in a pinch) that lets me write the overall style guide, character / setting notes, "previously on Buffy the Vampire Slayer in the last chapter" condensed history notes and then the notes for the scene (and the beginning of the next scene to help flow). The scene notes would be a combination of "A does X, B mentions Y" and occasional inserted sentences like "The silence was interrupted by a doorbell ring" (that are more ideas for the LLM and don't need to be preserved as-is). It should let me generate multiple variations, manually edit the previous prompts and generated results for continuation and crucially never delete unused generations so I can later steal sentences and ideas from them (but I don't need explicit branching from every response).

Does something that does all this exist as a local app? And one that has a usable UI without ridiculously small text and non-existent contrast?

I don't think that exists... Tooling is still at the point where you need to adapt your workflow for the tools available or build your own.

Everything will be a lot smoother if you can get a setup working with obsidian plugins or use claude code / codex on a git controlled directory.

I'll give some caveats right away: don't expect miracles in terms of prose quality, interesting plots, or coherency of metaphor. This is a moderately useful way to get a story framework without getting stuck on the rough spots, but you're going to have to do a lot of heavy lifting to get it into a state that's enjoyable even for you to read, and a lot of people will bounce off it no matter how aggressively you edit.

There's a few different approaches. But only a few have good UI, and only a few reliably maintain unused generations, and there's not a ton of overlap.

Most of the standard webuis or desktop ui chatbot implementations do implement what you want, just poorly. LMStudio's probably the best from an organizational perspective, since it supports simple drag-and-drop separate folders and subfolders. Biggest downside's just that it seems pretty abandoned and the company's pivoted to their agentic Bionic tool. Llama.cpp and Unsloth Studio support forking conversations, editing past prompts (or regenerating an answer to an existing prompt, and then switching back and forth from previous versions. They just suck at it, and it's really easy to lose past information if you aren't fastidious about labeling and forking, and even then it gets really messy. Get used to exporting old runs to JSON and sorting them into folders on disk, or you'll go nuts. I will caution that a lot of these want to provide their own endpoint, rather than interface with someone else's, though. Still, it might well be the easiest and most straightforward way to test out if that's actually helpful.

For an example of that approach with prompt-and-response, see here, Muse-Glimmer (imo, not great) or here, Qwen3.8 (almost passable, but definitely not what I was aiming for) (cw: Monty-Python-esque comedic gore, ARK fanfiction). You really don't need much in the way of fancy formatting for individual scenes or five-to-six scene series. For longer series, asking for an outline first, then prompting individual scenes (or parts-of-scenes) from the outline will let you expand into the mid-thirty scenes before it starts to get janky and around sixty scenes before the context starts going loopy.

There's a few dedicated writing tools. Unfortunately, they fail your crucial requirement: they technically support branching, but barely, and they don't enforce it. Just will list them so you know about them. SillyTavern is the roleplay power-user tool. Its support for flicking back and forth from past prompt edits is less good and the UI is a real pain to learn, but it has extremely strong tools for separating characters, worlds, lore, and actual prompts. I will admit I don't have too much experience with it, but I do give it props for not bundling in the inference engine. KoboldCPP is more writing-focused, but takes a similar approach. People tend to be happier about the user-interface, less happy about the inference engine (unfortunately, baked-in hard).

Agentic approaches are more complicated. I've been messing with an agent layout for Claude Code/pi-agent/whatever, and it's powerful, but it's also completely incompatible with the 'preserve past attempts' version. You can tell it to do so, and it might even try for a while, but then it'll just randomly ask for permission to rewrite a file or delete a directory, and it'll usually have good reasons, but it's a pain, and if you bulk-granted edit permissions then you're just out of luck. Maybe some agent framework will have better ways to set up and coerce permissions, but I think it's a dead end for writing prose, specifically. Still need to try out some of the more specialized stuff, just not optimistic.

I've started work on a procedural Java app that might be closer to what you're looking for, and very specifically targeted at writing short stories or novellas, but it's really ugly slop code and barely above a python script right now. If you aren't able to find something in the next week that you like, I'll see if I can get this into a better form.

I suppose I can always shill SillyTavern again, which despite a barely-responsive dev is still the most feature-complete LLM frontend I've tried; out of home for the week, I'm currently forced to use Agnaistic and I stumble over UI knobs and switches too often to be able to goon at all.

I don't have the Tavern UI before me but off the top of my head:

lets me write the overall style guide,

The prompts settings let you set the exact content and order of the entire wall of text you send to the LLM.

character / setting notes,

Main body of character notes goes into card definitions, setting can go into the additional info -> scenario field. It's sent as text anyway so this is more of a convenience thing.

"previously on Buffy the Vampire Slayer in the last chapter" condensed history notes

goes into Author's Note, optionally buried on some # message depth in the chat history; I usually just put PAST EVENTS (separated with newlines etc.) after char defs.

and then the notes for the scene (and the beginning of the next scene to help flow). The scene notes would be a combination of "A does X, B mentions Y" and occasional inserted sentences like "The silence was interrupted by a doorbell ring" (that are more ideas for the LLM and don't need to be preserved as-is).

If you don't want to actually self-insert chat, you can just put instructions in your "replies" as long as your prompt sequence/preset says that this can be a possibility. Usually LLMs are smart enough that I can just "say" something like [OOC: now draw describe her getting an education] and get the needed continuation.

It should let me generate multiple variations, manually edit the previous prompts and generated results for continuation and crucially never delete unused generations so I can later steal sentences and ideas from them (but I don't need explicit branching from every response).

All in by default, in fact I think "swipes" and generally complete control over chats is the comfiest feature of ST by far and I have no idea why this is not the default in corpo apps. I tried the Claude app for work after years of chatbotting and I frequently raged at the stupid frontend preventing me from editing old responses or forking chats for no reason far as I can see. Cursor feels much better in this regard.

Your call on how usable ST's UI is, it's fairly customizable and I've grown to like it against my will anyway.

I'll echo other posters and caveat that my own experience with genning OC things based on my own drafts has been mostly disappointing (too prone to getting stuck on character quirks and needing to show them off in some fashion in every fucking response), but I'm not the karma police here. Might be a prompt issue on my part.

What you're asking for might exist, for example Emily Youcis has been hyping up the "Slop Canon" which is some sort of specialized tool for making music videos that I couldn't find a direct link to), and all the various coding tools are basically that, but for code, but I've never heard anyone mention anything like that.

Barring some sort of magical orchestration tool, I think this is the completely wrong way of going about this. Like mentioned in the previous Tinker thread, I've been working on the story board for my game's intro. Most of the AI usage was images, but I also decided to solicit feedback from AI (various models). I agree with @Skulldrinker, it sucks as an editor. It's ideas suck, it's criticisms are better, but very hit or miss, and in the event it's a hit, the proposed fix sucks. I still found it useful, but I'd rate it's effectiveness as slightly higher than the Rubber Duck method.

The workflow I'd recommend is:

  • write the thing yourself, no matter how cringe.
  • if you're really insecure, ask AI for a rewrite, and pick and choose the parts you like
  • ask it for criticism, and fix only only the things you think it has a point on
  • if you're stuck, ask it for alternatives, but try to narrow the problem down as much as possible. It's capable of coming up with decent lines, but I wouldn't trust it with anything bigger than that.

it sucks as an editor. It's ideas suck, it's criticisms are better, but very hit or miss, and in the event it's a hit, the proposed fix sucks

Good thing I'm not looking to use it for any of these.

write the thing yourself, no matter how cringe.

This is not an option and that's not negotiable.

I'm explicitly not looking for AI to edit or polish my writing. I'm looking for AI to do the bulk writing (multiple variants of it) based on my detailed notes. I have the ideas, pacing and plot points and I also intend to do the editing myself (I have no problems handling either of those parts). All I need is something to help with the bulk writing process (which is where I've always had major problems - the only way I was able to get my masters thesis text written was with the help of prescription amphetamines and dedicating a month to literally nothing else than writing the remaining 25 pages of loosely spaced text based on work and theory that had long been finished).

It's capable of coming up with decent lines, but I wouldn't trust it with anything bigger than that.

This is pretty much all I'm asking it to do. "Here's the contents of a paragraph or two. Turn them into a draft that I will edit and rewrite into final text." Consider it the text writing equivalent of telling LLM to write a function that does exactly X and then cleaning it up by hand.

The bulk writing IS the writing. If you don't/can't enjoy the act of writing, you shouldn't be writing. I'm serious.
What you're going to get is an abomination that the AI will tell you is a masterpiece and should be shopped around to major publishers and will totally get optioned for a film adaptation, because every line is Chef's Kiss.

The only "legit" use of AI for writing fiction is personal (and ephemeral) use, so basically jerk-off material that will never be seen by any human other than you.

Dude, just fuck off. For good.

Who are you to gatekeep what other people are allowed to enjoy? Who made you the thought police who gets to decide what someone else can do on their computer?

I've never even hinted at publishing or posting the results anywhere. All I want is to do some AI assisted writing experiments in a way that suits my brain. Something which apparently you are allowed to do but others like me are not, for some vague reason.

Your input is not appreciated in the slightest and you should feel bad for being one of the worst gatekeepers I've ever come across here. Go beat up small kids or something like that if you're so desperate to cause misery because you certainly aren't contributing anything positive here.

Sorry, I suppose everyone has to reach the point of disillusionment with LLMs on their own terms. I'm being unnecessarily vociferous about it and making assumptions, my apologies.

This is pretty much all I'm asking it to do. "Here's the contents of a paragraph or two. Turn them into a draft that I will edit and rewrite into final text."

That's more than a line...

Have you tried doing that a couple times manually to see if it returns what you want, and can work with?

I think Youcis' Will Stancil Show is actually a good reference for you here, even if it's a different medium. I can imagine method working, but the end result might be a bit uncanny valley.

Have you tried doing that a couple times manually to see if it returns what you want, and can work with?

Yes, the method actually does seem to work reasonably well for my purposes based on initial tests. The UIs that I know of are just very poorly suited to it, making it more laborous than it should be. It's unlikely to win any awards but that's not the purpose and it's doubtful I'll even post such stories anywhere in the end.

It's similar to the highly succesful method we used on a bunch of group projects in my university days where I paired up with a friend. I would dictate the outline and throw out a whole bunch of ideas, points and individual sentences. My friend would write the draft based on those and we'd edit it together. It was a very fast way of working, easier than the normal way for both of us and we reliably got top grades doing it.

Another analogy would be that I'm taking the part of a script writer, director and film editor while the LLM handles the acting and camera work. Which is in fact what a bunch of people on /r/StableDiffusion/ are doing with the new Minimax H3 video model (second part). Only, I want to do that with story text instead of video.

What you're trying to do is bad and you shouldn't do it. Accept that you must either get better at writing, write at your current level of skill, or not write, not everyone is meant to be a writer and that's okay.

I've tried using an LLM as an editor for writing fiction and RPG adventures. It's bad at it.

You write like a disobedient LLM. Answer his question or I will delete the app altogether!

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats