What is Underfiction?
Underfiction is an interactive fiction platform. You describe a scene, direct what happens, and a frontier AI model continues the story — literary prose, not chatbot replies. Together you build something that reads like a novel.
Built by one person
Underfiction is designed and built by Stefan Hoschkara, an independent developer in Zürich. I have spent more than twelve years building iOS and macOS software and shipping my own products. I built Underfiction because the tool I wanted did not exist: strong models, creative freedom, prose worth reading, and stories kept off a company cloud.
Being a one-person product makes the responsibility legible. The product decisions, model choices, privacy architecture, support, and editorial writing lead back to the same person.
Why Underfiction?
Frontier models, uncensored. Full creative freedom with no refusals, no moralizing, no watered-down prose. Literary fiction routinely contains violence, desire, and moral complexity — an AI that can't engage with those can't write fiction worth reading.
Literary quality. Every response is written as prose. Scene-setting, interior monologue, tension, subtext — the kind of writing you'd actually want to read back.
Simple by design. No jailbreaks, no config files, no prompt engineering. Set a scene and start. The interface stays out of your way.
Private by design.
Stories, characters, and worlds stay on your device. Export a story with its media and connected cast and world, or move your complete library in one portable .underfiction file. Inference runs through Venice, which separates your identity from model requests at the infrastructure level. Learn more about privacy →
How models are evaluated
Models are evaluated as fiction writers, not as general assistants. I run candidates through the same repeatable scenes: dialogue-heavy exchanges, close narration, tonal changes, long-context continuations, and directions that require the model to preserve character and subtext while moving the story forward.
Automated regression checks catch structural failures: whether the production prompt is assembled correctly, whether dialogue and narration directions produce the intended kind of response, and whether required output metadata survives. I then read the prose. The practical questions are whether the model follows direction without flattening it, holds point of view and character voice, notices emotional implications, avoids assistant habits, and remains coherent as context grows. Cost, latency, and refusal behavior matter too, because a model that writes one fine paragraph but fails in an actual story is not a good product model.
Model and prompt changes are tested against the same scenes so that comparisons are based on behavior, not release notes or a single showcase response.
What “literary prose” means here
“Literary” is an operational standard, not a claim about prestige or genre. Underfiction asks a model to meet six concrete expectations.
Narrated scenes. Paragraphs, not chat transcripts, headings, lists, or asterisk actions.
Narrative control. A deliberate point of view, tense, tone, and narrative distance.
Character depth. Interiority, distinct voices, and dialogue that can carry subtext.
Concrete detail. Setting and sensory language when they sharpen the scene, not as decoration.
Direction with movement. The reader's instruction is followed, but the response still produces dramatic movement.
Prose hygiene. No assistant framing, moral commentary, repetition, or generic filler.
The prose can be spare, comic, romantic, gothic, or explicit. It does not need to be ornate. It needs to behave like prose inside a story.
How stories work
Stories are made of scenes. Each scene has a setting and a cast of characters. You direct what happens and the AI continues the story. When a scene ends, it gets summarized and the next scene picks up with full context but a fresh window. Within long scenes, older turns compress automatically so context stays sharp without ballooning your costs. Stories can run for hours without losing coherence.
Credits
Underfiction uses credits — simple pay-as-you-go. Frontier models are expensive to run. Credits let us offer them without subscriptions, ads, or compromises on model quality. You only pay for the prose — summaries and context compression are on us.