How to write AI video prompts that look good (with examples)

Most AI video that looks cheap was not made with a bad model. It started with a vague prompt. "A nice coffee shop" leaves the model guessing about everything that matters: what is in the frame, what moves, where the camera sits, what the light looks like. A good scene prompt makes those choices for it, in plain words, so the result looks planned instead of random.

This guide shows how to write AI video prompts that look good, with a simple six-part structure, camera and lighting vocabulary that suits vertical video, what to leave out, and before-and-after examples across six niches. Write the prompts themselves in English: most video models follow English prompts most reliably, even when the voiceover is in Arabic.

Why the prompt matters so much

A video model turns words into a few seconds of motion. Every detail you don't specify, it fills in with something generic. That is where the familiar "AI look" comes from: flat light, a drifting camera, a cluttered frame and a subject that does nothing in particular.

A clear prompt does three jobs. It tells the model what to show, it tells it how to film it, and it keeps each scene consistent with the ones around it. You don't need long prompts. One or two well-built sentences per scene usually beat a paragraph of adjectives.

The six parts of a scene prompt

Think of every scene as six decisions: subject, action, setting, camera, lighting, style. Here is one prompt built step by step.

1. Subject: what the viewer looks at. a ceramic cup of Arabic coffee
2. Action: what moves. Video needs motion, even small motion. coffee being poured from a brass dallah into the cup, steam rising
3. Setting: where it happens. on a low wooden table in a majlis, cushions softly blurred behind
4. Camera: how it is filmed. slow push-in, close-up, vertical framing
5. Lighting: what the light does. warm late-afternoon window light from the left
6. Style: the overall look. cinematic, shallow depth of field, rich warm tones

Put together:

A ceramic cup of Arabic coffee on a low wooden table in a majlis, coffee being poured from a brass dallah, steam rising, cushions softly blurred behind. Slow push-in, close-up, vertical framing. Warm late-afternoon window light from the left. Cinematic, shallow depth of field, rich warm tones.

Order matters less than completeness. If one of the six is missing, the model invents it, and invention is where things go wrong.

Camera moves that suit vertical video

A 9:16 frame is tall and narrow, and it is watched on a phone. Wide sweeping pans lose the subject off the edges. These moves work better:

• Slow push-in: the camera moves gently toward the subject. It adds focus and is the safest choice for a hook scene.
• Tilt up or tilt down: vertical movement fits a vertical frame. Good for buildings, outfits, a tall glass being filled.
• Top-down: looking straight down at a table, a desk or a plate. Clean, graphic and easy for models to keep stable.
• Close-up or macro: hands, textures, details. Small screens reward big subjects.
• Handheld feel: a slight natural sway makes a scene feel filmed by a person rather than generated. Ask for it softly, as in subtle handheld movement, not shaky.

Pick one camera move per scene. "Push-in while orbiting and tilting" usually produces a confused shot.

Lighting, style and one consistent look

Lighting words do more for quality than almost anything else. Useful ones: soft morning light, golden hour, warm window light, overcast daylight, neon reflections at night, bright studio lighting on a white background, moody low-key lighting.

Style words set the overall finish: cinematic, shallow depth of field, 35mm film look, clean commercial look, natural documentary style, muted earth tones.

The trick for a video that feels like one piece is a style tail: a short phrase you add, word for word, to the end of every scene prompt. For example:

warm golden-hour light, cinematic, shallow depth of field, soft film grain

When every scene ends with the same tail, colour, light and texture stay consistent even as the subject changes. Change the tail only when you want a deliberate shift, such as a night scene.

What to leave out, and why

Some things reliably go wrong in AI video, so the best prompt simply doesn't ask for them.

• On-screen text and Arabic letters. Most video models garble writing, and Arabic script comes out especially broken. Signs, menus, labels and titles all suffer. This is exactly why Vocla keeps captions as a separate track: your words are burned in cleanly by AI captions, not drawn by the video model.
• Logos and brand names. They come out distorted, and you don't want a near-miss version of a real brand in your video. Describe the object instead: a matte black wireless speaker.
• Real or famous people. Don't name celebrities or public figures. Describe a generic person: a young woman in a beige abaya.
• Too many subjects. Three people, a dog and a car in one 5-second scene is a recipe for warped hands and melting faces. One main subject per scene.
• Contradictory instructions. "Bright sunny night", "static camera with fast motion", "minimalist and full of detail". The model can only follow one of them, and you won't like which.

Before and after: six prompt pairs

Food

• Weak: delicious shawarma
• Better: Close-up of a chicken shawarma wrap being sliced in half on a wooden board, steam rising, garlic sauce glistening. Slow push-in, vertical framing. Warm kitchen light. Cinematic, shallow depth of field.

Real estate

• Weak: luxury apartment in Dubai
• Better: Bright modern living room with floor-to-ceiling windows overlooking a city skyline at sunset, sheer curtains moving slightly. Slow tilt up from the sofa to the view. Golden-hour light. Clean architectural style.

Product

• Weak: perfume bottle advertisement with brand name
• Better: A faceted glass perfume bottle on a stone pedestal, soft mist drifting around it. Slow push-in, close-up. Soft studio lighting on a dark background. Luxury commercial look.

Travel

• Weak: beautiful desert trip
• Better: A lone traveller in a light scarf walking along the crest of a red sand dune, footprints trailing behind. Wide shot, slow tilt down from the sky. Golden-hour light, long shadows. Cinematic, warm tones.

Fitness

• Weak: gym workout motivation
• Better: Close-up of hands gripping a kettlebell and lifting it from the floor in a minimal gym. Subtle handheld movement, vertical framing. Moody low-key lighting with a single side light. Gritty sports commercial style.

Tech

• Weak: phone app with text showing features
• Better: Top-down view of a smartphone on a clean white desk next to a coffee cup, a hand tapping the blank glowing screen. Static camera. Soft daylight. Minimal clean commercial look.

Notice the pattern: the better version names one subject, gives it something to do, picks one camera move, sets the light and drops anything with text or a brand.

How prompts fit the Vocla flow

In Vocla Video you don't start from a blank prompt box. You write or paste your script, choose a voice in any Arabic dialect or in English or French, and Vocla splits the voiceover into scenes of 3 to 8 seconds, each timed to the narration. For every scene it writes an English visual prompt that already avoids on-screen text, logos and real people.

Your job is to edit, not to start from scratch:

1. Read each prompt against the line it illustrates. Swap in your niche's details and your style tail.
2. Generate storyboard stills first. Stills are cheap compared with video, and one image per scene shows straight away whether a prompt works.
3. Render in the quality tier that fits the job: Lite, Standard or Pro. Higher tiers cost more credits per rendered second.
4. Regenerate a single scene if one misses, without touching the rest. A failed scene is refunded.

From Starter up you can also upload a product photo as the first frame, which is the most reliable way to keep a real product looking like itself. Plan details are on the pricing page.

If you post short vertical video, the AI TikTok video maker and its script planner help with the script before you ever reach the prompts. For opening lines, see 30 TikTok hooks in Arabic dialects, and for topics, 50 TikTok video ideas with Arabic AI voiceover. The full walkthrough, including a copy-paste prompt library, is in the guide on how to generate videos with AI.

Quick checklist

Before you render, check every scene prompt:

• One main subject, doing one visible action
• A clear setting
• One camera move that suits a vertical frame
• A lighting phrase
• The same style tail as every other scene
• No text, letters, logos, brand names or real people
• No instructions that contradict each other
• Storyboard still approved before rendering