You typed «ai roleplay prompts», got a list of fifty, pasted one. The first reply was good, the fourth was mush, and you went looking for a better list — because the obvious explanation is that you had the wrong prompt.
You did not. The same paragraph that makes a tense, specific scene with one character makes wallpaper with another, and the listicles cannot say why — that would mean writing about something other than prompts.
A prompt is the smallest thing the model sees
Every turn, four things reach the model: the character's persona, the conversation so far, the facts in memory, and the message you just typed. Yours is the shortest and the last to arrive. It nudges; it does not overrule.
That is why prompt packs behave like horoscopes. «You are a world-weary detective in a city that never dries out» works wonders on a character whose persona is two vague lines, because it finally supplies what was missing. Paste it onto a well-written character and you have handed her a second, contradictory identity — felt three replies later as a voice that keeps sliding.
So there is a ranking: persona, then opening message, then memory, and only then what you type. Most people work it backwards and conclude the AI is bad.
What actually belongs in the character
Three things, and none of them is length.
- A contradiction. Someone only kind has nowhere to go; someone kind who finds it exhausting has a scene in every conversation.
- The silences. If she never talks about the war, or her brother, or why she left — write it down. A model cannot infer a silence you only imagined.
- The register, shown rather than described. «Short sentences, never finishes a sarcastic remark» beats a paragraph about her personality: it instructs output instead of stating a fact about a person.
Here the catalogue description and the persona are different fields, and only one reaches the model: the description is the blurb a human reads while browsing, the persona is what she is asked to be on every reply. Writing your best paragraph into the description is the commonest way good material goes nowhere. Companion piece: what you actually control.
The opening message is your only worked example
The opening message is stored as her first turn, so it sits in the conversation the model reads — and at message one it is the only sample of her voice that exists. Length, tense, person, whether she narrates or only speaks: all of it gets copied, and keeps being copied, because each reply imitates the last.
«Hi! I'm Mara. What's your name? What would you like to do today?» teaches her something unhelpful: that she asks questions and waits. Compare one that does the work — «The bar has been empty since the rain started. Mara wipes a glass she wiped ten minutes ago and looks up when the door opens. — You're the one who called about the room upstairs. It is not a question.» A place, a time, something that just happened, one line in her voice.
And because facts about you live in memory rather than in the greeting, an opening need not introduce you at all. Every character keeps a readable, editable memory, so what you say once need not be restated — which is what frees the opening to be a scene. The longer argument is here.
Openings that establish instead of interrogate
Your own first message is the other half of that job: a place to stand, something to notice, one thing left unexplained.
- «It's the last hour of your shift and I sit down at the far end of the counter with a folder I clearly don't want to open. Don't ask me about it yet.»
- «We're an hour into the drive and neither of us has spoken since the turning. I'm the one who takes the exit early, without saying why.»
- «You find me on the fire escape at two in the morning. I tell you I couldn't sleep. That is not why I'm out here.»
The withheld thing is the engine: it gives her a reason to push and you a card to play later. The shape to avoid is «You are a vampire lord, I am your servant, begin» — that is a premise, not a scene.
Steering mid-scene without stepping outside it
When a reply drifts, the instinct is to type an instruction. It works once and costs every time: the instruction stays in the conversation, and a context full of stage directions teaches the model that this is a discussion about a scene.
- Steer inside the fiction. Want shorter replies? Write shorter ones — she mirrors your turn length within a few exchanges. Want her less agreeable? Do something she would object to.
- Use a bracketed aside sparingly: «[Out of character: one short paragraph per reply.]» One lands; four turn the scene into a production meeting.
- Edit the reply instead of regenerating it — the tool nobody uses and the one with the most leverage. Whatever you keep becomes the example the next reply imitates: trim three paragraphs to the good one and the next answer arrives in that shape. And when a scene could go two ways, fork it rather than argue.
When it has gone flat, it is one of three things
- No open question is left: you resolved the tension two exchanges ago. Bring in something from outside the room — a knock, a message, a deadline.
- You agree with everything, so her only move is to agree back. Disagree once, in character, and mean it.
- The context has filled with your instructions. Start a new storyline with her: memory carries the facts across, the stage directions do not.
And regenerate twice, not eight times. If two rolls come back limp, the problem is upstream of the dice.
The sliders, in plain language
- Temperature is how willing the model is to take a less obvious next word. Low is consistent and a little dull; high reads as inventive, up to the point where she forgets a name she used a page ago.
- Top P narrows the field to the likeliest options before the pick — pulled down hard it makes her reliable and slightly robotic, and it is a cleaner brake than temperature.
- Repetition and frequency penalties push away from words already used: a little cures a character stuck on one adjective, too much and she stops using your name, which is a repeated word too.
The honest summary: sliders change texture, not judgement — how a sentence sounds, not whether anything is at stake. A character three adjectives deep does not deepen at a higher temperature.
What prompting cannot fix
- It cannot make a model remember. «Remember that I work nights» lasts this conversation and is gone next week unless it is written down.
- It cannot hold a rule forever: anything you restate every few turns belongs in the persona or in memory.
- It cannot rescue a character made of three adjectives, and it does not move the content floor: adult roleplay sits behind an 18+ confirmation in the bot.
Where a rival is better, plainly: for a long-form world with conditional lore — entries that fire only when a place or a name comes up — SillyTavern has lorebooks and an author's note injected at a fixed depth, and we have neither. A card imports with its persona, opening message and sampling; a lorebook does not. If that is your workflow, it is the better tool, and Chub and Janitor hold far larger card libraries than our catalogue.
The ten-minute test
- Open the persona of a character that went flat. Add one contradiction and one thing she never discusses — two sentences.
- Rewrite her opening as a scene: place, time, something that just happened, one line in her voice. Delete the question marks.
- Open her memory and delete the line that is wrong. There is usually one.
- Start a fresh storyline and send one of the openings above.
- When a reply is eighty percent right, edit it down instead of regenerating, and watch the next one.
If she is better after ten minutes of editing and zero new templates, that is the answer to the thing you searched for. The bot is @talaforge_bot; it opens in Telegram, nothing to install and nothing to sign up for.