You have a story that is partly there. A character you can hear, a kitchen you can see, a scene that keeps rewriting itself on the walk home. What you do not have is the four thousand words between that scene and the end. The last time you asked an assistant for help, it handed back something smooth and kind and finished, every edge sanded off, the ending tied with a ribbon, and a narrator who sounded like a greeting card with a creative writing degree.
Fiction is the format where the model's defaults do the most damage. It was trained on a great deal of prose that resolves, explains, and reassures, so it pulls every draft toward resolution and sentiment unless you push the other way on every turn. It is also genuinely useful. It can hold a scene list, propose six ways into a chapter, play a character across a page of dialogue so you can hear what she will not say, and tell you where a reader stopped believing.
This guide is about keeping the division of labor right for fiction specifically. You direct the story, the point of view, and what stays unresolved. The assistant proposes. You judge sentence by sentence, because in fiction the sentence is the work. And you own it, which for a story means the reader hears one person and not the average of everyone the model has read.
What changes for this format
The method is the same everywhere. These are the places where this format bends it.
The pull toward resolution and sentiment
The model's strongest habit in fiction is to land every unit of text. A paragraph ends on a realization, a scene ends on a softening, a chapter ends on a lesson. Characters forgive in the same scene they were hurt in. The weather agrees with the mood. Grief arrives with its meaning attached. You have to name it in the brief and again in every scene request: no realization, end mid-gesture, she does not understand yet. Ask for scenes to stop one beat earlier than feels natural. Then cut the final sentence of the AI draft anyway. It is almost always the one doing the explaining.
Point of view drifts every time the text is regenerated
Close third slides into omniscient. A first-person narrator starts reporting what the other character is thinking. A limited narrator describes the expression on her own face. Each time the assistant regenerates a passage it decides the point of view again, so a draft that held in turn two can leak in turn five, in a paragraph you did not ask it to touch. Put the point of view rule in your accepted-decisions ledger as a single sentence with the limits spelled out: close third on Maren, she cannot know what Theo thinks, she can read his hands. Make a point of view check a named step in your examine pass after every bounded change, rather than something you trust yourself to notice.
Scene versus summary is your call, per beat
The model prefers summary because summary is safe. Over the following weeks they grew closer. The winter passed slowly. If you do not say which you want, it hands you a middle thing: a scene sketched in summary, which reads like a synopsis of a story rather than the story. Decide for each beat before you ask. If it is a scene, say so and give the constraints: real time, one location, dialogue allowed, no time skip, ends before the conversation does. If it is summary, say that too and give a length: three weeks in four sentences, no scene. A story is mostly the rhythm of these two modes, and that rhythm should be yours.
Dialogue tags and the stage-direction reflex
Said is invisible, which is why the model avoids it. It reaches for murmured, quipped, retorted, breathed, and attaches an adverb for safety. It also inserts a physical beat between every line, so people nod, run hands through hair, and glance at windows at a rate no human body could sustain. Ban both in the brief: tags are said or nothing, and a beat is allowed only when it changes what the line means. The model also writes dialogue in which everyone says precisely what they mean. Real dialogue is people not quite answering each other. Ask for the subtext by name: they are talking about the car, they are fighting about the money.
Keeping the unresolved thing
Every story worth finishing has something the writer is deliberately not resolving: the question the reader carries out of the last page. To the model, an open thread looks like an error, so it closes the thread, usually in a final paragraph you did not request, sometimes through a glance or an object that answers the question by implication. Name the unresolved thing in your brief as a protected decision in plain words: we never learn whether the letter was sent, do not answer it, do not hint. Then check for it specifically at the end, because the model will find a side door. A story that has lost its unresolved thing has usually also lost its reason to exist.
Voice is the whole product
In an email, voice is a courtesy. In fiction it is what the reader is buying. The model's literary register is a recognizable thing by now: lush, evenly rhythmic, fond of the weight of and something shifted and a quiet certainty. A Voice Capture session and a banned list are not optional on this path. The most reliable technique is also the least glamorous: write the first page yourself, badly if necessary, mark it mine, and make the assistant match it. The model is far better at imitating a sample in front of it than at following an adjective like spare or lyrical.
The process
- 01You
Write the story's spine by hand
Before you open a chat, write down what you already know. Who wants what. Where the story starts and in whose head. What you refuse to resolve. Names, ages, the season, anything a continuity error could later break. One messy page is enough. This is your source note, and it becomes the raw material block in the brief: the draft may not contradict it. If you do not know the ending yet, write that down as well. I do not know whether she stays is direction the assistant can work with. It means do not decide for me.
- 02You and AI
Capture your voice from your own fiction
Paste 600 to 1,000 words of your own fiction, even old work, even an abandoned draft, and run the Voice Capture mega prompt. Keep the five or six lines that describe your sentences and reject the ones that would describe anyone's. Then build the banned list in two halves: the model's tics (something shifted, a beat of silence, the weight of it all) and your own, which the model will amplify if you let it. Both halves go into the session brief. If you have no fiction of your own yet, write one page first. One bad page of yours beats a perfect description of a voice you have not got.
- 03You
Brief the session with point of view and the protected list
Use the Short Story mega prompt, or the universal brief with fiction additions: the point of view rule in one sentence, the scene list with each beat marked scene or summary, the unresolved thing as a protected decision, the tag rule, and a ban on scenes that end on a realization. Ask for questions before any draft. The questions are where you find out whether the assistant has understood your story or the synopsis of a different one.
Send this Before drafting, hold these as fixed for the whole session. Point of view: close third on [CHARACTER], who cannot know what anyone else thinks. Scene list: [BEATS, EACH MARKED SCENE OR SUMMARY]. Protected: [THE UNRESOLVED THING] is never answered or hinted at. Dialogue tags: said, or nothing. No scene ends on a realization. Use only the names, places, and facts in my notes; mark gaps [VERIFY]. Ask me up to three questions that would change the draft, then wait.
- 04AI proposes
Draft one scene at a time
Ask for one scene, never the story. Give the beat, the location, how much time has passed since the last scene, what the point of view character wants in this scene, and what she does not know. Expect the result to be roughly two-thirds usable at the sentence level and wrong at the ending. That ratio is normal. If the first draft of a scene is under half usable, do not fix it. Rewrite the scene by hand using the draft as a map of what to avoid, and ask for the next scene instead.
- 05You
Judge the sentences, then the ending
Read the scene three times, each time for one thing: point of view leaks, the ending, and tags and stage directions. Cut the final sentence of the AI draft before deciding whether it needs replacing. Usually it does not. Mark every sentence you would keep with keep so it is locked for the rest of the session. Then make one bounded request for the ending, using OLD and NEW so you can examine exactly what moved.
Send this Change only the last four sentences of this scene. The problem: it ends on [CHARACTER] understanding something, and she should not understand yet. End instead on an action or an object, mid-beat, with nothing explained. Keep every sentence above that point exactly as it is. Show OLD and NEW for the sentences you touch, then wait.
- 06AI proposes
Ask for a reader's critique, not an editor's rewrite
When a full draft exists, open a fresh chat, paste the whole story, and run the Critique Your Draft prompt with fiction questions: where did you stop believing, which character sounds like the others, where does the prose explain what the scene already showed, which sentence could be in anyone's story. Forbid rewriting. A critique in the drafting chat is contaminated by everything the assistant already decided. A fresh one reads the story as a stranger would. Decide which notes are real. Some will be the model asking for resolution again, dressed up as feedback.
- 07You
Final pass by hand, aloud
Read the whole story aloud once. Every sentence that sounds like a different writer gets rewritten by you, not sent back to the chat. Confirm the unresolved thing is still unresolved, including by implication. Check names, ages, the season, and the timeline against your spine. If you plan to submit the story anywhere, read the venue's policy on AI assistance before you do. Many magazines ask for disclosure and some decline AI-assisted work entirely, and that is their call to make.
Toolkit
The prompts that fit this format, and when to reach for each.
- Mega promptShort Story mega promptStart here for a full story session. It carries the point of view rule, the scene list, and the protected list in one brief.
- Mega promptVoice Capture mega promptRun once before the first draft, with your own fiction as the sample, and paste the result into every later brief.
- Stage promptCritique Your DraftAfter a complete draft exists, in a fresh chat, for reader-side notes. Not before, because it will critique the outline instead of the story.
- Stage promptBounded EditFor every change-only-the-ending turn. It enforces OLD and NEW so point of view cannot drift in the parts you kept.
Before and after
A paragraph the way an assistant tends to return it, and the same paragraph after a writer took it back.
Maren stood at the window, watching the rain trace silver paths down the glass. She thought about everything Theo had said, and everything he had not. Something inside her shifted. She realized, with a quiet certainty, that she had been holding on to a version of him that no longer existed. 'I think I understand now,' she murmured softly, more to herself than to him. Theo nodded, his eyes filled with a sadness that mirrored her own. Outside, the storm began to clear, and for the first time in months, she felt something like peace settle over her.
Maren stood at the window. The rain had been going since four. Behind her Theo was still at the table, turning the lid of the jam jar one way and then the other and not screwing it on. She had the sense that he was waiting for her to say the thing, and she knew what the thing was, and it was not true yet. 'Your mother called,' she said. 'Tuesday. I forgot.' The lid clicked. 'What did she want.' She did not turn around. The window gave her back a room with two people in it, the lamp, the jar, and nothing she could use.
Three changes carry the scene. The resolution is gone: the AI draft had Maren realize, understand, and find peace inside nine sentences, and the revision ends with her unable to say the thing, which is where the story actually is. Point of view is held: close third on Maren with her back to the room means the reader gets Theo's hands on the jam jar, not his sadness-filled eyes, which she could not see. And the weather stopped acting. The rain is rain with a time attached, the tags are said or nothing, and the dialogue is two people not answering each other instead of one murmured declaration.
Pitfalls
Every scene ends with a character understanding something.
Add no scene ends on a realization to the accepted-decisions ledger, ask for each scene to stop one beat early, and cut the final sentence of every AI draft by default before you read it for keeps.
The narrator knows things she cannot know, and it was fine two turns ago.
The passage was regenerated and the point of view was re-decided. Keep your own master copy outside the chat, use keep on finished passages, and re-read for knowledge leaks after every bounded change, including the sentences you did not ask it to touch.
Everyone talks in the same complete, articulate sentences.
Give each speaker a couple of habits in the brief (interrupts, never asks questions, answers a different question than the one asked). Name the subtext of the scene. Write the hardest exchange yourself and hand it over marked mine.
The prose is lush and rhythmic and not yours.
You described the voice instead of showing it. Run Voice Capture on your own pages, ban the model's literary tics by name, and make it match a paragraph of yours rather than an adjective.
Asking for scene six quietly changed scene five.
Work one scene per turn and paste only the last paragraph of the previous scene as context. Finished scenes live in your document, not in the chat, and are marked keep.
At 5,000 words you can no longer remember what was decided.
Keep the ledger: point of view, scene list, protected list, names and ages, the banned list. When rejected directions start reappearing, reset the chat with the ledger and the last finished scene, nothing else.
Questions
Can AI write fiction that does not sound like AI?
Not on its own, and not from a description of a voice. It can match a sample of your prose closely enough that the sentence-level work becomes editing instead of writing, provided you give it your pages, a banned list, and bounded turns. The parts it cannot do are the parts that make the story yours: what stays unresolved, what the narrator cannot know, which sentence to cut. Those stay with you on every turn.
Should I let the assistant draft whole chapters?
Draft scenes, not chapters. A scene is small enough to judge in one reading and to repair with one bounded change. A chapter is long enough for point of view to drift, for the model to resolve things you wanted open, and for you to stop reading closely. If you want to see the shape of a chapter, ask for a scene list rather than the prose.
Will using AI make all my stories sound the same?
It will if you let it draft from the same defaults every time. The defense is the ledger and the voice sample: a different point of view rule, a different protected list, and your own pages in the brief produce different stories. If two stories are converging, that is the signal to write the next opening page by hand before you open a chat.
Do I have to disclose AI assistance when I submit a story?
Read the venue's guidelines. They vary widely and they are the publication's decision to make. Some ask for disclosure of any AI use, some forbid it, some say nothing. If a magazine asks, answer honestly, including about editing passes. The ethics lesson on this site walks through how to describe what you actually did.