Most AI video that looks fake goes wrong in the same few ways: plastic skin, light from nowhere, a camera that floats as if nobody is holding it, and three things happening at once. Often the prompt asked for it. "Cinematic, epic, ultra realistic, 4K" gives a model nothing to film, so it falls back on its defaults, and those defaults are what viewers have learned to spot.
This guide is a method for prompts that produce footage that reads as filmed: seven things to specify every time, one shot per prompt, and before-and-after rewrites for UGC, product and faceless styles. It isn't tied to one model; the examples reference Seedance, Veo and Kling. Model facts link to sources and are as of October 2026.
Generated shots work best next to real footage. ClipSpeedAI finds the best moments in your streams, podcasts and YouTube uploads, cuts them to vertical 9:16, burns in captions and gives each clip a viral score.
Try ClipSpeedAI →A video model can't ask what you meant. Every detail you leave out gets filled with something typical, and enough typical choices add up to the glossy, weightless look people recognize. Many of the common tells trace back to the prompt:
Some problems a prompt can't fully fix: hands, on-screen text and fast action can still go wrong on any model. Regenerate, crop or cut around them.
Outside of multi-shot modes, a generation is one continuous take. Ask for a wide street shot, then a close-up of a cup, then someone laughing, and the model has to get between them without a cut, which is where faces melt and objects change shape. Write one camera position, one subject and one action, and build sequences in your editor.
Some models do offer a multi-shot mode: Kling's 3.0 user guide lists up to six shots per generation. Even then, write each shot as its own block with its own camera, subject and action.
Length matters too. Veo 3.1 makes 4, 6 or 8 second clips, and Seedance 2.5 runs 4 to 30 seconds in one pass. Longer takes give the model more time to drift. Act the shot out with a stopwatch: if it doesn't fit the length you picked, cut a beat or split the shot.
| Part | What to specify | Vague | Specific |
|---|---|---|---|
| Subject | Who or what, concrete details, one imperfection | "a woman" | "a woman in her early 30s, faded green sweatshirt" |
| Action | One main action you could direct | "using the product" | "flips the lid open with her thumb" |
| Camera | Shot size, angle, one movement, lens look | "dynamic camera" | "handheld close-up at eye level, slow push-in" |
| Light | Source, direction, time of day | "cinematic lighting" | "late-afternoon window light from the left" |
| Setting | A specific place, two or three real details | "a kitchen" | "a small rental kitchen, dish rack by the sink" |
| Timing | Order of beats, pace, held moments | "then stuff happens" | "first she pours, then looks up; hold the last second" |
| Audio | Ambience, specific sounds, any line in quotes | "nice audio" | "room tone, a kettle starting to hiss" |
A fixed order makes prompts easier to compare. We put camera first so every prompt opens with the framing; the exact order matters less than keeping it consistent. Drop the labels when you paste:
[Camera] shot size, angle, one movement, lens look.
[Subject] who or what, with concrete details.
[Action] one main action, in order if there are beats.
[Setting] a specific place, two or three real details.
[Light] the source, its direction, the time of day.
[Timing] pace, and what happens at the start and end.
[Audio] ambience, specific sounds, any spoken line in quotes.In most tools, aspect ratio, length and resolution are settings; typing "9:16" or "4K" into the prompt won't reliably change the output. Our Seedance 2.5 guide covers prompt structure, modes and limits for that model.
Describe what a camera would see. "A confident entrepreneur" is a feeling. "A man in his 40s in a navy quarter-zip, typing on a scratched laptop" is a picture. One or two imperfections (a faded sweatshirt, a chipped mug) help; a dozen turn into noise.
For action, use a verb you could give an actor: pours, lifts, unscrews, laughs. Put beats in order with "first" and "then". Abstract verbs like "showcases" get a generic guess. Settings work the same way: "a kitchen" gets you a showroom, while "a small rental kitchen, dish rack by the sink" gets you somewhere a person lives.
Don't name real people or copyrighted characters. Disney sent ByteDance a cease-and-desist over Seedance 2.0 in February 2026, and TikTok's community guidelines ban fake endorsements by public figures even when labeled.
A real camera has an operator, a position and a reason to move. Giving the model all three is one of the quickest ways to make a clip look shot rather than generated.
Name the shot size (wide, medium, close-up, macro), the height (eye level, low, overhead) and one move, or none: locked-off on a tripod, a slow push-in, a pan, a handheld follow. "Handheld, slight natural shake" suits UGC; "locked-off" suits product. In 9:16, close-ups usually hold up better on a phone than wide shots. Movement can cost consistency, too: Google says camera pans and scene cuts can break character consistency in Gemini Omni 1.1 Flash, so keep the camera still or slow when a face has to stay recognizable.
"24mm wide", "85mm portrait" and "shallow depth of field" work as style hints, not real optics, but they usually push perspective and blur the right way. For UGC, name the device instead: "filmed on a phone's front camera at arm's length" pulls the model off its cinematic default.
"Cinematic lighting" and "studio lighting" usually get you flat, sourceless light, a big part of the plastic look. Name the source and the time of day: a window on the left, low sun through blinds, a desk lamp. Light with a source casts shadows, and shadows sell a frame as real.
Timing is the slot people skip. Give the pace ("slowly", "in one quick motion"), order the beats, and ask for a held moment at the end so you have a clean cut point. For a hook shot, start with the action already under way.
Many models generate sound with the picture. Seedance 2.5 generates audio jointly with the video, and Veo 3.1 generates dialogue and effects, though Google calls natural speech in short segments "an area of active development". Prompt audio like a sound designer:
Listen to every spoken line before posting; models can swap or invent words. Our native audio guide goes deeper.
Before:
Beautiful young woman talking about how much she loves her new travel mug,
cinematic lighting, ultra realistic, 4K, trending on TikTokNothing to film. "Beautiful" and "ultra realistic" pull toward a textureless face, the lighting has no source, and "how much she loves" leaves the model to invent both the words and the claim.
After:
Selfie-style medium close-up, phone held at arm's length by the subject,
eye level, slight natural hand shake, phone front-camera look.
A woman in her early 30s, hair in a loose claw clip, faded green sweatshirt,
sits in the driver's seat of a parked car.
She holds a matte white travel mug up to the lens, flips the lid open with
her thumb, then presses it shut until it clicks.
Midday daylight through the windshield, slightly overexposed on the left
side of her face; the back seat is soft and out of focus.
Casual pace; she looks back at the lens for the last second.
Sound: muffled traffic outside, the lid clicking shut.
She says, casually: "Flip, click. That's the whole lid."One shot, one action with a sound, light with a source and a flaw, a phone look, and a line that demonstrates the product rather than claiming months of use.
Before:
Cinematic product video of wireless earbuds, epic lighting, camera flying
around the product, dynamic, high quality, logo visible"Flying around" plus "dynamic" gives a floaty orbit with no operator, and "logo visible" invites a garbled logo, because models often get text wrong.
After:
Locked-off close-up on a tripod, slightly above desk height,
100mm macro look, shallow depth of field.
White wireless earbuds in an open charging case on a light oak desk;
a ceramic coffee cup sits out of focus behind them.
A hand enters from the right, lifts the left earbud out and leaves frame.
Soft morning window light from the left, a gentle shadow to the right.
Slow, deliberate movement; the frame holds still at the start and end.
Sound: quiet room tone, a faint magnetic click as the earbud lifts out.The still camera reads like a real product shoot, the light has direction, the held frames leave room to cut, and it asks for no text. If the product must look exactly like yours, animate a real product photo and add the logo in your editor. See our image-to-video workflow.
Before:
A video about the history of coffee with coffee farms, people drinking coffee
in old cafes and a modern coffee shop, cinematic, with narrationAt least three shots and two eras in one take, so it will morph between them, and "with narration" asks the model to write and voice a script you haven't approved.
After, the first of several separate prompts:
Slow handheld walk forward at chest height between rows of coffee plants
on a hillside. Ripe red coffee cherries in the foreground; mist behind.
Early morning, overcast, soft even light, wet leaves.
Steady pace; the camera keeps moving forward the whole clip.
Sound: ambient only, birdsong, wind in the leaves, footsteps on damp soil.The next shot, say an overhead close-up of water poured over coffee grounds, gets its own prompt, again with ambient sound only, leaving room for a voiceover you record separately. Realistic b-roll can still need disclosure: YouTube's rules cover realistic scenes that never happened. Our faceless channel guide covers the full workflow.
When a take misses, find the slot that caused it and change only that one. Rewrite everything and you won't know what fixed it. Also:
For side-by-side tests, ClipSpeed's AI Creator has five video models: Seedance 2.5 for long single takes, Veo 3.1 for vertical hero shots, Kling 3.0 Turbo for high-volume variants, MiniMax H3 Max for fast image-to-video and Gemini Omni 1.1 Flash for editing existing clips. Nano Banana Pro, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst make start frames. Generations use ClipSpeed creation credits. ClipSpeed Create, in early access, brings this into Claude or ChatGPT (how to connect): every generation is quoted first, and nothing runs until the account owner approves it on a signed-in ClipSpeed page.
AI video looks like AI when the prompt leaves the model to guess. Write one shot per prompt and cover all seven parts: a concrete subject, one filmable action, a camera with a position and at most one move, light with a source, a specific setting, pace and held moments, and the sounds you want. Put format in settings and change one thing at a time.
You still have to check every clip and label realistic AI content where required. Generated shots are strongest next to real footage, and if you have long videos, ClipSpeedAI turns them into captioned vertical clips to build around.
How do I write an AI video prompt that doesn't look AI?
Cover seven things in every prompt: a concrete subject, one filmable action, the camera (shot size, angle, one movement, lens look), a light source with direction, a specific setting, timing (pace and held moments) and audio cues. Replace adjectives like "cinematic", "epic" and "ultra realistic" with things a camera could see, add one real-world imperfection, and describe one shot per prompt.
How long should an AI video prompt be?
Long enough to fill all seven slots, which is usually a short paragraph or one line per slot. Length on its own doesn't help: a long list of adjectives is still vague. If a prompt runs on, check whether it describes more than one shot or more action than fits in the clip length.
Can I describe several shots in one AI video prompt?
Usually you shouldn't. Outside of multi-shot modes, a generation is one continuous take, and asking it to move between shots without a cut is where faces and objects tend to morph. Some models do offer a multi-shot mode; Kling 3.0 lists up to six shots per generation. Even then, write each shot as its own block with its own camera, subject and action.
Should I write the aspect ratio or resolution in the prompt?
No. In most generators, aspect ratio, length and resolution are settings, and typing "9:16" or "4K" into the prompt won't reliably change the output format. Set them in the generator's controls. Choose the ratio you'll post in, such as 9:16 for Shorts, Reels and TikTok, rather than cropping later.
How do I prompt dialogue and sound in AI video?
Write ambience and sound effects tied to the action ("the lid clicks shut"), and put any spoken line in quotes with how it's said. Keep each line short, leave music out of the prompt and add licensed music in your editor. Listen to every line before posting, because models can change or invent words. See our native audio guide.
Do these prompts work in Veo, Kling and Seedance?
The structure carries across models, but each has its own limits. As of October 2026, Veo 3.1 makes 4, 6 or 8 second clips, Seedance 2.5 makes 4 to 30 seconds in one pass, and Kling 3.0 supports multi-shot generation. Veo 3.1, Seedance 2.5 and Kling 3.0 Turbo are available side by side in ClipSpeed's AI Creator, so you can run one prompt on each and compare.
Do I have to label AI video that looks real?
On the major platforms, often yes. YouTube requires disclosure of realistic altered or synthetic content, and TikTok requires a label on AI-generated content showing realistic scenes or people. Separately, the FTC's fake reviews and testimonials rule covers AI testimonials attributed to people who don't exist. This isn't legal advice; check current terms, and see our AI video disclosure guide.
Published by ClipSpeedAI · AI video generation and AI clipping in one place — create with Seedance, Veo, Kling and Nano Banana, then cut it into captioned shorts.