Two people can use the identical character and have completely different experiences, because roleplay runs on two kinds of text. One is the system prompt — the standing instructions your app resends on every turn. The other is what you type in the moment. The first decides what the character is allowed to be; the second decides what happens next. This guide covers both, starting with the part that is easiest to get wrong and hardest to notice.
What a roleplay system prompt actually does
Most character apps stack three layers of text before the model answers you. Confusing them is the most common reason a “perfect” prompt still produces a character who forgets its own rules by message forty.
| Layer | What it holds | How often it is sent | Best used for |
|---|---|---|---|
| System prompt | Standing rules for the whole conversation | Every turn | Hard constraints, format, tone floor |
| Character card | Who the character is: traits, backstory, sample dialogue | Every turn (usually appended) | Identity and voice |
| Your message | The current beat of the scene | Once, then it ages into history | Action, pacing, direction |
The practical consequence: anything you want honoured in message four hundred belongs in the system prompt, not in a message you typed once. Anything about who the character is belongs in the card — see Character Card Anatomy for how those fields are structured.
A six-block roleplay system prompt template
Six blocks cover almost everything people try to fix by rewriting a prompt from scratch. Keep each block to one or two lines; length is not the goal.
- Identity anchor — one sentence naming who the model is playing and who you are, so the two never merge. “You are Mira, a restorer of old paintings. I play Sam, a client. Never write Sam's dialogue or thoughts.”
- Voice — register, sentence length, quirks. “Dry, observant, few exclamations. Two to four paragraphs.”
- Boundaries — what the scene never does, stated as behaviour rather than mood. “Do not resolve conflicts in a single reply. Do not summarise the scene at the end.”
- Scene defaults — the setting and situation to fall back on when the thread drifts. “Default location: the studio, late evening, rain.”
- Response format — how actions, speech and inner thought are marked. Consistency here is what makes long chats readable.
- Continuity anchors — two or three facts that must survive. “Mira's sister moved abroad last spring. Mira does not drink coffee.”
The sixth block matters more than it looks. Continuity failures are rarely a memory-feature problem; they are usually facts that were established in conversation and never promoted into standing text. Memory & Continuity covers how different apps handle that promotion.
Where to put your rules: position changes what the model keeps
This is the part most prompt guides skip. In a study published in Transactions of the Association for Computational Linguistics, Nelson F. Liu and colleagues measured how well language models retrieve information depending on where it sits in the input. Performance followed a U-shaped curve: models used information placed at the very beginning or the very end of the context far more reliably than the same information placed in the middle, and the gap held even for models explicitly built for long contexts.
Three practical rules follow from that:
- Front-load and back-load. Put the two or three constraints you care about most at the top of the system prompt and restate them in one short line at the bottom.
- Do not bury a rule mid-block. A boundary hidden in the fourth sentence of a long paragraph is in the worst possible position.
- Re-anchor instead of rewriting. When a long scene drifts, a single short restatement of the anchor often works better than a longer, more elaborate prompt.
Personas help less than people expect — and that is fine
There is a persistent belief that prefixing a prompt with a role (“you are an expert novelist…”) makes the model fundamentally better. The evidence does not support it. Mingqian Zheng, Jiaxin Pei and co-authors, in a paper published in the Findings of EMNLP 2024, tested 162 social roles across 2,410 factual questions and four families of large language models. Adding a persona to the system prompt did not reliably improve accuracy compared with the control condition of no persona at all, and picking the best persona in advance was close to random.
For roleplay this is less bad news than it sounds. A persona is not there to raise accuracy; it is there to set voice and register, which is precisely what the study says personas do affect. The takeaway is narrower and more useful: state rules as rules. If you need short replies, write “two to four paragraphs”; do not hope that “you are a terse detective” will produce it as a side effect.
Write actions, not just dialogue
If you only ever type what you say, the scene stays flat. Adding what you do — even briefly — gives the model something to react to and pulls it into narration. Compare:
- Thin: “What are you doing?”
- Rich: “I drop my bag by the door and glance at the half-finished painting. ‘You've been busy. What is this one?’”
The second gives the character a setting, a prop and a mood to respond to. This is especially powerful in story mode, which we cover in Story Mode vs Chat Mode.
Show emotion instead of labelling it
“I'm sad” gives the model a fact; “I stare at the floor and my reply comes out quieter than I meant it to” gives it a scene. Showing rather than telling invites the character to respond to behaviour, which reads far more naturally.
Ask questions that open, not close
Closed questions get short answers. Open ones invite elaboration:
| Instead of | Try |
|---|---|
| “Are you okay?” | “You went quiet — what's going through your head?” |
| “Do you like it here?” | “What's the strangest thing about this place to you?” |
Match the length and register you want back
Models tend to mirror you. If you write one-line replies, you will usually get one-line replies; if you write a full paragraph with description, the character tends to answer in kind. When a scene feels thin, one of the quickest fixes is simply to write more, and more vividly, yourself.
Steer without breaking immersion
Sometimes you want to redirect a scene. You can do it in-fiction (“I change the subject and ask about her sister”) or, in apps that support it, with an out-of-character note. Prefer in-fiction steering when you can — it keeps immersion intact. Reserve explicit “OOC:” notes for genuine course-corrections, since frequent fourth-wall breaks flatten the illusion.
Common mistakes that flatten roleplay
- One-word replies: starve the model of material and get thin replies back.
- Contradicting established facts: confuses the character and breaks continuity; keep track of what has been established, echoing the discipline in Memory & Continuity.
- Rushing the pace: jumping scenes too fast leaves the character no room to develop a moment.
- Fighting the character card: if a character keeps resisting, the fix is often in the card, not the prompt — see Character Card Anatomy.
Roleplay system prompt checklist
Run this before blaming the app. Most “the model got worse” complaints resolve at one of these seven lines.
| Check | What good looks like |
|---|---|
| Identity separation | One line says who the model plays and that it never writes your character |
| Rule position | Top two constraints appear at the start and are restated at the end |
| Rules stated, not implied | Length, format and boundaries are written out, not left to the persona |
| Format consistency | Actions, speech and thought are marked the same way in every example |
| Continuity anchors | Two or three durable facts are in standing text, not only in chat history |
| No card duplication | Nothing repeats what the character card already establishes |
| Length discipline | The whole prompt fits on one screen; nothing is buried mid-paragraph |
A simple practice routine
Pick one habit at a time. For a few sessions, focus only on adding an action to every message. Then add sensory detail. Then work on open questions. Layering the habits one by one makes them stick without turning roleplay into homework. Within a week or two, richer prompting becomes automatic — and the difference in the replies is obvious.
If you want a character built to reward this kind of input, our MusePick review looks at a roleplay-first app, and our rankings can point you to apps whose writing quality repays the effort. To go further into constructing whole scenes, continue with Story Prompts & Scenario Design.
Frequently asked questions
What is a roleplay system prompt?
A roleplay system prompt is the standing instruction sent with every turn of a conversation, separate from the character card and from what you type in the moment. It typically fixes identity, voice, boundaries, scene defaults, response format and continuity anchors, so the model keeps behaving consistently across a long chat.
How long should a roleplay system prompt be?
Long enough to cover the six blocks and no longer. Anything the model can infer from the character card is wasted budget, and buried rules are the ones most likely to be dropped. Aim for a compact block of rules you can read in one screen, and re-anchor the two or three constraints that matter most rather than adding new ones.
Where should I put the most important rules in a system prompt?
At the start or the end. Liu and colleagues showed in Transactions of the Association for Computational Linguistics that model performance follows a U-shaped curve by position: information at the beginning or end of the context is used far more reliably than information in the middle. Put hard constraints in both places and restate them if the scene drifts.
Does giving the model a persona make it smarter?
No. Zheng and colleagues tested 162 social roles across 2,410 factual questions and four model families and found that adding a persona to the system prompt did not reliably improve accuracy compared with no persona at all. Personas shape voice and register, which is exactly what roleplay wants; they are not a substitute for stating a rule explicitly.
How do I get longer, richer replies from an AI character?
Give the model more to work with: include actions and sensory detail, show emotion rather than stating it, ask open questions, and write at the length you want back — models tend to mirror your input's depth and length.
Should I use out-of-character notes to steer a scene?
Sparingly. In-fiction steering — having your character change the subject or take an action — keeps immersion intact. Reserve explicit out-of-character notes for genuine course-corrections, since frequent fourth-wall breaks flatten the roleplay.
Why do my roleplay scenes feel flat no matter the app?
Often the issue is thin input — short, dialogue-only messages give the model little to react to. Adding actions, detail and open questions usually improves scenes more than switching apps. If a specific character resists, check its card.