Most roleplay sessions die in the first two exchanges, and it is usually the opening's fault rather than the character's. A character dropped into a featureless void with "hi" has no setting, no stakes and no direction, so it produces bland, hedging replies. A well-built scenario front-loads exactly enough structure to create momentum, then gets out of the way. This page is about that opening block — the scenario prompt. For turn-by-turn craft once the scene is running, see Writing Better Roleplay Prompts.
What role-play research says about openings
Scenario advice is usually folklore. It does not have to be: the academic role-play literature published through 2025–2026 measures the exact failures that a weak opening produces, and the fixes it proposes are things you can write into a prompt yourself. Three results are worth knowing before you write another opening.
| Study | Scale / method | Finding | What it means for your prompt |
|---|---|---|---|
| Rule-based role prompting, CPDC 2025 challenge (arXiv 2509.00482) | Four prompting approaches compared on a tool-augmented role-play agent | A structured "character-card / scene-contract" prompt scored 0.571 against a 0.519 zero-shot baseline — and beat automatic prompt optimisation | Writing the scene as an explicit contract outperformed letting an optimiser generate the prompt. Structure beats cleverness. |
| Memory-Driven Role-Playing / MREval (arXiv 2603.19313, March 2026) | Diagnostic across 12 LLMs, bilingual benchmark | Models "frequently fail to recall and accurately apply their designated persona knowledge without explicit cues"; the framework splits the skill into anchoring, recalling, bounding and enacting | Persona details you wrote once, elsewhere, are not reliably in play. Re-anchor the ones the scene depends on, in the opening. |
| Perspective-bounded memory / REVERIEMEM (arXiv 2606.25632, June 2026) | KBF-QA benchmark, approximately 4,386 questions over eight novels | Names two failure modes — factual overreach (a character using facts outside its perspective) and stylistic monotony; tagging facts by visibility improved knowledge-boundary fidelity by 34.6 percentage points | Say what the character does not know, and give the voice a situational anchor. Both are prompt-level fixes. |
None of this requires you to run a benchmark. It just tells you which slots in an opening are load-bearing — and the two that most people leave empty are the two that research keeps flagging.
The six slots of a scenario prompt
Setting
Where and when does this take place? Even a sentence — "a rain-soaked night market, stalls closing up" — gives the character sensory material and a mood. Setting is the cheapest way to lift a scene out of the void.
Situation
What is happening, and why does it matter? Scenes need a little tension or a goal: a problem to solve, a decision looming, a reunion, a deadline. Without stakes, there is nothing to drive the exchange forward.
Roles
Who are you to each other? Strangers, old friends, rivals, colleagues? The relationship frames how the character treats you and gives the model an immediate stance. This connects to the context field of the character card.
Knowledge boundary
State what the character does not know. This is the slot almost nobody writes, and it is the one that causes the most immersion damage: your character casually referencing your name, your job or last week's plot twist before you have told them. Role-play researchers call this failure factual overreach, and the perspective-bounded memory work found that tagging facts by who can see them raised knowledge-boundary fidelity by 34.6 percentage points over the strongest prior method. In a prompt it costs one line: "She does not know I was the one who sent the letter."
Voice anchor
Give the character one concrete speech habit for this scene rather than an adjective. "Curt, answers questions with questions, never says the other person's name" does more work than "mysterious." The same research names stylistic monotony — flattening a character into a single fixed voice — as the second recurring failure, and a situational anchor is the cheapest counterweight. Anchor the voice to the scene, not to the character's whole biography.
Open thread
Leave one deliberately unresolved element — a question hanging, a choice unmade, a mystery unexplained. This is the hook the story pulls on. A scenario that resolves everything up front has nowhere to go.
A reusable scenario template
Six lines, one per slot. Fill them in and you have a scenario prompt; skip the fourth and fifth and you have the version most people write.
| Slot | Prompt to yourself | Example |
|---|---|---|
| Setting | Where and when? | A night train, somewhere past midnight |
| Situation | What is at stake? | You have both missed your connection |
| Roles | Who are you to each other? | Strangers sharing a compartment |
| Knowledge boundary | What must they not know yet? | He does not know I recognise him |
| Voice anchor | One concrete speech habit? | Deflects every personal question with a joke |
| Open thread | What is unresolved? | He is clearly running from something |
Those six lines are enough to launch hours of roleplay, because they give the character somewhere to stand, something to want, and edges it cannot walk through.
Six roleplay prompts you can copy and adapt
These are scenario openings, not character cards — paste one into the scenario or first-message field, swap the specifics, and let the character respond. Each follows the six slots above, so you can see the structure working.
- Slow-burn drama. "Late at a station café, the last train gone. We were colleagues on a project that ended badly two years ago and have not spoken since. You do not know that I asked to be moved off that project, not fired from it. You speak in short sentences and change the subject when it gets personal. You are holding a folder you have not explained."
- Investigation. "A coastal town in off-season, a week of rain. I am a records clerk; you run the harbour office and have been asked to help me check a shipping log. You do not know which entry I am actually looking for. You are pedantic about dates and dismissive of guesswork. One page of the log has been cut out."
- Fantasy quest. "A mountain pass at first snow. You are a guide who has crossed it forty times; I hired you this morning and told you almost nothing. You do not know that I have made this crossing before. You narrate terrain in detail and people not at all. Something has been following our tracks since the second ridge."
- Slice of life. "A shared kitchen at 2am in a building where nobody knows anyone. We are neighbours who have nodded in the hall for a year. You do not know I moved in after leaving a job in the same field as yours. You talk mostly about food and deflect anything else. Your rent notice is on the counter, unopened."
- Science fiction. "A survey station eleven months into a fourteen-month rotation. You handle systems; I handle sampling. You do not know the resupply window has slipped. You answer with readings before opinions and rarely use my name. The station has started logging events neither of us initiated."
- Comedy of errors. "A wedding reception where I have been seated at the wrong table. You are the bride's cousin and assume I am a colleague of the groom. You do not know I have never met either of them. You are relentlessly enthusiastic and never let a silence sit. Someone is about to ask me to give a speech."
Notice that none of them resolve anything. Each ends on the open thread, which is the line that tells the character where to push.
Do you need a roleplay prompt generator?
Prompt generators are useful for one thing: breaking a blank page. They produce a setting and a situation, which are the two easiest slots to fill yourself, and they almost never produce a knowledge boundary or a voice anchor, because those depend on what you intend to do in the scene.
There is a relevant data point here. In the CPDC 2025 role-play challenge, researchers compared four prompting approaches including automatic prompt optimisation — an algorithm that rewrites the prompt for you — against a hand-built rule-based approach using an explicit character-card and scene-contract structure. The structured hand-built version won, scoring approximately 0.571 against a 0.519 zero-shot baseline, and the authors report it improved reliability compared with the more elaborate optimisation method. A generated opening is a starting draft. The slots you add afterwards are what makes it hold.
A practical compromise: generate for setting and situation when you are stuck, then write the last three slots yourself before you send.
Novel-style scenarios for story mode
Story-mode and novel-style tools want a slightly different shape from a chat opening. Instead of addressing the character directly, you write a short passage in third person and stop mid-scene, which signals the narrative register you want continued. The six slots still apply — they are just embedded in prose rather than listed.
Two adjustments matter. First, keep the passage short; a long opening burns context and, per the memory-driven role-play diagnostic, persona details still need explicit cues to stay in play, so a wall of scene-setting does not protect them. Second, put the knowledge boundary in narration rather than dialogue — "she had not yet been told who was waiting" reads as prose, while "you don't know who is waiting" breaks the register. If you are choosing between the two formats, Story Mode vs Chat Mode covers the trade-offs.
Common scenario mistakes
- The blank room: no setting, no situation — the character has nothing to react to.
- Over-scripting: dictating exactly what the character thinks and does removes their agency and flattens the collaboration.
- Pre-resolved plot: a scenario with the conflict already settled leaves no forward motion.
- Too many threads at once: a dozen dangling mysteries overwhelm the scene; one strong hook beats five weak ones.
- No knowledge boundary: the character knows everything you know, so nothing can be revealed and every reveal lands flat.
- Adjectives instead of habits: "cold and mysterious" gives the model nothing to execute; a speech habit does.
- Assuming the character card is enough: details written elsewhere are not reliably recalled without cues, so re-anchor the two or three the scene depends on.
Escalation: building an arc
Once a scene has momentum, shape an arc by introducing complications gradually rather than all at once — a new obstacle, a revelation, a change in stakes. Pacing matters: give each beat room to breathe before the next. Knowledge boundaries are also what make escalation land, because a reveal only works if the character genuinely did not have the fact a moment earlier. If you are running a longer story across sessions, this is where memory becomes decisive, since the character needs to recall earlier beats; see Memory & Continuity. For richer settings that persist, combine this with World-Building for AI Roleplay.
Matching scenario to mode
Scenarios pay off most in story mode, where the app narrates the world around your prompt; in plain chat they still help but the descriptive payoff is smaller. For multi-character scenes, the same six slots apply, but each character needs its own role, knowledge boundary and voice anchor — which is precisely why group scenes drift faster, as covered in Group & Multi-Character Roleplay.
A 60-second scenario check
Before you send an opening, run these seven questions. Any "no" is a line you can add in a few seconds.
- Can I picture where this takes place and roughly when?
- Is something at stake, unresolved or about to happen?
- Is the relationship between us stated, not implied?
- Have I written at least one thing the character does not know?
- Is the voice described as a habit rather than an adjective?
- Is there exactly one open thread, not five?
- Have I left the character something to decide in their first reply?
Start simple, then build
You do not need an elaborate world to begin — the six-line template is enough, and four of the six lines are ones you would have written anyway. Run it, see where the character takes it, and add complications as the story earns them. To try this with an app built for narrative roleplay, our MusePick review is a useful reference, and our rankings highlight apps with strong story-mode support.
Related on CompanionRank
Writing Better Roleplay Prompts World-Building for AI Roleplay Character Card Anatomy MusePick ReviewFrequently asked questions
What should an AI roleplay prompt include?
Six slots: a setting, a situation with tension, the roles you hold toward each other, a knowledge boundary stating what the character does not know, a voice anchor describing one concrete speech habit, and a single open thread left unresolved. The last three are the ones most openings skip.
How do I start a roleplay so it doesn't stall immediately?
Give the opening a setting, a situation with some tension, defined roles, and one unresolved thread to pull on. A character dropped into a blank room with no stakes produces bland replies; a scenario with direction gives it something to do.
Why does my AI character know things I never told it?
Because nothing in the prompt marked those facts as out of view. Role-play research calls this factual overreach, and work on perspective-bounded memory found that tagging facts by who can see them improved knowledge-boundary fidelity by 34.6 percentage points. In practice, one line stating what the character does not know fixes most of it.
Are AI roleplay prompt generators worth using?
They are useful for breaking a blank page, but they mostly produce settings and situations — the two easiest slots. In the CPDC 2025 challenge a hand-built rule-based role prompt scored approximately 0.571 against a 0.519 zero-shot baseline and outperformed automatic prompt optimisation, so treat a generated opening as a draft and add the knowledge boundary and voice anchor yourself.
Can I over-plan a roleplay scenario?
Yes. Over-scripting — dictating exactly what the character thinks and does — removes their agency and flattens the collaboration. Seed a hook and enough structure for momentum, then leave room for the character to surprise you.
Do scenarios matter more in story mode or chat mode?
They help in both but pay off most in story mode, where the app narrates the world around your setup. In plain chat the descriptive payoff is smaller, though clear roles and stakes still improve the exchange.