From a one-line idea to a cinematic AI shot: a practical prompt tutorial

If you have generated AI video from a short phrase like "cinematic drone shot of a city," you already know the usual result: a pretty but generic clip that could belong to anyone. The model is not failing so much as filling in every decision you left out—lens, movement, lighting, pacing, and what the subject is actually doing.

This tutorial is a practical way to write prompts that behave more like a shot list than a slogan. The same structure works for text-to-video, for animating a still, and for planning a short multi-shot sequence. The examples below are ready to copy, then tighten.

I used this workflow on KlingVideo, a free Kling AI video generator that turns text prompts, reference images, and motion-control inputs into cinematic clips with Kling 3.0 plus Seedance, Veo, Wan, Hailuo, Grok and more. The habits below are model-agnostic; they are just easier to practice when you can try the same brief in more than one engine.

A simple shot structure

Before you write a paragraph of style words, lock five decisions:

1. Subject — who or what is on camera, including a distinguishing wardrobe or facial detail.
2. Setting — place, time of day, weather, and a few tangible props.
3. Action — a single, visible verb that can play out in a few seconds, not a whole plot.
4. Camera — shot size, lens feel, and one camera move.
5. Light and grade — key-light direction, contrast, and color temperature.

Add an optional sixth line for constraints: no morphing faces, no jumpy cuts, no extra people walking through frame. Write the five decisions as short clauses, then combine them. If a first take looks messy, change one clause at a time.

Example 1: text-to-video, one controlled move

A weak prompt usually sounds like this:

Cinematic video of a woman walking in the rain, highly detailed, 4k, masterpiece.

A stronger prompt specifies one action and one camera job:

A woman in her early 30s with a short black bob and a mustard raincoat walks toward camera down a narrow Tokyo side street at night. Neon pharmacy and vending-machine signs reflect in shallow puddles. She holds a transparent umbrella slightly tilted against the rain. Camera: slow dolly backward at chest height, 35mm look, shallow depth of field, raindrops catching street light in the foreground. Lighting: wet asphalt, mixed neon magenta and cool fluorescent spill. Realistic motion, no jump cuts, no extra pedestrians crossing in front of her.

This works because the action is one verb (walk toward camera), the camera has one job (dolly back), and the environment has specific, light-catching surfaces: puddles, neon, an umbrella. "Cinematic" is implied by the shot, not used as a magic adjective.

Example 2: animate a still without breaking the face

Image-to-video is less about inventing a new scene and more about telling the model how little should change. Start from a clear portrait or mid-shot still. Restate the person and wardrobe so the model does not "improve" them, then add a small, physically plausible motion.

Keep the same woman as in the reference image: mustard raincoat, short black bob, transparent umbrella. She stands in the same alley. Subtle motion only: she inhales, her eyes shift toward a neon sign on the left, and a breeze tugs the raincoat hem. Rain continues to fall with the same intensity. Camera locked off, no reframing, no wardrobe change, maintain facial identity and lighting from the still.

If the face drifts, shorten the action further—eyes only, or hair only—and keep the camera static. Animating a still is usually more convincing when the background is allowed a little life, such as rain, steam, or a flickering sign, while the subject barely moves.

Example 3: a three-shot sequence with a consistent character

You can plan a short scene as three prompts that share a character description and change only shot size and action. This is useful when you want to keep characters consistent across clips.

Paste a short character bible into every shot:

Same character throughout: Japanese woman, early 30s, short black bob, mustard raincoat, small silver hoop earrings, transparent umbrella. Night, rainy Tokyo side street, neon reflections on wet pavement.

Then vary only the coverage:

• Shot A, wide: She appears at the far end of the alley and walks closer, umbrella up. Slow push-in from a wide 24mm view. Hold the environment: vending machines, hanging laundry, one cat under a shop awning.
• Shot B, medium: She stops under a pharmacy sign, tilts the umbrella back, and checks a phone in her left hand. Camera at chest height, almost static, slight handheld feel.
• Shot C, close-up: Rain beads on the umbrella; she looks up, and a neon highlight moves across her cheek. Tight 85mm look, extremely shallow focus, no cutaways.

Keep continuity boring on purpose: same clothes, weather, and time of night. A coat-color change will read as a different person, even when the face is close.

Example 4: describe reference movement clearly

When you transfer reference movement from a clip—someone turning, a product orbit, a handheld follow—say which part of the motion you want, and what should not come along for the ride.

Transfer the camera path and walking cadence from the reference clip. Replace the location with the rainy Tokyo alley described above and dress the subject in the mustard raincoat. Keep the original timing and stride. Do not copy the reference wardrobe, faces, or background. Smooth stabilization, no morphing mid-stride.

Motion-control inputs help most when the reference is clean: one subject, even lighting, and a single obvious move. A chaotic reference with lots of cuts will usually transfer the chaos as well.

Iterate like an editor, not like a gambler

A practical loop looks like this:

1. Generate a first pass with a complete five-part prompt.
2. Watch it once for story (did the action happen?) and once for craft (did the camera do what you asked?).
3. Change a single variable: lens, move, or light.
4. If identity slips, add a short character bible and reduce the motion.
5. Keep the take only when the motion and face both hold.

New users get free starter credits, which is enough to learn this loop on a handful of short clips. Paid plans unlock more capacity if you later need it. On supported exports, there is no watermark, so the version you share can match the take you approved.

A reusable template

Copy this and fill the brackets:

[Subject, including one wardrobe detail] in [setting, time, weather]. They [one visible action]. Camera: [shot size], [lens feel], [one move]. Lighting: [direction and color]. Style: naturalistic motion, consistent identity. Constraints: [what must not change].

The habit to unlearn is stacking empty praise words. The habit to practice is making one photographic decision per line. Do that, and even a short AI clip starts to feel directed. If you want a single place to try the same prompt with Kling 3.0 and the other models mentioned above, KlingVideo is a straightforward starting point.