How to Stop Your AI from Speedrunning the Story

You built it carefully. The glance held a beat too long, the argument that was really about something else — and then, six turns in, the confession. The kiss. The rescue. Scene over, tension spent, and the two hundred turns of slow burn you wanted are gone.

This isn't the model being dumb. It's the model being agreeable. Language models are trained to resolve: a question gets an answer, a tension gets a release. Fiction runs on the opposite — tension is the product, and resolution is what you spend it on, as late as possible. Left alone, every model speedruns, some worse than others.

Name what must not resolve

The single most effective pacing move is a direction that names the withheld thing: this is a slow burn, and the feelings stay unspoken until I say otherwise. Models follow explicit non-resolution rules surprisingly well; what they can't do is infer that you wanted the tension kept. In Underfiction that's a custom direction — written once, applied every turn, synced with the story. Elsewhere, restate it whenever the scene starts leaning toward a confession.

Direct the beat, not the plot

When a scene accelerates, don't argue with the reply — direct the next beat smaller. "Slow down. Let the silence do the work." "Stay in this room." "She almost says it, and doesn't." A beat-level direction gives the model something to write instead of somewhere to arrive. The scenes you remember from novels are almost all beats a lesser writer would have skipped.

End scenes before they sag

Pacing isn't only slowing down — it's knowing when a scene's question is spent. When it is, cut: end the scene, open the next one hours or days later. A hard cut resets the model's urge to resolve, because the new scene has a new question. Scene structure is why long stories hold in apps built for chapters, and why they sprawl in a single endless chat.

The escalation ratchet, used deliberately

Models escalate because each reply mildly outbids the last. You can ride that instead of fighting it: keep your own turns one notch below where you want the scene, and let the model close the gap. If you write at intensity seven, the reply lands at nine. Write at four when you want a six. Your input is the thermostat.

When it's the tool, not the technique

Some platforms make patience structurally hard: replies chopped short, context windows that forget the withheld thing existed, or models tuned so agreeable they experience an unresolved scene as a bug. If you've set standing rules and directed beats and it still speedruns, switch models — pacing discipline varies enormously between them — or switch tools. Underfiction lets you change the model on any turn, so you can hand the delicate scene to the writer with the best sense of restraint.