Uncensored AI for horror roleplay that actually scares
Horror lives or dies on restraint, and most chat models have none: they either refuse the scene or explain the monster in the first reply. Here is how to set a model up so the dread builds.
What horror asks of a model
Good horror roleplay is mostly about what the narrator does not say. The model has to hold information back, describe the wrong detail at the right time (the hallway light that is on when it should be off), and let a quiet turn sit without filling it with action. It also has to follow through: when the thing in the cellar finally gets someone, the scene needs weight and consequence, not a cutaway.
That means two separate skills. The first is pacing, where the model slows down, uses short paragraphs, and lets your character notice things. The second is permission, where violence and gore can be written plainly when the story earns them.
Why mainstream models ruin the scare
Filtered models fail horror in predictable ways. They soften injuries into vague "wounds", add a reassuring line at the end of a frightening turn, or step out of the story to ask if you are comfortable. Some refuse outright once a character is hurt badly. Others do the opposite of withholding and narrate the creature's full anatomy and motives in turn two, because they are tuned to be helpful and complete.
An abliterated model removes the refusal behavior, which fixes the first problem. The second problem, over-explaining, is a prompting job, covered below.
Character card and prompt tips
- Give the threat rules, then hide them. Put the monster's nature in a lorebook entry or the author's note, and tell the model the player does not know it yet. Models are far better at withholding a secret they have been handed than at improvising one.
- Write sensory anchors into the scenario. A specific smell, a sound, a recurring object. The model will reuse them, and recurrence is what makes a place feel haunted.
- Cap reply length. 150 to 250 tokens per turn keeps the pace tight. Long replies drift into summary.
- Ban the comfort beat. Say explicitly that scenes may end unresolved and characters need not feel safe at the end of a turn.
Sample horror system prompt
You are the narrator of a slow-burn horror story. Stay in the story at all times.
Reveal the threat gradually: describe effects before causes, sounds before shapes.
Never explain what the entity is unless {{user}} discovers it in play.
Keep replies under 200 words. End many turns on an unanswered detail.
Violence and injury are described plainly and with consequence when they happen.
Do not reassure {{user}}, summarize feelings, or add closing morals.Put this in the system prompt field in SillyTavern, not inside the character card, so it survives card swaps.
Sampler settings and which model to start with
Horror wants slightly lower randomness than most genres. You want odd details, not incoherent ones. A good starting point:
- Temperature 0.8, enough variety without the narrator losing track of the house layout.
- Min-p 0.08, which trims nonsense tokens while keeping unusual word choices.
- Repetition penalty low, around 1.03. Horror relies on deliberate repetition of motifs, and a heavy penalty breaks that.
Start with outlaw-1. It is the house pick, built on GLM 5.3 Flash with the refusal direction removed, and at $0.30 input and $1.00 output per million tokens a long haunted-house campaign stays cheap. The 1M context window means the early clues are still in memory when they pay off. If you want sharper plotting for a mystery-shaped horror story, try glm-5.3-outlaw.
Common failure modes and fixes
- Monster revealed too early. Move the description out of the card and into a lorebook entry keyed to a word the player has to say first.
- Every turn escalates. Add "alternate tense turns with quiet ones" to the prompt. Dread needs troughs.
- Purple prose. Lower temperature to 0.7 and ask for plain, concrete description.
- Narrator controls your character. Add "never write {{user}}'s actions, thoughts or dialogue" and trim reply length.
FAQ
Can an uncensored model write graphic horror violence?
Yes, fiction involving violence between adult characters is allowed. The model will write injuries and deaths plainly when the scene calls for it, and you control how graphic it gets through the system prompt.
Why does my horror roleplay feel like a summary instead of a scene?
Usually the replies are too long and the temperature is too high. Cap replies around 200 tokens, lower temperature to about 0.8, and tell the model to describe effects before causes.
Does this work with JanitorAI?
Not today. Browser-only chat sites call the API from the tab and the API does not answer CORS requests. SillyTavern, which makes requests from its own server, works.