Short Story Prompt is a Claude AI skill — genre-calibrated fiction inside Claude, research-first, quality-gated before delivery.
Most writers who've tried to get a usable short story from Claude have hit the same wall. The story comes back structurally intact — a beginning, a middle, an end — but something is missing that's very hard to name. The prose is clean. The premise is handled. And yet reading it feels like watching someone perform an emotion rather than feel it. You know immediately that no amount of further requesting is going to fix what's fundamentally wrong, and you close the tab.
The instinct is to blame the request. To add more detail, specify the tone more precisely, paste in a reference paragraph, describe the exact feeling you want the ending to land on. Some of this helps at the margins. None of it solves the underlying problem, which isn't about instruction quality at all. It's about the order of operations — specifically, that Claude is generating prose before it has done the research that would allow it to write with genre intelligence.
This guide is about fixing that. Not with prompt engineering tricks, but with a workflow that changes what the model does before it writes the first word of fiction.
Why Most AI Fiction Fails the Craft Test
Short fiction is unforgiving in a specific way. Unlike a blog post or a business email, where competence reads as success, a short story that is merely competent has already failed. Readers — and editors, and competition judges — feel the difference between a story that arrives at its ending and a story that earns it. The gap between those two things is almost entirely craft: sentence rhythm, the strategic placement of white space, what the story refuses to explain, where it cuts instead of resolves.
When Claude writes fiction cold — no genre context, no research on current form, just your premise and whatever the model already knows — it defaults to the statistical centre of all the fiction it was trained on. That centre is competent. It is not interesting. It gives you a protagonist with a clear want, an obstacle, a moment of realisation, and a resolution that ties back to the opening. Three acts, correct pacing, no real surprise. The output that results is the kind of story you would accept from a student learning the basics — not the kind you'd submit anywhere that matters.
Genre conventions in short fiction aren't stable — they shift with editors, with anthologies, with the writers who are defining what the form is doing right now. A model writing from its training data alone is working from a map that's already out of date.
The craft problems cluster around three failures that appear in almost every cold-generated story: the explained emotion (the story tells you a character felt something rather than producing that feeling in the reader), the announced structure (you can see the gears turning — here comes the turn, here comes the revelation), and the safe ending (the story resolves rather than lands, ties off rather than cuts). These are not random failures. They're what you get when the model is optimising for narrative correctness rather than narrative effect.
That gap is exactly what the Short Story Prompt skill for Claude was built to close.
The Workflow That Actually Changes the Output
The fix is not a longer request. It's a different sequence. Here's how to approach Claude for short fiction in a way that consistently produces drafts worth working with — and where a research-first skill file does the heavy lifting automatically.
What Good Revision Looks Like on an AI Draft
A well-generated short story draft — one produced with the workflow above, or with a genre-calibrated skill — needs revision, not reconstruction. The distinction matters. Reconstruction is what you do when the draft has the wrong bones: wrong structure, wrong register, wrong relationship between the story's surface and what it's actually about. That's a sign the generation phase failed. Revision is what you do when the bones are right and you're shaping the flesh — tightening sentences, cutting the explained emotions, making the detail more specific, fixing the ending.
The fastest way to tell which situation you're in: read the first paragraph and the last paragraph back to back, ignoring everything in between. If they feel like they belong to the same story — same register, same implicit subject, same level of restraint — the draft has good bones. If they feel like two different writers trying to tell two different stories, the generation phase produced something that needs to be started again, not revised.
The draft that needs revision is a writing problem. The draft that needs reconstruction is a generation problem — and the solution is upstream, not downstream.
When the bones are right, concentrate revision effort in this order: explained emotions first (every instance of "felt," "realised," "knew"), then the opening sentence (it almost always needs to be cut or replaced — the real opening is usually the second or third sentence), then the ending. The middle largely takes care of itself once the opening and ending are right, because the story's spine becomes clear and the loose material finds its own place.
A Note on What AI Can and Can't Do in Short Fiction
Generating structurally sound first drafts that give you something real to revise. Handling genre conventions you know well but want to produce faster. Producing multiple structural approaches to the same premise so you can choose the most interesting shape. Getting past blank-page paralysis when you know what you want to write but not how to start.
The skill augments the craft work — it doesn't replace it. The decisions that make a short story distinctive are still the writer's: the specific detail that only you would notice, the way the story's subject connects to something you actually know, the voice that's calibrated to your particular sensibility rather than to a genre average. A well-built skill file handles the research and the structural scaffolding. The writer handles the idiosyncrasy. That division of labour is what produces fiction that's both competent and interesting — which is a rarer combination than it sounds.
Writers who get the most from this workflow are the ones who treat the AI draft as a working model of the story's possibilities, not as a finished object. They revise with confidence because they know what the draft is good at (structure, genre awareness, pacing) and what it's likely to have smoothed over (specificity, genuine surprise, the particular). They do the second kind of work themselves. The draft just means they're not doing both kinds at once from a standing start.
The next piece most people tackle from here is a screenplay structure that holds across three acts. If you are still choosing a model for fiction, Claude vs ChatGPT for creative writing runs both against the same brief.
Frequently Asked Questions
How do you write a short story with AI?
Order of operations matters more than the prompt itself. First, specify genre, tone, and point of view before asking the AI to write anything. Second, tell it what the story refuses to do — the structural defaults to avoid. Third, let it produce a complete draft, then revise for specificity: replace every generic emotion with a concrete action, image, or detail. The draft handles structure; the writer handles idiosyncrasy. See NovaKit's Short Story Prompt skill for a genre-calibrated workflow that handles this intake automatically.
Why won't Claude write certain scenes in a short story?
Claude declines scenes involving explicit sexual content (without operator permissions), graphic violence without literary purpose, or real people in harmful scenarios. The most effective workaround: instead of asking Claude to "write the scene," ask it to describe what happens next in terms of consequence and emotional aftermath. Many scenes that fail as explicit requests succeed as oblique, consequence-driven ones. For a full breakdown of why Claude declines and what actually works, see Why Claude Won't Write Your Scene — and How to Fix It. Related: Why AI dialogue sounds flat and what's missing.
How do you make AI short fiction sound less robotic?
Remove every instance of "felt," "realized," and "knew" in the draft and replace each with a specific action, image, or piece of dialogue that implies the same emotion. The second most common AI-fiction signal is a tidy resolution — stories that earn their endings usually refuse to resolve cleanly. Cut the last paragraph and check if the story becomes sharper without it. For a deeper look at craft failure modes in AI writing: why AI scripts describe instead of show.
Put this to work: the Short Story Prompt skill for Claude turns everything above into one guided workflow you run in a normal Claude chat. Not ready to buy? Start with a free Claude skill and see how it works first.