Scary AI Video Prompts for Halloween That Actually Unsettle

Most AI horror clips fail for the same reason: the prompt asks for a monster. Video models draw creatures badly (too glossy, too many teeth, zero dread), but they are very good at empty rooms, stillness and small wrong details. So the prompts in this guide describe a hallway, one light bulb, a camera that does not move and a noise from somewhere off-screen. The viewer fills in the rest, and that is where the fear comes from.

You get ten shots to copy, each paired with the model I would run it on, a poster prompt for the invitation, a list of things that get prompts rejected, and a budget for a 30-second short.

How every prompt here is built

Each prompt is one shot and follows the same order: framing, subject, action, camera, light, style, sound. If camera terms are new to you, my guide to AI video prompts and camera moves explains them. Run each shot at the lowest resolution a model offers first, and render the keeper at a higher one. Model names link to pages with current prices.

Ten shots to copy

Shot 1: a door that opens on its own (Veo 3.1 (Quality))

Wide static shot down a narrow apartment hallway at night, locked-off camera on a tripod.
At the far end a bedroom door slowly swings open by itself and stops halfway.
Nobody is there. The only light is a dim bulb in the hallway ceiling, the bedroom beyond is pitch black.
Realistic, muted colors, slight film grain.
Audio: low fridge hum, a slow door creak, then silence, no music.

Shot 2: footsteps above the kitchen (Veo 3.1 (Quality))

Medium static shot of a dark kitchen, lit only by the light of an open fridge.
A woman in a sweater stands with her back to the camera, holding a glass of water.
She freezes and slowly lifts her head toward the ceiling. Footsteps cross the floor upstairs.
Locked-off camera, cold fridge light, deep shadows. Realistic indie horror style.
Audio: fridge hum, three slow heavy footsteps upstairs, the glass trembles slightly, no music.

Shot 3: something behind the trail camera (Veo 3.1 (Quality))

Night-vision trail camera footage, black and white infrared, fixed camera strapped to a tree.
A deer grazes at the edge of a forest clearing. It suddenly raises its head, stares
past the camera and bolts away. The clearing stays empty for a moment.
Timestamp-style framing, low resolution texture, no on-screen text.
Audio: crickets, a branch snaps somewhere behind the camera, the crickets stop.

Shot 4: candles going out step by step (Seedance 2.5)

Static wide shot of a wooden porch on Halloween night, six carved jack-o'-lanterns
lined up on the steps, candles flickering inside.
From left to right the candles go out one by one, as if something is walking up the steps.
The porch light flickers once. Locked-off camera. Warm candlelight against blue night.
Audio: wind in dry leaves, a soft puff with each candle, a wooden step creaks, no music.

Shot 5: a whistle answered from the woods, one 15-second take (Seedance 2.5)

One continuous shot, 15 seconds, static camera on a tripod at the edge of a pine forest at night.
0-5 s: a man with a flashlight stands with his back to the camera and whistles two notes into the dark.
5-10 s: silence, the flashlight beam shakes slightly on the tree trunks.
10-15 s: from deep in the forest, the same two notes are whistled back. He does not move.
Single flashlight as the only light, cold mist. Realistic, documentary feel.
Audio: distant wind, his breathing, the answering whistle, no music.

The timestamps let Seedance pace a long take instead of rushing everything into the first seconds.

Shot 6: a child in an old photo turns his head (Grok Imagine Video 1.5)

On Kadraris this model animates a start image, so make the still first, for example in Nano Banana Pro: a faded black and white photo of a family of five outside a farmhouse. Then send:

The camera stays completely still. The old family photograph is motionless,
except the small child at the edge of the group, who slowly turns his head
toward the camera. Nothing else moves. Paper grain and scratches remain.
Audio: an old clock ticking, a faint breath close to the microphone, no music.

Shot 7: one swing moving in the fog (Grok Imagine Video 1.5)

Start image: a playground in fog at dusk with a single swing in the middle of the frame.

Static shot. The fog drifts slowly. One swing begins to move on its own,
higher with each swing, while the swings around it hang perfectly still.
Sodium streetlamp light from the left, everything else in grey fog.
Audio: a chain creaking in rhythm, distant traffic, no music.

Shot 8: a whisper in bed (Kling 3.0, sound on)

Close-up of a teenage girl lying in bed under a blanket in a dark room, lit only by her phone screen.
She lowers the phone, eyes wide, and whispers: "Did you hear that?"
Behind her, the wardrobe door is open a few centimeters, the gap is completely black.
Static camera, cold phone light on her face. Realistic, shallow depth of field.
Audio: her whisper, a quiet knock from inside the wardrobe, no music.

Kling 3.0 renders silent clips unless you turn on generate_audio. You need it here for the whisper and the knock, and the rate with audio is higher than without.

Shot 9: the mirror that lags behind (Kling 3.0)

Medium shot of a steamy bathroom mirror, a man wipes the fog off with his hand
and looks at his reflection. He turns his head to the left. His reflection turns
half a second later. Static camera behind his shoulder.
Single bulb above the mirror, warm light, dark tiles. Realistic horror style.
Audio: dripping tap, extractor fan hum, no music.

Shot 10: trick-or-treaters at the last house (Seedance 2.5)

Static wide shot from the sidewalk at night: three children in Halloween costumes
with candy buckets walk up to the last house on the street. Before they knock,
the front door opens slowly on its own. Warm light spills out, the hallway is empty.
The children stop. Locked-off camera, streetlamp and porch light only.
Audio: dry leaves, a bucket handle rattles, the door hinge creaks, no music.

Shot 4 makes a strong opener and shot 10 a strong ending if you cut them into one short.

Why these prompts work

The same habits run through all ten shots. Use them when you write your own:

• Imply the threat. A door open wider than a moment ago, a silhouette behind frosted glass, a dog staring at an empty corner. Every detail you do not ask the model to draw is one it cannot get wrong.
• Lock the camera. static shot, locked-off camera, tripod makes the frame feel like security footage, so any small movement inside it jumps out. Moving AI cameras tend to drift, and drift kills suspense.
• Describe the sound before anything else. A ticking clock, rain on glass, a creak upstairs, and no music. Real horror scares mostly through the ears. Veo 3.1, Seedance 2.5, Grok Imagine Video 1.5 and Kling 3.0 with audio enabled all create ambient sound from text.
• Name the light source. a single bare bulb, light from an open fridge, a phone screen, a flashlight beam, a flickering streetlamp. Whatever that light does not reach becomes the scary part of the frame.
• One event, 5 to 8 seconds. A long clip gives the model room to wander, and a short film is made of cuts anyway.

Things that ruin a horror clip or get it blocked

• Blood and explicit violence. Model vendors filter wounds, gore and attacks, sometimes even mild ones. A rejected job is free, but you lose the time. Suggest the violence instead: a scream off-camera, a flashlight rolling on the floor.
• A creature in plain view. If something must appear, keep it soft-focus, far back, or a one-second silhouette.
• Sweeping camera work. Orbits, cranes and drone flights make the scene look like a game trailer.
• Asking for music. Generated scores are bland and hard to edit. Write no music and add a track in the edit.
• Writing in the frame. Dates, titles and messages on walls come out misspelled. Add text in the edit or on a still poster.
• Stacking events. "The door opens, a ghost appears, she screams and runs" is four shots, not one.
• Real people and famous franchises. No photos of people without consent, and no characters from well-known horror films.

A poster or invite with lettering (Nano Banana Pro)

Lettering inside video still falls apart, so I keep titles and invitation details on a still. Nano Banana Pro gets short lines right in most attempts. Keep the text brief and put it in quotes:

Vertical poster for a Halloween house party, 4:5. A dark wooden front door at night,
slightly open, warm light and thin smoke coming out of the gap, two carved pumpkins
on the steps. Large condensed serif headline at the top: "HALLOWEEN NIGHT".
Below the door, smaller text: "31.10 / 8 PM / COSTUMES REQUIRED".
Muted orange and black palette, film grain, no other text.

Switch the ratio to 16:9 and put your film title in the quotes, and the same prompt gives you a YouTube thumbnail. More tips on lettering are in AI text in images.

Budget for a 30-second short

Thirty seconds is about 4 to 6 shots. The figures below cover only the footage you keep; with tests and throwaways, count on two or three times that. Veo 3.1 always outputs 8 seconds, so you need four clips.

These are each model's lowest rates. Higher resolution and Kling with sound cost more, and Studio shows the exact amount before you start. Failed jobs and jobs blocked by a vendor filter cost $0. With no subscription and a balance that never expires, whatever remains after October 31 stays for your next project. All rates are on the pricing page.

Batch shots through the API

To render many variants of one scene, say twenty versions of the hallway, use the AI video generator API with one request per shot:

{
  "model": "veo-3.1",
  "input": {
    "prompt": "Wide static shot down a narrow apartment hallway at night, locked-off camera on a tripod. At the far end a bedroom door slowly swings open by itself and stops halfway. Nobody is there. Dim ceiling bulb, the bedroom beyond is pitch black. Audio: low fridge hum, a slow door creak, then silence, no music.",
    "aspect_ratio": "9:16",
    "resolution": "720p"
  }
}

Post it to POST /v1/generate, then collect the file with GET /v1/tasks/{id} or let a webhook notify you.

Questions

Which model should I start with for a horror short?

It depends on the shot. Veo 3.1 for short realistic clips with ambient sound, Seedance 2.5 for long single takes with a timeline, Grok Imagine Video 1.5 when you want to animate an old photo, and Kling 3.0 with audio for whispered lines. All four run from one balance in Kadraris Studio.

My prompt was refused. What should I change?

Take out blood, injuries and weapons, and hint at the danger instead. Vendors are stricter with horror than with most genres. You are not charged for a refused job.

How do I keep drafts cheap?

Write and test on Veo 3.1 Fast at low resolution, and only render the approved shots on Veo 3.1 (Quality). There is no free plan, but failed generations cost $0 and the minimum top-up is $10.

Can I mix shots from different models in one film?

Yes. The shots are separate clips, so you can cut a Veo hallway next to a Seedance porch in any editor. Keep the light and grain words similar across prompts so the cuts do not jump.