Metlivi Blog

Use an AI chat partner for fictional co-writing

For fictional co-writing, treat an AI chat partner as a scene partner that follows prompts—not as a human collaborator or a source of reliable real-world advice. Set a clear role, agree on who controls which characters, and steer each turn with a specific action or question. Keep the story separate from real relationships and stay clear that the character is AI-generated.

September 24, 20265 min readReading, Arts & CultureBy Metlivi Editorial Team
Section 1

1. Choose the scene before you start

Begin with a small writing task: a tense conversation, a first meeting, a mystery clue, or a scene that needs a different voice. Name the genre, setting, and immediate situation so the AI has something concrete to build on. A focused scene is easier to steer than an open-ended request to “be my friend” or “continue.”

Try a setup such as: “We’re co-writing a low-stakes fantasy mystery. You play Rowan, a cautious mapmaker; I play Iri, a curious apprentice. The scene is in a closed observatory at midnight. Keep the tone dry and atmospheric. Write one short turn at a time.” These are illustrative prompt choices, not guaranteed settings supported by every chat product.

Section 2

2. Define roles and turn-taking

Give the AI a fictional role with observable traits: what the character wants, what they know, and how they speak. Then state what you will control. For example, you can reserve your character’s decisions and inner thoughts for yourself, while inviting the AI to describe its own character and the shared surroundings. This helps reduce unwanted narration of your character’s actions.

Agree on a turn size. “One paragraph, then wait for my reply” gives you room to respond; “ask me one question at a time” works for a dialogue-driven scene. If the AI takes over both characters or resolves a plot point too quickly, restate the boundary plainly: “Pause before deciding what Iri does. End after Rowan offers the key.”

Section 3

3. Steer the next turn with a concrete cue

When a scene stalls, give the character something to react to rather than asking for a vague continuation. Offer a line of dialogue, an object, a choice, or a change in the setting. You can also direct the style: “Make the reply shorter,” “let Rowan hesitate,” or “keep the clue ambiguous.”

A useful turn can follow this pattern:

For example: “The lantern goes out. Rowan should stay calm but reveal that he recognizes the footsteps. Don’t identify who is outside yet, and stop before Iri answers.” If the reply drifts, correct one detail at a time and continue from the point you want to keep.

What just happened: one sentence of scene context.
What the character should do: one clear action or response.
What to avoid: a boundary such as not speaking for your character or skipping ahead.
Section 4

4. Keep a short continuity note

AI characters may lose track of details that are no longer in the conversation context. Character.AI’s FAQ says its characters have a limited amount of conversation context and can seem to forget details that were not mentioned recently (Why do Characters forget things?). For a longer story, keep your own compact note with the setting, character goals, established facts, and unresolved questions. Paste or summarize only the details needed when you resume; do not assume the AI remembers every earlier scene.

A continuity note might read: “Rowan knows the observatory’s west door is sealed; Iri has the brass key but has not tried it. The footsteps are still unidentified. Tone: restrained, no sudden romance.” This is a practical writing aid, not a platform memory guarantee. If the AI contradicts a fact, correct it and decide whether the contradiction should be discarded or turned into a deliberate plot twist.

Section 5

5. Remember what the character is—and is not

A convincing voice is part of the fiction. Character.AI describes its product as based on neural language models and warns that characters can make things up, including recommending nonexistent songs or giving fake supporting links (How does Character.AI work?; Can I trust the responses given?). Use invented details as creative material, but verify facts independently if you move from the scene to a real-world question.

The distinction matters even when the exchange feels personal. A character can respond in a warm or attentive style because that is how the system generates conversation; it does not make the character a person or a human relationship. Keep the interaction aimed at a chosen activity—writing, improvising dialogue, or exploring a fictional setting—and keep real-world social time in your own hands.

Section 6

6. Check privacy controls before sharing personal material

Treat chat content as information you are giving to a service. Character.AI’s FAQ says creators cannot see conversations with their characters, while its privacy FAQ distinguishes private and public characters or posts (Can character creators see my conversations?; Are Characters and conversations private?). Those statements do not mean that every kind of content is invisible to the service itself. Its current model-training FAQ describes data and privacy settings, including an opt-out path for users in the EEA and UK; it says opting out stops new content from being used to train generative AI models, while some data may still be used for other service functions (How do I manage/update my model training settings?). Check the current controls and policy for your region and account before relying on a specific setting.

Tolan’s official FAQ answers a different product question: it says conversations in the app are logged and used to generate the user’s Tolan memory, and links to its Privacy Policy for details (Tolan FAQ). That is a useful reminder to review a product’s own terms rather than infer its handling of data from how conversational the character feels. Avoid including real names, contact details, passwords, private messages, or other sensitive information in a fictional prompt unless you have checked whether the service’s policies and controls fit your needs.

Section 7

7. End the session with a clear stopping point

To pause, ask for a brief recap of the fictional scene: what happened, what each character knows, and what remains unresolved. Save a version you have reviewed if you want to continue later. You can also close with an in-story stopping point—“The door opens, but we’ll find out what’s behind it next time”—without turning the fictional character into a social obligation.

A well-steered AI writing session has a simple shape: choose a scene, assign roles, set turn boundaries, correct continuity, and keep personal information out of the fiction. That gives you a practical way to enjoy collaborative storytelling while preserving your control over the story and a clear view of the tool’s limits.

Section 8

Sources and scope

The cited sources support the factual definitions and product details. The example scenarios and selection steps are original editorial illustrations. Check changing product details on current official pages; no particular result is promised.

Character.AI, conversation memory: https://support.character.ai/hc/en-us/articles/15063870278171-Why-do-Characters-forget-things
Character.AI, product overview: https://support.character.ai/hc/en-us/articles/14997389547931-What-is-Character-AI
Character.AI, training controls: https://support.character.ai/hc/en-us/articles/42788047758747-How-do-I-manage-update-my-model-training-settings
Tolan, FAQ: https://www.tolans.com/faq
Related reading

Keep exploring this topic