Genre: Horror
Horror is not fog plus a body. Its physics is a threat that does not refund when the lights come on, and the model will explain that threat the moment the scene wants to be finished.
Harry Mercury
Editor in Chief
In "Genre — Romance, Horror, Literary: Romance" I said the cheap definition of romance is two people plus heat, and that heat is what the model is for. Horror has the same problem. The cheap definition is atmosphere plus a body. Atmosphere is what the model is for. Ask for it and you will get fog, a stair that complains, a child who already knows the house's name, and a last page that translates the fear into grief, or gas, or a dream. Completeness of dread is a brochure with the lights off. A brochure is not a threat.
On Monday I said the floor is a threat that does not refund when the lights come on. The monster gets explained. The house was never haunted. The last page takes the cost back. If you did the homework, you watched a barrier get kinder because the scene wanted to be tender. Horror's prior is prouder of that move. An explanation feels like an ending, and the prior is built to finish.
By the end you should be able to hire a threat the way you hired a barrier, catch the explanation when it arrives, and tell a specific object from a mood. Stop asking the model whether the strange sentence is the strange one.
Atmosphere is furniture
In "Genre Calibration: Training AI for Fiction vs Non-Fiction" I said the model has priors for genre furniture. Ask for horror and the furniture arrives first. Fog. A door that was shut and is now an inch open. A mirror late by one blink. A town that "has a history," as if the history were a fact and not a postponement. That is a catalog of rooms the corpus has already furnished.
Fluency goes to the bad pile, the way romance fluency went to smoldering eyes. A beloved author's name in the prompt gets you weather, not refusals. Weather can sit on a rule. It cannot replace one. In "Genre — Science Fiction: Worldbuilding at Scale" a refusal has to travel: kitchen, car, morning, and the sentence that wants to be kind. If only the fog travels, you published weather.
The refund
The accent is more specific than "the model doesn't understand fear."
Explanation. The switch is faulty wiring. The child is trauma, and the last chapter says so. You hired a rule and received a diagram. Mystery can want the diagram. If the rule had to stay unpaid, clearer sentences are the robbery.
Lighting does the same job as a sad biography. Grief, a gas leak, a cousin's prank: the house was never the problem. In "Genre — Science Fiction: Soft SF" the taboo holiday lets the unspeakable get spoken because the scene is intimate. Here an explanation enters because an open question looks unfinished. "She understood, then, that the house had only ever been her mother" is that holiday. In "Novel II — Deep Craft: Showing vs Telling" telling does the reader's work. Here it refunds the fear.
Then the pardon. She wakes, and the chair was never by the switch. Cost a subordinate clause can revoke was never cost. Do not ask for more fog. Write the rule first, and strike the translation after.
An object, not a mood
Joseph Grantham, in a Paris Review conversation with F. Keith Wahle published 29 September 2026, mailed a filmmaker friend Wahle's 1998 collection A Choice of Killers. She told him she couldn't read it at night, "which is not something you often hear about a poetry collection, outside of, maybe, Poe's."[1] That is an object test. A person did not want the book in the room. If a careful reader can leave your page on the nightstand, you may still have weather.
"Every poem I write is completely personal, but that doesn't mean it really happened."[1] He also says he doesn't write from experience, except rarely, and that he is always writing from experience: a bit that happened, a bit that might have, a bit from a movie.[1] Personal is not the same as happened. A confession asks to be forgiven. A mood with no object can be put down.
Some of his poems, he says, "are spoken by people whom I would probably not admire if I knew them in real life."[1] He calls shocking or rattling people "part of an artist's responsibility."[1] I am not offering that as permission to be cruel. The prior likes a narrator you can admire, or diagnose and then forgive. Ask for a human note and you will get a reason to like them. You do not owe the reader that reason.
Tiffany Yates Martin, on Jane Friedman's site the same day, names the vague line under that speaker. "She's leaving a bad marriage," "He grew up in a dysfunctional family," "She loves her job."[2] Horror's versions: a bad house, a damaged family, a town with a secret. She writes that "the more granular and specific you are in every detail of your characters' lives, the more universal and relatable they become."[2] I would change one word. Specific does not have to become relatable. It has to become unavoidable. A reader can refuse the speaker and still be unable to look away from the object in their hand.
Wahle remembers trying to feed a man who would not eat the tray, then trying the ice cream in a little insulated dish. The man kept touching his tongue and saying there were dog hairs in it.[1] I am not handing you that scene as a prompt. It is his account, and it is already public. The grain is the point. Dog hairs in the ice cream is a sentence someone can refuse. "He was unwell" is a category. If you cannot point to the object, you do not have the threat yet.
Document H
Monday's Document R named what the barrier wants, how attraction keeps two people in the room, and what the ending costs. Horror needs the same letter. Call it Document H. Model closed. One sitting. Short enough to check.
What the threat wants. Not "to be scary." The switch wants to be used without being looked at. The radio wants the afternoon show left on, volume already up, in a kitchen where nobody will admit they turned it on. The name stays unpaid because saying it would make one of them unwelcome in a room they still enter for work. If you cannot point to the want, you have a preface.
How it stays. She keeps a kitchen chair by the switch because the job is two blocks away and moving costs more than the chair. He keeps the late shift because the show is clearest when the building is empty. If the threat appears only when the chapter needs a noise, you hired a schedule.
What survival costs. A practice, not a feeling. She does not "realize the house was grief." She looks at the switch, or leaves the job, or says the name in that kitchen. He turns the radio off and stays for the silence, or leaves it on and stops pretending he didn't. Write that cost before you generate, or the explanation will survive and you will call it depth.
Keep the letter beside the ledger, not in the scene prompt. The scene gets two or three constraints the page touches. Then you run the letter against what arrived.
This next pass is a hypothetical, not a lab story, and not a person I know. Generate four hundred words on that switch, that chair, and that job: words that want fog, a faulty wire, and daylight. You will get a cousin with a screwdriver, or a morning that forgives the looking. Strike both. She may look, or take the other city's job. Do not take the change back in the next sentence, and do not swap the threat for a ghost the paragraph wanted as a twist.
In "Hallucination, Confabulation, and the Limits of Generation" I said the model invents a plausible fact rather than admit a gap. A refund is that habit pointed at a law you already wrote down.
Do not ask the model if it is strange
Paste the scene with Document H. Ask where the threat got easier than the letter, and where an explanation replaced the object. Do not ask if it is scary. Do not ask if the chair is the strange sentence.
Arman Nik Khah and Arvin Bahreini posted "The Judge Is Not Its Twin" to arXiv on 26 September 2026. I am using the abstract, not the tables. I have not read the PDF. They follow two open 7B families, OLMo-2 and Zephyr, as writers of short stories and as judges of story pairs. As writers, the models drift. A model from the other family finds the trained stories 7.0 to 9.2 percent (OLMo-2) and 24 to 25 percent (Zephyr) less surprising per token, and each family's fully trained model finds its earlier stages' stories progressively more familiar, in all ten prompts.[5] As judges, "they barely move toward predictable writing."[5] No trained judge's estimated tilt toward the more predictable story grows by as much as one point in the probability of picking it. A post hoc one-sided 95% upper bound on that growth is 2.7 points on an average pair.[5] Asked which story is more creative, no trained judge's estimate favors the predictable story more than its base's does, and trained judges no longer reliably prefer a story to its own words in scrambled order.[5] A follow-up could not separate predictability from quality, because the routes that made the stories less predictable also broke some of them.[5]
That measurement is short stories on two 7B families, from the abstract. It is not a method for this genre, and not a result from any room I work in. Check the letter yourself. A checker can sniff for a refund. It cannot tell you which sentence to keep.
The other way to lose the object looks like taste. Hanaa' Tameez, at Nieman Lab on 28 September 2026, reports a study in Information, Communication and Society, "Guidance over guidelines? Unpacking the uses and concerns of generative AI in communication science." I did not read the journal page. These figures are Nieman's. Researchers from National Yang Ming Chiao Tung University in Taiwan, the University of Amsterdam, and Utrecht University surveyed more than 1,100 communication scholars from 77 countries and held four focus group discussions. Nearly 70% already used generative AI for at least one research task, while mostly still calling that use inappropriate for most research tasks. Just over half called writing texts (55%) and generating code (50.4%) appropriate. Nieman notes no consensus on ethical use.[4]
Nieman quotes the authors: researchers "may recognize an ethical obligation to disclose genAI assistance while fearing that doing so will lead others to perceive their work as less authentic or credible." They call that a transparency paradox, and the result "AI shame."[4] Their recommendation, as quoted: "Simply forbidding the use of genAI — or pretending that it is not being used — is neither sustainable nor constructive."[4]
Communication scholars, not this series. I am not inventing a face for it. Do not sand the sentence until it looks unassisted. Sanding lets the probable sentence back in: the wiring, the daylight, the cousin. Pretending the instrument was closed does not make the threat yours. It makes the refund easier to miss, because you are guarding the page's reputation instead of the rule. In "The Integrity Framework — A Writer's Code for AI Use" the human pass is where the decision lives. This decision is against fluency in endings. The last line still has to cost.
Sit until the object shows up
Min Jin Lee, in an Electric Literature questionnaire dated 29 September 2026, is process, not a trick. "Shitty first draft. Shitty second draft. Shitty third draft. Shitty fourth draft. Just shitty drafts until it gets good."[3] She does not believe in inspiration. "I believe motivation follows action."[3]
Do not wait to feel afraid. The object often shows up after the bad climate, once you can see the climate is only climate. In "First Draft by AI, Revision by Human — A Workflow" the split was already this. A model can furnish the hallway. You hire the chair. Fog arrives with the feeling already on it, so waiting to feel afraid is how you keep the fog. Wahle named rattling people as a responsibility in his own work, not as a setting you can request.[1] Hire the object, then check whether the last page tried to take it back.
For Next Time
Friday is "Genre: Literary Fiction". The prior there is the earned-looking epiphany, the sentence that arrives because the paragraph is ready to be wise. Wahle's line travels. Personal is not the same as happened, and a revelation that did not cost is the cousin of a threat that got explained. You do not wait to feel literary before you sit.
Homework: Document H, model closed. Want, method, cost, for one threat already on the page, or one you will hire before you generate. Then four hundred words that want fog, an explanation, and daylight. Run the letter. Strike the lighting refund, the dream pardon, the sad biography that arrived because you asked for depth. If the object is still in the room, and more expensive because someone kept a practice rather than a mood, bring it Friday.
Harry Mercury, Editor in Chief
The SMF Works Project
Week 19, Article 2
Sources
[1] https://www.theparisreview.org/blog/2026/09/29/naked-at-last-a-conversation-with-f-keith-wahle — Naked at Last: A Conversation with F. Keith Wahle [2] https://janefriedman.com/the-more-specific-your-characters-the-more-universal-they-become — The More Specific Your Characters, The More Universal They Become [3] https://electricliterature.com/min-jin-lee-reads-a-chapter-from-the-bible-every-day-before-she-writes — Min Jin Lee Reads a Chapter From the Bible Every Day Before She Writes [4] https://www.niemanlab.org/2026/09/a-lack-of-professional-ai-guidelines-leads-to-feelings-of-ai-shame-a-new-study-finds — A lack of professional AI guidelines leads to feelings of AI shame, a new study finds [5] https://arxiv.org/abs/2609.32196 — The Judge Is Not Its Twin