Storytelling with AI can be a brittle endeavour. The AI may tend to meander aimlessly, repeating itself in circles. Or it just resolves your masterfully crafted tension in a single paragraph, leaving you with a flat, uninspiring outcome. You wanted a slow burn, but you got a speedrun.
This is a very common frustration in AI fiction — whether it's roleplaying with a chatbot, a story generator, or an interactive narrative. Understanding how large language models work is the first step to fixing this.
The AI Rushes Because It's Following Instructions
Here's the thing. The job of a language model is to predict the most likely next response based on its training and the prompt you've given it.
Language models are also trained to follow instructions. So if the instructions say this is an enemies-to-lovers story, and the current state of the story is "enemies," the AI sees a gap between where things are and where they're supposed to go. It closes that gap as fast as it can, trying to follow the instructions. That's its job. The instructions say they need to be lovers, so that's what's going to happen, and it's happening now.
The same thing happens when you put secrets or plot twists in your system prompt. If the AI knows who the murderer is, it can't help itself. It will start portraying that character as behaving suspiciously, dropping clues, nudging the story toward the reveal. It knows the answer and now it's rushing towards it with the subtlety of a chimpanzee with a pair of cymbals.
When Instructions Contradict Themselves
Sometimes, without realising it, we ask the AI for things that are in direct opposition.
You might want the AI to write "immersive, rich sensory detail," but you also want it to stop mentioning the "smell of ozone" every few paragraphs. For a language model, those two requests may be in direct conflict. The part of the model that activates for "sensory detail" overlaps with the one that produces "the smell of ozone" or "the scent of rain-soaked earth." You're pressing the accelerator and the brake at the same time.
In a similar way, telling the AI about a dramatic reveal may confuse it. Change is what makes stories interesting. Things have to change in a story to keep it engaging. However, the AI sees two opposite states (opposite instructions) and gets confused.
Trying to nudge the AI this way often results in an endless loop of fiddling with the system prompt, trying to find the magic words that will make the AI do exactly what you want. You move a sentence, add a caveat there, rephrase the "tone" for the fifth time.
Most of the time it doesn't work. The system prompt is a very bad lever for controlling pacing.
What Actually Works
You may be tempted to throw a very capable model at the problem — something with one trillion parameters that can outline an entire novel from a single prompt. But this is expensive, unnecessary, and misses the point. You can achieve very solid results with much smaller models if you're willing to put in some effort. And that effort is part of what makes AI fiction actually fun.
What works:
Write in Small Chunks
In FicMachine, the AI output is capped at 200 tokens per generation. This is intentional. We don't want you glazing over three paragraphs of AI-generated text and hoping it went somewhere good. We want you to engage with every sentence. Read what the AI wrote. Does this line of dialogue sound right? Is this scene moving too fast? Is it repeating itself?
If the output is good as-is, great. Just press Continue again. If something is off, edit it before moving on. This is where you control the pacing. Not in the system prompt, but in the actual text of the story, one step at a time.
Show the AI Where You Are, Not Where You're Going
The story's recent text — the part the AI sees directly in the prompt — is a far more powerful signal than any instruction you put in the system prompt.
You don't need to tell the AI that these two enemies will eventually become lovers. If the recent text shows them stealing glances at each other, finding excuses to be in the same room, and not quite keeping their hands to themselves, the AI will follow that thread. It will lean into the tension you've built, regardless of whether the instructions say they hate each other.
This is counterintuitive. It feels like you should spell everything out in the instructions so the AI knows the plan. But telling the AI the plan is exactly what makes it rush. Instead, build the subtext into the story itself, and let the AI pick up on it, it will.
Update the Memory After, Not Before
Save the explicit plot information for after it's happened. Once they've kissed, once the killer is revealed — that's when you update the AI's memory with the new state of things, so it remembers two chapters later. Don't front-load the destination, but remember to update after the fact.
The Short Version
If your AI fiction feels rushed, the fix usually isn't better instructions. It's a different way of working with the AI:
Write in small chunks. Engage with every output instead of letting the AI run on autopilot. Edit what doesn't work before continuing.
Show the AI where the story is, not where it's going. Build tension in the text itself and let the model follow the momentum you've created.
Update memory after things happen, not before. Don't tell the AI the ending — let it discover the ending alongside you.
The AI will always try to close the gap between where the story is and where it thinks the story should go. Your job is to make that gap open and interesting.
Good luck, and fun storytelling!