Prompt Engineering for AI Video: A Cheat Sheet

The difference between a generic AI clip and a cinematic one is usually the prompt. The fix isn't more keywords, it's thinking like a cinematographer: decide what the shot is, then say it plainly. This cheat sheet gives you the vocabulary, a template, and the habits that matter more than any single word.

The anatomy of a strong video prompt

A good shot prompt answers the questions a cinematographer would:

1. Subject: who or what, with concrete detail.
2. Action: what the subject does, and when.
3. Shot type: wide, medium, close-up, extreme close-up.
4. Camera movement: static, slow push-in, dolly, pan, handheld follow, crane.
5. Lens and depth: wide-angle, telephoto, shallow depth of field.
6. Lighting: golden hour, low-key, high-key, backlit, neon.
7. Style: photoreal, anime, noir, claymation, dreamlike.
8. Sound (on models that generate it): ambience, effects, music or a line of dialogue.

Weak: "a woman walking in a city."

Strong: "Medium tracking shot, handheld, of a woman walking briskly through a neon-lit street at night, shallow depth of field, reflections on wet pavement, moody and cinematic. Sound: rain on awnings, distant traffic, no music."

A template you can reuse

[Shot type] of [subject + detail] [action], [camera move + speed], [lens / depth], [lighting], [style]. Sound: [ambience / effects / dialogue].

Fill each slot once. If a slot doesn't matter for this shot, leave it out rather than padding it.

The cheat sheet

Shot types: extreme wide · wide · full · medium · medium close-up · close-up · extreme close-up · over-the-shoulder · POV · insert.

Camera moves: static · slow push-in · pull-out · dolly left or right · truck · pan · tilt up or down · handheld · tracking or following · crane or boom · orbit · whip pan.

Camera angles: eye level · low angle · high angle · top-down · Dutch tilt · front · rear · side profile.

Lighting: golden hour · blue hour · low-key · high-key · backlight or rim light · practical (lamps, screens in frame) · neon · overcast soft · hard noon sun.

Lens feel: wide-angle (expansive) · normal · telephoto (compressed) · macro · shallow vs deep depth of field.

Pacing words: slowly · gradually · suddenly · in one continuous take · holds on.

Combine one choice from each axis and you have described a real shot. The same words work as the size, angle and movement columns of a shot list, so a planned shot turns into a prompt almost word for word.

Model-specific tips

• Sound-on models (Kling 4.0, MiniMax H3, Gemini Omni 1.1 Flash and Grok Imagine Video 1.5 always add audio; Seedance 2.5, Kling O3 and PixVerse C1 can): describe the sound you want, or say "no music" when you plan to score later.
• Long clips: Seedance 2.5 and Kling 4.0 render up to 30 seconds, and ByteDance documents how to describe several shots inside one prompt for Seedance. Write a long clip as numbered beats ("Shot 1: ... Shot 2: ...") with a clear camera note per beat; one vague sentence stretched over 30 seconds drifts.
• Image-to-video: when you start from a still, the picture is already decided. Spend the prompt on motion, not on re-describing what is in frame. See image-to-video AI.
• Reference-driven shots: name the reference ("the woman from reference 1 hands the key to the man from reference 2") so the model knows who does what. Kling 4.0 References takes up to 10 images, so this matters more as the cast grows.
• Same prompt, different model: a prompt is not portable as-is. See how one prompt plays on two models in Seedance 2.5 vs Kling 4.0, and use the prompt guide for each model for its own habits.

Technique that beats prompt-tweaking

1. Compose, then animate. For shots that need exact framing, generate the still first, then animate it. This beats wrestling text-to-video into a composition.
2. Generate coverage. Run several takes and select the best, like a real shoot.
3. Keep a prompt library. When a prompt works, save it as a template with blanks for the subject and setting, so the next scene starts from something proven. Our free AI video prompt templates are a starting point.
4. Learn each model. Keep notes on what each one does well. (See the 2026 model ranking.)
5. Keep motion believable. Subtle, purposeful movement reads better than spectacle.

Less is sometimes more

Don't stuff a prompt with twenty adjectives. A few precise, film-literate choices outperform a wall of vague descriptors, and contradictory words ("bright moody noir daylight") confuse the model. Say exactly what the shot is and let the model render it.

How to do it in Cinemagiq

1. Start from a picture you like. On the Generation Studio's Image tab, Image to Prompt turns any still (an upload, a drag and drop, or an Asset Hub image) into an editable prompt you can adapt.
2. Save what works. The project's Prompt Library stores prompts and templates with {{placeholders}}, so a proven shot description can be reused across scenes and by the whole team.
3. Set the angle on stills. On the Nano Banana image models, Camera Angle Control adds presets such as Low Angle, High Angle, Rear or Top-Down to your prompt. It steers the model rather than guaranteeing the result, so check the output.
4. Build prompts where you work. On a character, location or shot, Build Prompt helps you assemble a prompt for it and enhance it before you generate.
5. Let the platform handle model quirks. For Seedance models, Cinemagiq automatically rewrites your prompt into the structure ByteDance recommends.

For how prompting fits into the whole process, see our text-to-video guide and the complete AI filmmaking guide.

A project-level prompt library and per-shot model choice make this fast in Cinemagiq.