Why does AI end every reply with a question?
Updated
AI characters end on a question because chat models are trained to hand the turn back politely, and every reply that closes on "What will you do?" teaches the next to do the same. Trim the trailing questions from recent replies, give the character something they want so they end on their own move, and ask for endings on action. On UnsavedAI, a quick Edit removes the habit before it sets.
Try it now, free, no account neededStart a story with LokiThe trickster god needs an accomplice. You'll do.The question is fine; the same question every time is not
Ending on a hook is good roleplay. It is how a character hands you the turn. The trouble starts when every reply closes the same way: a recap of the scene, then "What do you think?", "Shall we go?" or the game-show "What do you do next?". It reads like a form, and it puts every decision on you, so the character feels passive however vivid the rest is.
Why the habit forms
Assistants are trained to finish by checking in with the user, and that reflex survives in fiction. The chat then reinforces it: three replies ending in a question make a fourth more likely. It is worst when your own turns end in questions too, because the conversation becomes an exchange of questions in which nobody acts.
Other ways a reply can end
| Ending | Example | Use it when |
|---|---|---|
| An action that changes things | *She pockets the key and walks out.* | The scene needs to move |
| A statement to push against | "I already told them you'd say yes." | You want conflict |
| An interruption | *The lights go out.* | The moment is stalling |
| A choice with a clock | "Door or window. Ten seconds." | A question fits, but it should be sharp |
| Silence | *He waits, and doesn't help.* | You want pressure on your character |
With Loki, the trickster god shopping for an accomplice:
Before: *Loki grins.* "So: the vault, the guards, the very angry goddess. Quite the plan. What would you like to do first?" After: *Loki grins and drops the stolen key into your palm before you can refuse it.* "Too late. You're holding it, so you're an accomplice." *Somewhere above, a horn sounds.* "That'll be the very angry goddess."
The second version still hands you the turn, but on his move, with a problem attached.
Fixes that work anywhere
- Trim the question. Edit the last line off the next few replies; most of them end well one sentence earlier.
- Give the character an agenda. A character with a goal ends on their own move. Captain Mireya Black, who has captured you and wants to recruit you, never runs short of things to push for.
- Vary your own endings. Close some turns on a statement or an action so the chat is not a string of questions.
- Paste once: (OOC: end your replies on an action, a line of dialogue or an event, not a question to me. Let your character make choices too.)
Standing rules and card fixes
For a lasting fix, put the rule where it is read every turn: the Main Prompt or post-history instructions in a SillyTavern Chat Completion preset, or the custom prompt in your JanitorAI proxy settings. A phrasing that works: "Do not end replies by asking {{user}} what they do or want. End on {{char}}'s action or words."
Then check the card. If the greeting and every example line close on a question, the model is only copying them. Rewrite one or two to end on a move. On UnsavedAI, those are Their first message and Example lines in the character builder; on house characters, editing the first few question endings is enough.
Questions people ask
Should a narrator ever ask "What do you do?"
At a real fork, yes. Game-master characters such as Grimm offer choices by design; the fix is for that to happen at decision points, not after every description.
The character asks my opinion on everything. Same fix?
Yes. Give them opinions in the card ("decides fast, rarely asks") and trim the check-in questions from a few replies.
