The character sheet: six fields, no more
- Name and role — what she does all day, not what she means to you.
- Three opinions she holds strongly, at least one you disagree with.
- Speech pattern — sentence length, formal or blunt, three words she overuses.
- What she wants in this scene — a goal independent of your character.
- What she will not do — two or three hard limits, stated up front.
- One contradiction — she says she is fine and clearly is not. This is what makes her read as a person.
Memory: three rules that keep a scene alive
- Pin the sheet. Re-send the character sheet at the top of every new session — most drift is just context loss, not model failure.
- Keep a state line. After each scene, write one line: where they are, what just happened, what she now believes. Feed it back next time.
- Never retcon in chat. If the character contradicts herself, treat it as intentional and let it become a story beat.
The escalation ladder
| Stage | What the character does | Common mistake |
|---|---|---|
| Setup | Establishes place, mood, one concrete detail | Rushing to intimacy in message two |
| Friction | Disagrees about something small | Making her agreeable — removes all tension |
| Turn | Reveals something she has been hiding | Announcing the reveal in narration instead of dialogue |
| Aftermath | Changes her behaviour based on what happened | Resetting to the default personality |
Most sessions die at friction, because people steer the character toward agreement. Agreement is comfortable and completely unmemorable. The scene only becomes worth remembering after someone in it says no.
Keeping your own voice in it
A common failure on the human side: you stop writing and start prompting. The moment your messages read like stage directions, the model starts narrating for you and the scene turns into a novel you are reading instead of a conversation you are in. Write what your character says. Let the other side carry the description.
FAQ
Q1Why does my AI character forget things?
Almost always a context-window issue rather than a model problem. Keep a compact state summary and re-send it at the start of each session instead of relying on the model to remember forty messages ago.
Q2How do I stop the AI from writing my lines?
Give an explicit instruction: describe only the character’s own actions and speech, never the user’s. Then hold the line — one slip teaches the model that it is allowed.
Q3Do longer prompts make better characters?
Up to a point. Six specific fields beat two pages of prose. Long descriptions tend to add atmosphere and no behaviour, and behaviour is what the model actually plays.
Based on desensitised conversation samples from real private-domain and customer-service work, plus public communication research. Each page is revised as reader feedback comes in. How we work