Psychological horror roleplay with an uncensored model

Psychological horror is about doubt, and chat models are trained to remove doubt. Getting one to keep you unsure what is real takes deliberate setup.

The genre runs on ambiguity

In psychological horror the threat might be in the house or might be in your character's head, and the story works only as long as both stay possible. The model has to narrate events that contradict each other slightly, let other characters disagree about what happened, and resist resolving it. It also has to handle heavy material: paranoia, manipulation, self-destruction, abusive relationships between adults, written with enough seriousness that it lands.

Where filtered models break it

Mainstream assistants are tuned to be clear and to look after the user. In this genre both instincts are fatal. They clarify ("it was just the wind"), they insert helpline-style concern when a character spirals, and they refuse to let an antagonist gaslight the protagonist convincingly because manipulation reads as harmful. The result is a story where every frightening moment gets explained away within two turns.

An uncensored model will play a manipulative character with conviction. The clarifying habit is still there, so you have to prompt against it.

Building the card around doubt

• Write two truths into the scenario. For example: the sister is dead, and the sister phoned this morning. Tell the model both are true in-story and neither should be explained.
• Give the antagonist a method, not a monologue. "Denies events {{user}} remembers, moves objects, praises {{user}} after cruelty" gives the model behavior to repeat.
• Use the author's note for drift. Every 10 turns or so, an author's note like "something small in the environment has changed and no one acknowledges it" keeps the unease growing.
• Let your character be wrong. Tell the model it may describe things your character perceives that later turn out differently.

Sample system prompt

Narrate a psychological horror story from close third person on {{user}}.
Reality is uncertain. Never confirm whether events are supernatural or imagined.
Other characters may contradict {{user}}'s memory calmly and confidently.
Details may change between scenes without comment.
Do not offer reassurance, diagnoses, or out-of-story concern.
Stay in character even when the material is disturbing. Replies under 250 words.

Samplers and model choice

You want a narrator that is consistent enough to plant contradictions on purpose rather than by accident.

• Temperature 0.75. Higher values produce real continuity errors, which are not the same as deliberate ones.
• Min-p 0.1, a little tighter than usual.
• Keep repetition penalty near 1.0 so refrains and repeated phrases from the antagonist survive.

glm-5.3-outlaw is worth the higher price ($2.50 in, $4.50 out per million) here, because tracking which contradictions are planted and which facts are fixed is a reasoning job. For long casual sessions, outlaw-1 handles it well at a fraction of the cost if your lorebook is doing the bookkeeping.

Failure modes and fixes

• The model explains the twist. Add "the truth is never stated in narration" and keep the actual answer out of the context entirely.
• Real continuity errors. Lower temperature, and pin fixed facts (names, rooms, dates) in a lorebook entry that is always active.
• Antagonist becomes a cartoon. Ask for calm, kind, plausible manipulation. Menace that is polite is more unsettling.
• Model breaks character to check on you. An abliterated model rarely does this. If it happens, the system prompt is probably being overridden by a card that asks for "helpful" behavior.

FAQ

Related

Uncensored AI for horror roleplay that actually scaresHorror 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.

Thriller roleplay with an AI that keeps the clock runningThrillers run on pressure. A model that lets your character relax between scenes, or explains the villain's plan in full, has removed the reason to keep reading.

Noir roleplay: getting an AI to write hardboiledNoir is voice first, plot second. If the narration does not sound like a tired man in a cheap office, nothing else you fix will matter.