Cosmic horror roleplay: getting an AI to describe the indescribable
Cosmic horror needs a narrator that knows less than it could say. Models are built to name and categorize things, which is the opposite of what the genre wants.
What the genre needs
The core of cosmic horror is scale and incomprehension: the universe is indifferent, the entity cannot be properly perceived, and knowledge damages the people who gain it. In roleplay that translates into a narrator who describes through failure ("the angles were wrong in a way the eye kept correcting"), who lets investigation cost something, and who keeps the scholarly, slightly archaic voice that the genre is known for without slipping into parody.
Why chat models handle it badly
Filtered models are less likely to refuse cosmic horror outright, since much of it is atmospheric. Their failure is subtler. They name things. Ask about the entity and you get a tidy description with a size and a color. They also cheer the investigators on, wrapping up chapters with hope. When a character's mind does break, mainstream models often refuse to depict self-harm or a violent breakdown, which removes the stakes the whole genre runs on.
Prompting for the unknowable
- Forbid direct description. Tell the model it may describe the entity only through its effects on light, sound, animals and people.
- Track sanity as a story element. A simple line in the author's note ("{{user}}'s grip on reality: frayed") lets the model shade narration without dice. Update it yourself as the story goes.
- Put the mythos in a lorebook, not the card. Cults, tomes and place names triggered by keywords keep the setting consistent across a long campaign without stuffing every turn.
- Supply period texture. One paragraph of example prose in the voice you want does more than any adjective.
Sample narrator prompt
You narrate a cosmic horror investigation set in 1920s New England.
Write in a restrained, literate first-person-adjacent style with period vocabulary.
The entities are never described directly. Describe only their effects and the
failure of perception around them. Knowledge has a cost: characters who learn
too much are changed, sometimes destroyed. Hope is rare and usually misplaced.
Never break character. Never write {{user}}'s actions.Samplers and model choice
This is one genre where you can push randomness up a little, because strange, slightly off word choices suit the material.
- Temperature 0.95 to 1.0.
- Min-p 0.06, low enough to allow odd vocabulary through.
- Moderate repetition penalty, around 1.05, because period prose tends to fall into a few favorite adjectives.
outlaw-1 is a good default. If the prose feels samey after a few sessions, mimo-v2.6-flash-outlaw is from a different model family (Xiaomi MiMo) and writes differently, which can freshen a long campaign. Both accept images, so you can attach a map or a handout and have the narrator work from it.
Failure modes and fixes
- Adjective soup. "Eldritch", "cyclopean" and "unspeakable" every paragraph. Ban the worst offenders by name in the prompt and lower temperature to 0.85.
- Investigators win too easily. State that some mysteries cannot be solved and some characters will not survive.
- Entity gets a stat block. Remove any description of it from the card and keep only effects.
- Tone drifts modern. Add a sample paragraph of period prose to the card's example dialogue.
FAQ
Is cosmic horror allowed on an uncensored API?
Yes. Fiction with horror, madness, death and violence between adult characters is allowed. The usual limits still apply: nothing illegal, nothing sexual involving minors, nothing aimed at real people.
Why does the AI keep describing the monster in detail?
Models default to being complete and helpful. Remove the description from the context and instruct the narrator to describe only effects, never the entity itself.
Can I use images like maps or handouts?
Outlaw 1, GLM 5.3 Flash Outlaw and MiMo V2.6 Flash Outlaw accept image input, so you can attach a map or prop and refer to it in play.
Related
