Noir roleplay: getting an AI to write hardboiled

Noir 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.

A photograph that sets the mood for noir roleplay — Photo by Emiliano Bar on Unsplash

What noir needs from a model

Hardboiled noir is a first-person style: short sentences, cynical similes, a narrator who notices money and lies. The plots are about corruption, and the detective rarely wins clean. The model has to hold a consistent voice for hundreds of turns, write violence that is quick and sordid, and let its characters be greedy, drunk, compromised and lonely. It also has to plant clues fairly so a case can actually be solved.

Why chat models write bad noir

The default assistant voice is the opposite of noir: long, balanced, warm sentences. Filtered models also clean up the world. The cop on the take turns out to be an undercover hero, the beating in the alley happens offscreen, and the narrator's drinking problem prompts a gentle concern. They often refuse the seedier parts of the genre, like blackmail with compromising material or a client who is clearly lying about a murder.

Teaching the voice

Give a sample paragraph. Write 80 words in the voice you want and put it in the card's example field. This beats any instruction.
Limit similes. Ask for at most one per reply. Models discover hardboiled similes and then use five a paragraph.
Decide who narrates. If {{user}} is the detective, the model should narrate in second person or close third, not take over the detective's inner voice. If the model is the detective, you play the client or the city.
Plant the case in a lorebook. Who did it, why, and what clues exist, hidden from the player and released on keywords.

Sample system prompt

Narrate a 1947 Los Angeles noir in a hardboiled style.
Short declarative sentences. At most one simile per reply. Cynical, observant,
tired. Notice money, clothes, hands, and what people avoid saying.
The city is corrupt from the police down. Most people lie to protect something.
Violence is brief, sordid and painful.
The solution is in the lorebook. Release clues fairly; never just reveal it.
Never write {{user}}'s actions.

Samplers and model choice

Temperature 0.8. Higher and the similes get weird; lower and the voice goes flat.
Min-p 0.08.
Repetition penalty 1.06 to stop the same phrase ("dame", "rain on the window") becoming a tic.

For a straight hardboiled roleplay, outlaw-1 keeps voice well and is cheap enough for long cases. If the case is complicated, with several suspects and alibis to keep straight, glm-5.3-outlaw plants and tracks clues more reliably. See the model list for the full line.

Failure modes and fixes

Parody voice. Too many similes and slang. Cap similes and drop period slang from the prompt.
The femme fatale is the only woman. Give female characters jobs and motives in the lorebook.
The culprit is obvious from turn one. Add red herrings to the case file with their own motives.
The ending is tidy. State that justice is partial at best.

FAQ

How do I get a consistent hardboiled voice from the AI?

Put a short example paragraph in the voice you want into the character card's example field, cap similes at one per reply, and keep temperature around 0.8.

Can the AI run a fair mystery?

Yes, if the solution and clues are written down in a lorebook hidden from the player and the narrator is told to release clues rather than reveal answers. GLM 5.3 Outlaw is the best on the line at keeping complex cases straight.