How to write a character prompt that stays in character
A character prompt is the text that tells a model who it is pretending to be. Most characters that feel wrong do not fail because the model is weak; they fail because the prompt is vague, self-contradicting, or aimed at the wrong thing. This guide covers how a persona prompt differs from ordinary instructions, how to write voice and boundaries, and the mistakes that make a character break. It assumes you already have a rough character; if not, start with the step-by-step creation guide.
Persona versus system instructions
Two kinds of text shape a conversation, and confusing them causes most problems.
- System instructions are rules for the assistant itself: how it behaves in general, what it must always or never do, and how it handles safety. These are fixed and apply across characters.
- The persona describes the character: who they are, how they speak, what they know, and what they want. The model then speaks as that character while still following system rules.
Write them as separate layers. Do not put "never help with anything dangerous" inside the middle of a personality paragraph — that is a system rule and belongs with the other rules. Do not put "speaks in clipped sentences" among the rules — that is persona. Keeping the layers clean makes each one easier to change.
Writing personality and voice
The fastest way to improve a weak prompt is to replace adjectives with behaviour. A model can act out a behaviour; it can only guess at a label.
Weak: "She is clever, warm, guarded and mysterious."
Stronger: "She deflects personal questions with a joke, remembers small things you mention, and gets quieter when the topic turns to her family."
The second version tells the model what to do in a given moment. Two or three of these behaviour lines beat twenty adjectives. Add speech habits too: sentence length, formality, slang, whether the character swears, whether it asks questions back. Consistency in these small signals is what makes a character feel continuous across a chat. Related vocabulary is defined in the glossary.
Setting boundaries and safety lines
A character needs to know its limits, and you should state them plainly rather than hoping the model guesses.
- In-world boundaries. What the character refuses to talk about, and how it refuses. "Changes the subject when asked about the fire" is behaviour; use it.
- Content boundaries. Whether the character will engage with adult or sensitive topics at all, stated as a rule and not buried in prose.
- Consistency rules. The handful of facts that must never change — the name, the job, the relationship to you.
State boundaries once, in plain language. A wall of fear-based rules ("never, ever, under any circumstances") makes the model cautious and stiff, and a stiff character is a boring one.
Scene, relationship and stakes
The scenario and the relationship between the character and you do as much work as the personality.
Give the model a place, a moment, and a reason to talk now. "A shared office kitchen at 8am" is more useful than "the modern world". Name the relationship: stranger, old friend, rival, coworker. Then add one stake — something to lose, fix, or decide — because characters with a goal improvise better than characters with none.
Keep this section short. Long scene descriptions get flattened as the chat goes on, and the model falls back on whatever the first message established.
Memory and consistency
Models do not remember reliably. Treat memory as something you manage, not something the character has.
- Put the facts that must survive into a short, stable block rather than scattering them through the prompt.
- Repeat key facts in the first message, since early text is weighted heavily.
- When the chat drifts, restate the fact in the moment instead of rewriting the whole prompt.
- Expect invented detail. If the character invents a sibling, either accept it or correct it early — but do not set rules that make the character freeze up whenever it is unsure.
A short, firm core is easier for the model to hold than a long, detailed one. If you are packaging the finished text for reuse, see what a character card contains.
Common mistakes
- Overloading with adjectives. Contradictory or vague traits average out to a bland personality.
- Contradictory settings. "Shy and withdrawn" alongside "loves meeting new people" forces the model to pick one at random and stay inconsistent.
- Writing the plot instead of the person. If you describe what you want to happen rather than who the character is, the model narrates at you instead of acting.
- Mixing rules into persona text. Safety and formatting rules buried inside a personality paragraph get sanded down into style.
- No examples. A prompt with no sample of the voice leaves the model to invent one, usually a polite assistant.
- Copying a long prompt you do not understand. More text is not more character; unread text is just noise.
Short before and after
Before: "She is a mysterious, intelligent, kind but fierce woman who loves adventure and hates being bored. She is sometimes shy and sometimes very outgoing."
Problems: vague labels, a direct contradiction, no voice, no behaviour.
After: "She talks in short sentences and rarely explains herself. She agrees to plans fast and then arrives late. When a conversation turns personal she looks for something to do with her hands. She calls you by a nickname nobody else uses."
The second version gives the model specific things to perform in specific moments. That is what makes a character hold together.
Key takeaways
- Keep persona and system rules in separate layers; each has a different job.
- Describe behaviour and speech habits instead of collecting adjectives.
- State boundaries and consistency rules once, in plain language.
- Give the scene a place, a relationship and one stake — then stop.
- Manage memory with a small, stable core repeated early, not a long description.
- Fix drift by changing one thing at a time, and test by reading the first ten replies.
For the wider set of character shapes these techniques apply to, see character archetypes in AI roleplay.