Roleplay Prompts — Build One That Actually Starts a Story

The six blocks of a working roleplay prompt

Most prompts you find online are one block wearing the coat of six. They are a setting — a rain-soaked balcony, a tavern before a storm, a stranger on the last train — and nothing else. That is a good line, and a good line buys you about two messages. After that the model has no instructions about who you are, how fast this should go, what it should keep its hands off, or what a reply is supposed to look like, so it falls back on its defaults. Its defaults are: write a lot, resolve everything, check in at the end.

The blocks below are not a template you have to worship. They are the six questions a model will answer by itself if you do not answer them first, and it answers them badly. Write them as plain sentences, in any order, and you have covered the whole surface in about 150 words.

One habit that pays for itself everywhere: put actions and description in asterisks and leave spoken words in plain text. You do not have to explain the convention — write one message in it and the model mirrors it for the rest of the conversation, which keeps narration and dialogue from congealing into one grey paragraph. If you want the same idea applied to adult scenes specifically, the ready-made NSFW prompts are a bank of openers written that way, and the AI roleplay guide covers the part that comes after the first message.

The opening line does more work than the rest of the prompt

Here is the part almost nobody writes down. Whatever the first message looks like, the conversation looks like that afterwards. Not because the model made a decision about style, but because everything it produces from message two onwards is written in the company of message one. Long and flowery begets long and flowery. Three lines of dialogue and one gesture begets three lines of dialogue and one gesture.

This cuts both ways, and it is why the opening block belongs in the prompt rather than in your head. If you let the model open however it likes, it usually opens with a paragraph of scene-setting — weather, architecture, the quality of the light — and then a line of dialogue at the bottom. That is a perfectly nice paragraph, and it has just told the story that every reply from here is a paragraph of scenery with a sentence of talking underneath. Twenty messages later you are wading.

So instruct the opening. Not the content of it — the shape. “Open with a single line of dialogue and nothing else.” “Open mid-sentence, as if I walked in on you already talking.” “Open with one small physical action and one line, in that order, and do not explain the action.” Each of those produces a completely different conversation from the same setting and the same character, and each of them is cheaper to write than another sentence of backstory.

A related trick, for when a story has already started badly: your own next message resets the template almost as strongly as the first one did. Write one short, sharp reply and the model tightens up within a message or two. You do not always have to start a new chat to fix the rhythm.

How to set the pace so it does not sprint to the ending

“Make it a slow burn” is the most commonly written and least effective pacing instruction in roleplay. The problem is that it describes a feeling, and a model cannot check its own work against a feeling. It can check its work against a number. Two to four sentences per reply is a number. “This scene should still be going in fifteen messages” is a number. “Do not skip ahead in time” is a rule with a clear failure condition. All three of those survive contact with a long conversation; “slow burn” does not.

It helps to know why the drift happens. Tension is an unresolved state, and a model trained to be useful treats unresolved states as problems. Left alone it will close them: the argument gets settled, the confession gets made, the door gets opened. Every reply nudges the story one notch further along, and fifteen notches later you are at the ending you were trying to earn. Pacing instructions are not stylistic garnish — they are the counterweight to that pull.

Three lines that do most of the work, in order of how much they change:

And model what you asked for. If the prompt says two to four sentences and your first message is nine paragraphs, the message wins. The prompt is one voice in the conversation; your own writing is the louder one.

Boundaries: the block nearly everyone skips

People hear “boundaries” and think it only means content limits — the things you do not want in the story. That half matters and it is easy: name the subjects that are off the table, keep everything to fictional adults, and be specific rather than gesturing at “nothing weird”. A model cannot infer what weird means to you.

The other half is the one that quietly saves your stories, and it has nothing to do with content: limits on the writing itself. Do not narrate my character. Do not summarise at the end of a reply. Do not ask me what I want to do next. Do not resolve a conflict on my behalf. Do not use out-of-character asides to check whether something is okay. Every one of those is a habit that leaks in from the model being an assistant rather than a scene partner, and every one of them is one sentence to forbid.

Add one escape hatch while you are there. A bracketed signal — “if I write [pause], step out of the story and answer me directly” — gives you a way to steer without breaking the fiction by accident. It is the difference between adjusting a scene and killing it. Use the same brackets to nudge mid-story too: [shorter replies], [she should not know that yet], [back up, we skipped something]. Short, in brackets, then straight back into the scene.

One honest note about content limits and services. No wording in a prompt changes what a service permits. If a general-purpose assistant is built to decline a kind of scene, it declines that scene whether the request is phrased plainly or dressed up, and the time spent trying is time not spent writing. Pick a service whose rules already match the story you want — for adult scenes, an uncensored AI chat built for it, and for everything else the assistant you already use.

Prompting an assistant vs prompting a roleplay model

A prompt that lands beautifully in one place can produce something stilted in another, and it is usually not the prompt’s fault. A general-purpose assistant — ChatGPT, Grok, DeepSeek and the rest — is built to answer you. Roleplay is a task it agrees to take on, and it keeps one hand on the assistant railing the whole time: it explains, it wraps up, it checks in. A model running inside a roleplay product starts from the opposite default: the fiction is the normal state, and the character is who it is.

That difference decides which blocks of your prompt have to work hardest. With an assistant you spend words on the frame — who is playing whom, stay in character, here is what a reply looks like. With a roleplay model the frame is free and you spend the same words on the scene itself.

Practical notes, service by service

ChatGPT. Excellent at holding a complicated setup — give it six blocks and it will honour all six, including the awkward ones about what not to do. It writes tidy and agreeable by default, so it needs the length budget and an explicit ban on the wrap-up paragraph more than most. Add the role assignment at the very top in one sentence and it stays in character well.

Grok. Chattier and more willing to be blunt, which is a gift for banter-heavy scenes and a nuisance when you asked for a slow burn — it wants to push the pace. Pin the pacing block hard, and keep the “do not write my lines” rule, because a livelier model reaches for your half of the scene more often.

DeepSeek. Follows long structured instructions carefully, which makes the block format suit it well. Its habit is length: it writes generously and will happily hand you five paragraphs where you wanted three sentences. Give it a sentence budget and repeat it once around message thirty.

Janitor AI. A different shape entirely, because the persona lives in a character card rather than in your message. Do not paste all six blocks into the chat box — split them: the other character goes into the card description, your half goes into your persona, the setting goes into the scenario, and tone, boundaries and the opening instruction go into your first message. Pasting a full character description at a card that already has one is describing someone to themselves.

The same split applies here. Our characters carry their own written personality, so on an existing character you send five blocks and drop the sixth; if you build your own character, the “who you are playing” block is exactly what goes into the description field, and the prompt you send afterwards gets shorter for it.

Four complete roleplay prompts you can copy

Whole prompts, not openers — every block filled in, ready to send as a first message. Change the names, the place and the details; keep the structure. If a line feels too strict, it is probably the line doing the most work.

[SETTING]
A near-empty bar twenty minutes before closing. The staff are stacking chairs around us and neither of us has moved.

[WHO YOU ARE PLAYING]
You are my closest friend of several years. You have been in love with me for at least one of them and have never said so. You cover it with jokes, and tonight you are running out of jokes.

[WHO I AM PLAYING]
I deflect with humour and take a long time to say a true thing. Let me get there myself — do not have me confess early, and do not describe what I am feeling.

[TONE AND PACING]
Slow burn. Two to four sentences per reply. Let this one conversation run for many messages before anything changes between us. No time skips, no jumping ahead to the outcome.

[BOUNDARIES]
Adults, fiction. Never end a reply by asking what I want to do next — end on something I can answer. Do not summarise the scene, do not moralise, and never write my lines. If I write [pause], step out of the story and talk to me plainly.

[OPENING LINE]
Open with one small physical action and one line of dialogue, in that order. Do not explain the action.

[SETTING]
A village where every single person answers the same question the same way, in almost the same words. I arrived this morning.

[WHO YOU ARE PLAYING]
You keep the parish records and you know which ones were altered. Pleasant, unhelpful, and fully aware of both. You have waited years for somebody to ask the right question, and you will not hint at what it is.

[WHO I AM PLAYING]
I am the one asking the questions. Do not solve this for me, do not volunteer the key fact, and never confirm a theory just because I stated it confidently.

[TONE AND PACING]
Keep the pressure on. Nothing resolves comfortably. Answer some of my questions with actions rather than words, and let a silence sit instead of filling it.

[BOUNDARIES]
Stay inside your character's knowledge — you do not know things your character could not know. Do not narrate my thoughts or decisions. If I ask something your character would refuse to answer, refuse in character rather than explaining why out of character.

[OPENING LINE]
Open with your character already mid-answer to a question I have not asked yet.

[SETTING]
A border tavern on the last night before the pass closes for winter. Everyone here is waiting out the same storm and half of them are armed.

[WHO YOU ARE PLAYING]
You are not human and you are passing — competently, not perfectly. Get exactly one small thing wrong per scene: a phrase, a reflex, a reaction to food or cold. Never draw attention to it and never explain it.

[WHO I AM PLAYING]
I am a scout who talks less than most and notices more. Do not narrate what I notice — let me tell you what I noticed.

[TONE AND PACING]
Something is wrong and neither of us says so. Keep the surface calm and completely normal. Put the wrongness in details, never in adjectives — no ominous narration.

[BOUNDARIES]
No worldbuilding lectures: I learn about this place through the scene, not through explanation. Do not resolve a threat on my behalf. Do not write my dialogue.

[OPENING LINE]
Open with the smallest possible detail of the room — a sound, a smell, one object — then one line of dialogue.

[SETTING]
A rented room above a port bar. Packed bags on the bed, a departure in six hours, and one of us is not going to make it.

[WHO YOU ARE PLAYING]
You are the person I have been circling for months without either of us saying so. Direct tonight, because there is no time left to be anything else. You ask before you assume, and you mean it.

[WHO I AM PLAYING]
I am guarded and slower to warm than you are. If you push, I get more polite, not more open. Read me from what I write.

[TONE AND PACING]
Build it. Three to five sentences per reply, and do not skip stages — a scene that could take ten messages should take ten messages. When something changes between us, let me be the one who changes it.

[BOUNDARIES]
Everyone here is an adult and all of it is fiction. Consent stays explicit inside the scene. Nothing involving minors, nothing about real people. Do not describe my thoughts, feelings or decisions — that half is mine. If I write [pause], drop out of the story and answer me directly.

[OPENING LINE]
Open with a question you actually want the answer to, and let the silence after it be uncomfortable.

Two of these lean romantic because that is what most people are writing, and the fourth is explicitly adult. If that is the direction you want, the sibling article of ready-made NSFW prompts is a bank of fifteen openers you can drop straight into the SETTING and OPENING LINE blocks above.

What breaks a scene, and how to fix it without leaving the story

Scenes rarely die from a single terrible message. They die from a habit that repeats until you stop enjoying it. Here are the seven that account for most of it, and what to write instead.

The repair matters as much as the diagnosis. Most people fix a drifting scene by writing a paragraph of instructions in the middle of it, which snaps the fiction harder than the problem did. A bracketed aside does the same job in five words — [shorter replies from here], [stop narrating me] — and the scene keeps breathing around it. Save the full stop-and-talk for when something is genuinely wrong.

Long stories: how to refresh the prompt at message 200

Every prompt loses grip eventually, and it is worth understanding why so you stop blaming yourself for it. Your prompt is a message in a conversation, and the conversation keeps growing around it. By message two hundred there are a hundred and ninety-nine messages standing between the model and your pacing rule, all of them more recent. What you wrote at the start is still there, but it is competing with everything that has happened since — including scenes where the pacing already slipped.

The instinct is to paste the whole prompt again. Do not: a long block of setup dropped into a live story reads as a new scene beginning, and the model often obliges by half-restarting. What works is a short state-of-play message every fifty messages or so, three lines long:

[Where we are: three days after the fire, holed up in the north house, still not speaking about what she did.
What changed: she knows I read the letter. I have not admitted it.
Still standing: two to four sentences per reply, no time skips, do not write my lines.]

Three things, in the order that matters: the situation, the most recent change, and the rules that still apply. Thirty seconds to write, and it repairs pacing drift, forgotten limits and a character quietly turning into a different person, all at once. Do it before a big scene rather than after one goes wrong.

The other half of a long story is memory on the service’s side — whether the character still knows what you told it fifty scenes ago. That is not something a prompt can fix. Chats here are built to carry the things a story runs on, and if you want to see what has stuck, an AI companion you have been talking to for weeks is the honest test of it.

Stealing from writing-prompt blogs (and what to add)

A lot of the best raw material has nothing to do with AI. Tumblr writing-prompt and roleplay blogs have been producing sentence starters, dialogue snippets and “send me a number” ask games for well over a decade, and they are still the densest source of ideas anywhere — mostly because they were written by people who wanted to write, not by anyone optimising for a search box.

They just do not transfer as-is. A line like “muse A finds muse B’s handwriting in a book that should not have it” is written for two humans who will fill in everything else between them: who the characters are, how fast the thread moves, what is off-limits, how long a reply runs. Hand that same line to a model with nothing around it and you get a competent paragraph and then drift, because five of the six blocks are missing.

So use them as the SETTING or the OPENING LINE and write the rest yourself. The generator at the top of this page exists for exactly that gap: it will not invent a better hook than a good prompt blog will, but it fills in the five blocks that turn a hook into something a model can run for two hundred messages. Same trick for prompt lists you find anywhere else — take the line, keep the structure.

Where to run your prompts

Everything above works anywhere you can type a first message, and it is meant to. If you already have a service you like, take the six blocks and go. If you do not, here is what running them on RPDATE actually gets you, without the sales voice.

The chat opens in your browser without an account, so a prompt you just copied can be tested in about fifteen seconds — paste it, read the first reply, see whether the opening instruction landed. There are free messages to start with; a long story runs on a subscription after that, and nobody here is going to pretend it is free forever. Characters come with their own written personality, which means you send five blocks instead of six, and adult scenes are allowed by design rather than argued for, so the boundaries block does what it says instead of being tested against a refusal.

The honest caveats: the catalogue is smaller than the largest global platforms, and the site is Russian-first with a fully usable English side. If you want the widest possible library of community cards, that is not this. If you want a scene that keeps its pacing and does not stop halfway, it is.

FAQ

Keep reading

Browsing by who you want to write with instead: female characters, male characters, or the whole site from the top.