The best AI prompts for story writing aren't clever one-liners. They hand the model the things it can't guess: who is in the scene and what they're called, what happened just before, what the scene is for and where it stops, and a paragraph of your own prose to match. The eight templates below work in ChatGPT, Claude, Gemini or any model on OpenRouter, and each one is built on what Anthropic's and OpenAI's own prompting docs recommend, or on a default we counted in 1,321 scenes that AI models wrote.
A prompt that leaves a gap gets the model's default in that gap.
Why most AI story prompts produce the same story
Read the pages that rank for this query and you'll find prompts like "What if humans could communicate with animals?" (AI Stack Saturday, June 2024) or "Generate three distinct character backstories for a protagonist hiding a secret in a small coastal town" (INSIDEA, February 2026). Neither page shows what came back. Here is what tends to come back when a prompt names nothing:
| What the prompt left out | What the model fills in | Source |
|---|---|---|
| A character's name | The same few names. Asked for 34 different scenes, three times each, with no name given, GPT-6.1 Sol called a character Mara in 23 of its 102 scenes | AI Slop Index, KudoWrite Research, October 2026 |
| A character's name, across models | "Elara Voss": Amazon listed 62 books credited to an author of that name | Max Read, 7 August 2025 |
| How the prose should sound | Stock phrases: "casting long shadows" was the most frequent cliché in our count (16 times), then "first time in my life", "one last time" and "air hung thick" | AI Slop Index |
| Anything at all | Patterns that appear "over 1,000x more frequently in LLM output than human text", for some of them | Antislop paper, Paech et al., October 2025 |
The models aren't broken. They're averaging. A prompt is how you stop them averaging over every story ever written and start writing yours. The same goes for the idea itself: these horror story ideas were each changed three times to get past the first premise.
The five things every AI prompt for story writing needs
OpenAI's prompt engineering guide puts it in one line: "Give the model any additional information it might need to generate a response", and adds that this material "is usually best positioned near the end of your prompt". For fiction, "any additional information" comes down to five things:
- The people in the scene, by name, with the one fact about each that matters here (what they want, what they're hiding). A premise from a plot generator rarely has the want: add it before you paste the idea in.
- What the reader has just read: the last paragraph or two of the scene before, pasted, not summarised.
- What this scene is for and where it ends. A model with no stopping point writes until it can tie a bow on it.
- How the story is told: person, tense, whose head we're in.
- A sample of your prose. Anthropic's prompting best practices calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure". Two hundred words of your own chapter does more for voice than any list of adjectives.
The templates put the instruction first and your material last, in the order OpenAI suggests. Fill the brackets, delete what doesn't apply.
Eight AI story writing prompts to copy
1. Write a scene from a one-line beat. The workhorse. Run it six times, one beat each, and you have a 2,000-word story; our guide to writing a short story with AI has a six-beat plan to start from.
Write the next scene of my novel, about 600 words, third person past tense, close on Nadia. What happens: Nadia's landlord turns up unannounced while she is hiding an eviction letter; she lies about it, and he leaves believing her. End on the door closing, not on what she feels about it.
People in the scene: Nadia Petrova, 34, a night-shift nurse, proud, hasn't told her sister she's broke. Mr. Varga, the landlord, polite and never in a hurry.
The last paragraph before this scene: [paste it].
A page of my prose, for the voice (match its sentence length and plainness; don't copy its content): [paste 200 words].
2. Continue a chapter that stalled.
Continue from the last line below for about 400 words. Keep the scene in the same room and the same hour. The chapter needs to get to [the next thing that must happen], but not in this passage: move one step toward it. Same point of view and tense as the text. [Paste the last 1,000 words.]
The last 1,000 words aren't always enough on their own; getting AI to continue your story without drift covers the summary and bible notes to send with them.
3. Dialogue with subtext.
Write the conversation in which [A] asks [B] for [thing]. A wants [X] and won't say it. B knows and pretends not to. Neither names the real subject. Mostly dialogue, short tags, one physical action per beat at most. Stop when one of them leaves the room.
4. Rewrite a paragraph in your voice.
Rewrite the paragraph marked PASSAGE so it reads like the two paragraphs marked MINE: same rhythm, same plainness, same habit of [ending on a concrete detail]. Keep every event and every line of dialogue in PASSAGE. MINE: [paste]. PASSAGE: [paste].
5. Expand a scene that's too thin.
This scene is too short and reads like a summary. Expand it to about twice the length by slowing down the moment where [X happens]: what she notices, what she does with her hands, what she nearly says. Don't add events, people or a new ending. [Paste.]
6. Names that aren't the defaults.
Give me ten surnames for a family of fishermen in [place, decade], the kind found in local records of the time, with a one-line note on where each comes from. No invented fantasy names.
Then check the ones you like against the list in your head of names you've seen in too many AI stories. A name the model chose for you is a name it chooses for everyone.
7. Find the slack in a scene.
Read this scene as an editor. List the three places where tension drops, quoting the sentence where it drops and saying why in one line. Don't rewrite anything. [Paste.]
8. Five ways the scene could go.
Here is the situation at the end of chapter [n]: [three sentences]. Give me five different ways the next scene could go, each in two sentences, each changing what one character wants. No twist endings; I'm looking for the next step, not the last one.
If there's no situation yet, only a scene in your head, where five famous novels got their first idea shows how one image turned into a book.
Say what you want instead of listing what to ban
The tempting fix for clichés is a list: "don't use 'shiver down her spine', don't use em dashes, don't call anyone Elara". Anthropic's guide says to "Tell Claude what to do instead of what not to do", and its example swaps "Do not use markdown in your response" for "Your response should be composed of smoothly flowing prose paragraphs". The same guide adds that a reason helps: "NEVER use ellipses" works less well than saying the text will be read aloud by a text-to-speech engine that can't pronounce them, because "Claude can generalize from the explanation."
For fiction, that means "plain, concrete sentences; describe what she can see and touch" rather than a list of phrases to avoid, and "this is a quiet scene; the reader should finish it unsure whether he believed her" rather than "no melodrama".
Lists also run out. On Google's developer forum, a writer who built a list of names Gemini had already used, and told it to avoid them, reported that the list kept growing with variants like "Lyraella" and "Lyraia" (thread, July 2025). The researchers behind the Antislop work didn't solve the problem with instructions either: they built a sampler that backtracks when a banned phrase appears and a fine-tuning method that cut slop by 90%. A prompt can't do either. What a prompt can do is leave the model fewer gaps to fill.
Prompts for spicy and dark scenes
ChatGPT won't write explicit sex scenes, whatever the prompt says; why ChatGPT won't write NSFW covers what it refuses and where writers go instead. On models that will write them, the same five things matter more, not less. In our count, the two uncensored models were among the most cliché-heavy of the 13: 4.69 stock phrases per 1,000 words for Venice Uncensored and 2.95 for Cydonia 24B, against none at all for Claude Sonnet 5.5, GPT-6.1 Sol and Grok 4.7 (AI Slop Index). A permissive model given nothing to work with falls back on "air hung thick".
Two lines do most of the work in a romance prompt. State the heat level in plain words ("the scene stays on the page through the kiss and cuts at the bedroom door", or "explicit, on the page"), and state what the scene changes between the two of them, because a love scene that changes nothing reads as filler at any heat. How to write smut has the craft side.
Where pasting prompts into a chatbot falls short
- You carry the context by hand. Every scene needs the names, the last page and a voice sample pasted again, and when you change something in chapter 4 it's on you to change it in every prompt after.
- Long chats forget. Past a certain length the start of the conversation falls out of what the model can see; why AI forgets your story explains how.
- ChatGPT refuses part of the genre. For romance and dark fiction, the prompt isn't the limit; the model's rules are.
- Nothing keeps the bible. Who knows what, who's dead, what colour the car was: a chat holds none of it unless you paste it.
KudoWrite writes the prompt for you
Of the five things every prompt needs, KudoWrite gathers the first two on every beat, and your beat supplies the third. You write the beat as a line, a rough idea or a full passage, the way template 1's opening sentence is written. Before anything is drafted, the engine plans a brief from the story bible entries for the people in it and whatever else the passage needs, the story so far, the chapter's current moment and the beats before and after. No pasting. If the draft misses, a one-line steer edits the beat in place, and the steer outranks everything else in the brief.
The bible and the whole-book summary keep up as you write, so the next brief starts from what you kept rather than from what you remembered to paste. There's no content filter of KudoWrite's own, so the heat line from the romance prompts above is yours to set, within the law. You get one free credit to try it after you confirm your email, then one-off packs from $3.50 for 100 credits, no subscription.
The next step
Take the scene you're stuck on and write its beat the way template 1 is written: who's there, what happens, where it ends. Paste it into the chatbot you use, with your last page and your own paragraph for the voice, and compare it with what your usual prompt gets. Then put the same beat into KudoWrite, where the names, the last page and the bible go in with it on their own. If the result reads like AI, AI writing that doesn't sound like AI is the editing pass.