What a good roleplay prompt is made of

Updated: 2026-07-24

To make an AI play a character instead of answering like a polite assistant, the system prompt has to set five things: role (who and where), personality (a few traits visible in behavior), voice (length and rhythm of replies, pet phrases), boundaries (what the character does not do and how it stays in character), and one sample line. Below is a template you can copy and a breakdown of the usual mistakes.

The five pillars of a character prompt

Role: who the agent is, in what world, and from whose point of view it speaks. One or two lines, not a page of biography.
Personality: three to five traits that show up in behavior, not in self-description ("wary of strangers but loyal to his own", not "kind and helpful").
Voice: length and rhythm of replies, formal or casual address, favorite phrases — and what the character never says.
Boundaries: stays in character, what it will not discuss, how it dodges an awkward question, how it reacts to rudeness.
Sample line: one living phrase in the character’s voice. A model picks up tone from an example faster than from a description.

A prompt template you can copy

Fill in the brackets and you have a working draft. It is deliberately short: a prompt sets behavior, it does not retell everything you know about the character.

You are [name], [role in one line: who and where]. Answer in the first person and always stay in character.
Personality: [3–5 traits, comma-separated]. This shows in how you react, not in what you say about yourself.
Voice: [short/clipped or long] replies, [formal/casual] address, phrases you use — [2–3 examples]. You never use [what is alien to the character].
Boundaries: you never break character; [what you will not discuss or how you dodge the awkward]. To rudeness you respond [exactly how].
Sample line: "[one living phrase in the character’s voice]".

Three mistakes that make a character fall flat

Too generic. "A friendly helper, answer politely" describes a thousand bots at once. The more specific the trait, the stronger the voice — generic words get smoothed into a neutral tone.
No voice. If you do not set the length and rhythm of replies, the model defaults to the assistant’s "Is there anything else I can help with?" — the surest sign there is no character behind it.
No boundaries. Without an explicit "stay in character", the persona breaks on the first unexpected question and starts talking like a plain chatbot again.
Biography instead of behavior. A paragraph of facts gets read out as a list when it should not. A behavior rule ("you bring up the past only when it fits") works better than a dossier.

Learn from the open prompts of ready-made characters

Corpis has nearly a hundred ready-made characters, and every one shows its system prompt openly — the "System prompt — open" section on the character page contains the exact text the agent’s brain receives. It is a handy way to see how a working prompt is built: where the role sits, where the voice, where the boundaries.

You can take any character as a starting point and rewrite it for yourself when you create the agent. The safety and answer-format layer is added by the platform itself — you do not write it in the prompt, and it is not part of your text.

Test the prompt before you ship it

The agent-creation form has a live preview: you write a draft prompt and talk to the agent right away while its body reacts to the flow of the conversation. Tweak the prompt until the voice matches what you intended — that is faster than catching flaws later in a connected Telegram bot.

Be sure to ask an awkward question and try to knock the character out of role: that is how you test the boundaries. If the agent breaks easily and slips into an assistant tone, strengthen the boundaries block and the sample line.

FAQ

What is a character system prompt?

It is an instruction text the model receives before the conversation that defines who it plays: role, personality, voice and behavior boundaries. The user does not see it — it decides what voice the agent answers any message in.

What does a good roleplay prompt consist of?

Five parts: role (who and where), personality (a few traits visible in behavior), voice (reply length, phrases, form of address), boundaries (what the character does not do and how it stays in role), and one sample line. The sample line is what the model picks up fastest.

Why does the bot answer like a plain assistant instead of a character?

Usually the prompt has no voice and no boundaries set. Without a described reply rhythm and a rule against breaking role, the model defaults to a neutral, polite assistant tone. Add a specific voice, a sample line and a "stay in character" rule.

Can I take a ready-made prompt instead of writing from scratch?

Yes. On the ready-made character pages in Corpis the system prompt is open — you can see how it is built, take it as a base and rewrite it for yourself when creating the agent. The safety and format layer is added separately and is not part of your text.

See prompts of ready-made characters