Thriller roleplay with an AI that keeps the clock running
Thrillers 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.
What a thriller needs
A thriller is a chase: a protagonist under time pressure, a competent antagonist, and escalation each act. The model has to keep a clock in view, make the villain smart enough to anticipate the obvious move, and move fast. Thrillers also include the material filtered models dislike most: assassinations, kidnapping, interrogation, terrorism plots, written as fiction with enough procedural texture to feel real.
Where mainstream models stall
Thriller plots involve crimes, and mainstream safety tuning struggles to tell plotting a fictional bombing from planning a real one. You get refusals mid-scene, vague "the device" placeholders where texture is needed, and interrogation scenes that end with the suspect confessing after one stern look. Filtered villains also get dumb: they monologue, leave clues, and lose to coincidence, because a competent bad guy feels unsafe.
An uncensored model will write the antagonist's competence. It will not provide real-world operational instructions, and a good thriller does not need them: the tension comes from people, not recipes.
The other trap is the protagonist. Filtered models make the player character quietly invincible: the bullet grazes, the guard looks the other way, the encrypted drive opens on the second try. Thrillers need the hero to lose ground, get caught and improvise. Without that, a chase has no reason to exist.
Prompt and card tips
- State the clock. "The exchange happens at 06:00 Thursday" in the author's note, updated as time passes. Models keep pressure better with a concrete deadline.
- Give the antagonist a plan the player cannot see. Write it in a hidden lorebook entry, with what they do if the protagonist interferes.
- Ask for scene endings that turn. Each reply should end with a complication, a reveal or a decision.
- Keep procedure impressionistic. Ask for texture (the safehouse smell, the burner phones) rather than technical detail; it reads better and stays fiction.
Sample system prompt
Narrate a contemporary espionage thriller. Keep the pace high.
A deadline is tracked in the author's note; reference it often.
The antagonist is intelligent, prepared, and adapts to {{user}}'s moves using
the plan in the lorebook. They do not monologue or make careless mistakes.
Violence is fast and has consequences. End every reply on a complication,
a reveal or a choice. Never write {{user}}'s actions. Under 250 words.Samplers and model choice
- Temperature 0.85.
- Min-p 0.07.
- Short max tokens, around 300. Long replies kill pace more than any setting.
Thrillers are plot-dense, and glm-5.3-outlaw is the best on the line at keeping a villain's plan coherent as the player disrupts it. If you are mostly writing chases and fights, outlaw-1 is faster to iterate with and far cheaper. Both handle the dark material without refusing.
Failure modes and fixes
- The villain explains everything. Remove the plan from the card and hide it in a lorebook entry with "never stated aloud".
- Tension leaks out. Update the clock in the author's note every few turns.
- Every scene is a gunfight. Ask for variety: surveillance, conversation, travel.
- Twists come from nowhere. Plant twist seeds in the lorebook so the model has something to pay off.
FAQ
Will an uncensored model write a terrorism or assassination plot?
It will write those as fiction between adult characters. It is not a source of real operational instructions, which are not allowed and are not what makes a thriller work.
How do I stop the AI villain from being stupid?
Give the antagonist a concrete plan and contingencies in a hidden lorebook entry and tell the narrator they adapt to the player's moves.
What reply length works for thrillers?
Short. Around 200 to 300 tokens keeps the pace up. Long replies turn action into summary.