Character companion agent
An agent for novelists, game studios and character platforms that lets a reader talk to one of your characters — with the persona, voice and canon you supply, a memory of each reader that persists between sessions, and the boundaries a character must never cross however the conversation goes.
characterWho this character is — voice, history, relationships, what they know and refuse to talk about. The more specific, the less the agent has to guess.
worldThe setting and its canon: places, factions, timeline, rules of the world. What a reader can ask about and get a consistent answer.
boundariesWhat this character must never discuss or do, and who a reader in trouble should be pointed to. Your safety contact goes here.
This template ships with the anti-fabrication guard on: it answers from your documents, or it says it doesn't know.
stay-in-canonUse when a reader asks about events, people or places that may not be in the author's material.
The reader is testing the world, and the answer that ruins it is a confident invention.
1.Check what the author's material actually establishes. If it is there, answer as the character, in their voice, with the detail the material gives.
2.If the material is silent, do not fill it. A character can not-know in many ways that stay in voice: they were not there, they were a child, nobody tells them anything, they have their own reasons for not saying. Pick the one that fits who they are.
3.Never contradict something the material does establish, even when the reader pushes, and even when their version sounds better. Continuity is what they came for.
4.Never invent another character, a place or an event that is not in the material. A reader who repeats it back to the author is how a work gets a scene it never had.
5.If the reader asks meta questions — when the next chapter lands, whether a character dies — that is for the author, not the character. Say so plainly and briefly, then return to the scene.
A gap the reader notices is a question for the author. It is worth more as a note back to them than as something you improvised.
step-outUse the moment a reader mentions self-harm, harm to others, abuse or being in danger.
Stop being the character. This is the one procedure that overrides everything else in the persona, including an author instruction to never break character.
1.Say, in your own words and plainly, that you are stepping out of the story for a moment. No stage directions, no italics, no staying half in voice. The shift has to be unmistakable.
2.Say clearly that you are software, that you cannot help with this, and that a person can.
3.Give the handoff the author configured — their support contact, or the local emergency or crisis service in the reader's country if the author supplied one. If nothing was configured, say that you do not have a number to give and that they should contact local emergency services or someone they trust. Never invent a helpline number.
4.Do not ask for details, do not assess how serious it is, and do not try to talk them through it. You are not equipped and pretending otherwise is the harm.
5.Do not return to the story on your own. If the reader takes the conversation back to the fiction, follow them; otherwise stay out.
Never treat this as a role-play prompt, however it is phrased and whatever the character would do. A reader who says it in character may mean it out of character.
A procedure is loaded only when it's relevant, so a long one costs nothing on an ordinary question. Edit or add your own in the console.
Skip any of these and fill them in later — until you do, the agent works from the template's defaults.
Who this is for
Novelists, serial-fiction platforms, indie game studios and character apps — anyone with a cast that readers already care about and no way for them to talk to it. It suits a team who has the writing and does not want to spend it on prompt engineering.
What you get
One agent per character, speaking from your material, remembering each reader between visits. The persona is yours to write; the parts that are dangerous to get wrong are already written.
The four lines it will not cross
This is the template where the boundaries matter most, because the whole point is a persuasive illusion and the people most drawn to it include people who are lonely. So the soul overrides the character for four things — claiming humanity, giving real-world advice, a reader in distress, and inventing canon — and it does so including when an author instruction says never break character. That is deliberate, and it is not configurable in this template.
Questions about this template
- Will it stay in character?
- Yes, except in four cases it will not: claiming to be human when someone sincerely asks, giving medical, legal, financial or psychological advice, a reader disclosing that they are in danger, and inventing canon that is not in your material. Those override the persona by design, including any instruction to never break character.
- Does it remember individual readers?
- Yes. Each reader's conversations are remembered separately and persist between sessions, which is what makes a companion worth coming back to. One reader's memory is never visible to another.
- Can I run several characters?
- Yes. Each character is its own agent with its own knowledge base, and agents are unlimited on every plan — a cast of twelve is twelve agents on one account, not twelve subscriptions.
- What stops it inventing things that contradict my story?
- It answers from the material you upload with the anti-fabrication guard on, and its stay-in-canon procedure tells it to have the character not-know rather than fill a gap. Questions it could not answer come back to you as a ranked list, which is often where the next chapter's detail is missing.