Talk to Characters in AI Stories, Not to the AI

You talk to characters in AI stories the way a novelist does: by writing the line, not by describing your intention to write it. Type the words your character actually says, in quotes and in their register, and the narrator has speech to answer with speech. Type I ask her about her sister and what comes back is a summary of a conversation you never got to hear.

• A quoted line is answered as speech; a described request is answered as summary.
• Paraphrasing your own dialogue hands the narrator the job of choosing your character’s words.
• Meta-questions addressed to the narrator pull you out of the scene and usually stall it.
• One intention per line: a spoken move carrying three questions gets one of them answered.
• Branching games hand you dialogue an author already wrote; ink’s bracket syntax exists to shorten those authored lines for the menu.

How do you talk to characters in AI stories?

Write the sentence your character speaks and nothing around it. Not I ask the ferryman where the bridge went, but “Where did the bridge go?” The second version is already fiction. It arrives in the narrator’s hands as a line of dialogue that a person in the room has just heard, and the only thing left to compose is the answer.

Most apps in this category give you two input modes, usually labeled do and say. The distinction is not cosmetic. A do move states an attempt, and the attempt happens whether or not it works. A say move states a line, and the line is spoken. Choosing the wrong one is the most common reason a scene comes back thinner than the move deserved.

What you write	Mode	What the page gives back
“Say her name.”	say	She says it, or refuses to, and the refusal is a scene
I ask her about her sister	do	A paragraph reporting that you asked
Can I ask her about her sister?	neither	An answer addressed to you; the story does not move

The middle row is the one worth staring at. It is not wrong, and a competent narrator will produce something readable from it. But you have delegated your own dialogue. Whatever your character said in that room, you will read it secondhand, in the narrator’s phrasing, and the words will not be the ones you would have chosen. The wider habit of writing moves as single specific sentences is covered in how to play AI text adventures.

Why a quoted line comes back differently

Because the narrator is completing prose, not answering a query. A line in quotation marks is a finished piece of the scene; it needs a reply, a gesture, a silence. A described request is an instruction that has to be converted into speech first, and the conversion is where your voice goes missing.

Only the top row leaves the story further along than it was. The middle row advances it too, but at the cost of your character’s voice. The bottom row does not advance it at all, and readers who spend an evening in that row often conclude the app is dull when what they have actually been doing is interviewing it.

There is a practical tell. If you can read your move aloud and it sounds like something a person in the room would say, it is a line. If it sounds like something you would type into a search box, it is not.

Which chat habits flatten a scene

Habits carried over from chat assistants are the main thing that makes an AI story feel like a chatbot with scenery. They are all reasonable behavior somewhere else. In a scene they cost you turns. The habits have a structural source, traced in why a roleplay chat and a piece of interactive fiction are built to different shapes.

• Asking permission. Can I search the body? You can. The attempt is yours to make, and in a well-built narrator an impossible move becomes a scene rather than a refusal.
• Addressing the system. Make her angrier. This is direction, not dialogue. Say something that would make her angrier instead.
• Stacking questions. Three questions in one spoken line will get one answer, usually the least interesting.
• Out-of-register politeness. Chat trains a flat, courteous voice. Characters rarely have one, and a smuggler who talks like a support ticket will be written like one.
• Correcting the narrator mid-sentence. A contradiction belongs in the story’s record, not inside a spoken line. Auditing and fixing that record is its own task, covered in what the AI narrator remembers.

Fabledrift, an interactive fiction app for Android, separates these deliberately: you write your move as a do or a say in your own words, and asking the narrator a question is a third action that does not advance the story. The page has no chat bubbles, which is a design decision before it is a visual one. Bubbles teach you to write turns; a page of prose teaches you to write lines.

What a branching game writes for you instead

In a pre-written game your dialogue is authored in advance, and the interface exists to compress it. This is not a limitation anyone is embarrassed by. It is the whole craft, and the tooling shows how carefully it is done.

Ink, inkle’s narrative scripting language, has a square-bracket syntax for exactly this problem. Its documentation states that “the square brackets in fact divide up the option content. What’s before is printed in both choice and output; what’s inside only in choice; and what’s after, only in output.” The worked example is a line of dialogue: * "I am somewhat tired[."]," I repeated. The menu shows a clean sentence; the page shows the same sentence with its attribution.

Parser games solved it differently. Threaded Conversation, an Inform 7 extension modified from Emily Short’s original and based on her game Alabaster, models talk as quips — its source defines one as “a comment for the player to make”, paired with a reply. The system offers a running list of what you could say, and accepts DISCUSS, SAY, ASK and TELL as ways to pick one.

Approach	Who writes your line	What you gain	What you give up
Branching menu	The author	Every line is deliberate and tested	You choose, never phrase
Parser topics	The author, indexed by subject	Freedom of subject, authored quality	A hidden list you must guess at
AI narrator	You	Any sentence, in your own register	Nobody vetted it before you sent it

That last cell is honest and worth sitting with. An authored line has been read, cut and rewritten before it reached you. Yours has not. The trade is that yours is yours, and the character has to deal with a sentence no author anticipated.

When not to say anything

Silence is a move, and it is often the strongest one available. A character who has just told you something enormous does not need a follow-up question; they need a beat in which nobody fills the room. Writing I say nothing and keep my hands where he can see them gives the narrator a gesture to write against, which is more than most questions supply.

The same goes for interruption, refusal and leaving. Across the two launch worlds the pressure runs in opposite directions — a forest that keeps track of what is promised beneath its branches rewards saying less, while an empty station rewards saying anything at all out loud — and both punish the reader who treats every scene as an interview.

Say it the way she would say it

The test is small enough to run in your head before every spoken move. Read the line as your character. If it sounds like a person in a room, send it. If it sounds like a request for information, rewrite it until the words are the ones she would actually use, then send that.

Everything else follows from the habit. You will stop asking permission, because the line has already been said. You will stop stacking questions, because people rarely do. And the record of your story will fill up with sentences you chose rather than sentences composed on your behalf, which is the only real reason to talk to characters in AI stories instead of reading a book where somebody else already has.