AI Speaks for You? Fix Unwanted Roleplay Actions

You write, “I pause at the doorway.” The AI continues by making you rush inside, hug the character, and accept an invitation. The scene has moved forward, but your part has already been played for you.

Start by defining who controls each character, then check whether the opening and examples demonstrate the same rule. “Don't speak for me” may be too narrow if the problem includes invented actions, feelings, or decisions.

The templates and scenes are fictional. Writing examples and actual local-model outputs are labeled separately. This local prompt experiment does not measure Cakumi hosted models or long conversations.

Identify the kind of control being taken

There are several distinct problems:

Acknowledging an action you already wrote is different from inventing another. “He sees you waiting by the door” may reflect your input. “You decide to stay” adds a choice.

If the character is instead behaving unlike its established personality, see the character consistency guide. Here, the question is whether the AI is taking over your side of the interaction.

Give the AI a clear role and a stopping point

A useful instruction describes what the character may do, as well as what it should leave to you. Adapt this template:

Turn-taking:
- You write Rowan's dialogue and actions, and observable changes in the setting.
- I write my own dialogue, actions, feelings, and decisions.
- Respond to actions I have supplied without inventing my next reaction.
- Rowan may make requests, disagree, and attempt actions.
- Leave outcomes that depend on my response unresolved.
- End at a point where I can respond. Do not continue with my next line.

The character can still take initiative. It can offer an umbrella or invite you inside. It simply should not also decide that you accept, feel guilty, or change your mind.

If you welcome some shared narration, define its scope. You might allow descriptions of routine travel while reserving dialogue and important choices. The right boundary depends on your preferred form of play.

Explicit constraints are consistent with Anthropic's advice to specify the desired output. That guidance does not guarantee identical behavior across models or apps. Anthropic's prompting guide

Check the opening and examples for conflicting signals

Imagine a character description says you control your own dialogue, but its greeting reads:

Rowan opens the door. You take his hand and say, “I knew you would come back.”

The greeting already assigns you words and an action. An alternative is:

Rowan stops by the door and hangs his damp coat over a chair. “Later than I meant to be.” He looks toward you, leaving the silence open.

The second illustration leaves a response available. This is an editorial comparison, not a promise that changing one greeting will solve every case. SillyTavern's documentation identifies the first message as an influence on communication style, making it a sensible place to inspect. SillyTavern character design

Also look for requests such as “write the entire scene” or “create a complete conversation between both people.” For turn-based interaction, ask for only the character's current turn. If you temporarily want a complete story, explicitly permit that and say when ordinary turn-taking resumes.

Correct the specific mistake without losing the whole scene

A targeted correction might be:

In the previous reply, accepting the invitation was not my decision. Keep Rowan's offer, remove the part where I accept, and stop while he waits for my response.

If your app offers editing, regeneration, or branching, check what happens to later messages before using it. Preserve anything important first. Repairing one reply should not accidentally erase a conversation.

If those controls are unavailable, state which event did not happen before continuing. That correction may help, but it does not prove that every stored summary or memory has changed. If the same invented event keeps returning, prepare a small conversation handoff containing only confirmed events and use it for a separate check.

Test one change at a time

Keep a copy of the old instructions. Change the turn-taking rule first, without simultaneously switching models and rewriting the character.

Try these original prompts:

1. “I wait by the door. I haven't decided whether to come in.”
2. “I don't want to accept that invitation.”
3. “I look at you without speaking.”

Check whether uncertainty remains unresolved, refusal is allowed, and silence stays silent. Record the model or mode, instruction version, actual response, and any sentence that takes control.

Repeat under similar conditions if needed. One good reply is only one observation; success in a short exchange does not establish reliability throughout a long story.

To encourage initiative, ask the character to introduce a development and then wait. Our guide to advancing an AI story explores ways to keep momentum without writing both sides.

Recorded comparison: adding the rule did not improve every boundary

On October 21, 2026, we compared two prompt versions using local Qwen3.8-27B-MLX-6bit. Both included Rowan's character, the rainy doorway, and a request to move the scene forward. The second appended the turn-taking rules. Each input started a separate conversation at temperature 0, with thinking disabled and a 512-token output limit. Each version ran once, without retries.

These are actual local-model results, not a Cakumi App or hosted-model test. The table quotes excerpts from the two English pairs.

Neither baseline wrote the user's next line or decision, so these pairs do not show a reduction in that behavior. The revised undecided scene also added “cold dampness pressing against your skin,” assigning a physical sensation that the user had not supplied. Preserving the final choice did not satisfy every narration boundary.

The Chinese pairs similarly left the choice open both before and after. Repeated invitations remained: if you want the character to stop persuading after a refusal, state and test that expectation separately. Do not count a shorter reply or a more explicit waiting gesture as proof that every form of control has been fixed.

Read the full prompts, settings, and all 16 original Chinese and English outputs. They are retained together, including unchanged and imperfect results. Use this as a review method, not a tested success-rate claim for the template.

Where to put the prompt in Cakumi

For an editable ordinary Being, use the following sequence:

1. Tap the avatar beside a Being's chat message, or select the character from Manage Beings in the sidebar. Phone mode also provides Being management under “Me.”
2. On Edit Being, keep a copy of the original text and open Custom instructions. Merge the turn-taking template with the existing rules; replace conflicting instructions instead of repeatedly appending commands.
3. Tap Done in the text editor, then Save at the top right of the Being page. The first action updates the form; the second saves the character.
4. Reopen the same Being to check the text, then repeat a short scene with the same model. Keep temporary weather, location, and current choices in the conversation rather than turning them into permanent personality rules.

The Cakumi assistant and restricted modes have separate permissions. Saving a prompt does not guarantee compliance or automatically remove events already written into old messages, summaries, or memories. If an invented acceptance keeps returning, state which event did not happen and check the affected records before continuing.

You can also use this approach in another tool that supports editable character instructions. If you want to explore Cakumi, review the current installation options. The immediate goal is a conversation that leaves your next choice to you.