AI Video Prompt Guide: Write Prompts That Work

What Makes a Good AI Video Prompt?

A good AI video prompt answers five questions in a single, ordered sentence block: what is the subject, what is it doing, how is it filmed, where does it happen, and how should it look. Video models like Runway Gen-3, Luma Dream Machine, Kling, and Pika respond far better to structured, concrete language than to loose descriptions. A prompt such as "a woman walking through rain-soaked Tokyo streets at night, handheld camera following behind her, neon reflections on wet asphalt, cinematic, shallow depth of field" will outperform "cool video of a woman in Tokyo" almost every time.

The core difference from image prompting is motion. Every video prompt should contain at least one explicit motion instruction, either for the subject ("she turns her head slowly") or for the camera ("slow dolly-in"). If you leave motion undefined, the model chooses it for you, and the result is often subtle drifting or warping that reads as low quality.

The Five-Block Prompt Structure

Use this order every time. Consistency makes prompts editable: if a clip fails, you know which block to change.

1. Subject block: who or what, with one or two visual details. "A gray heron with folded wings standing in shallow water."
2. Action block: one clear motion, with speed or manner. "It spreads its wings slowly and lifts off."
3. Camera block: shot type plus movement. "Low-angle tracking shot, camera tilts up as it takes flight."
4. Scene block: environment, time of day, weather. "Misty lake at sunrise, fog over the water."
5. Style block: look and mood. "Shot on 35mm film, muted colors, soft natural light, cinematic."

Keep the whole prompt under about 60 words for most models. Overly long prompts dilute attention, and models tend to drop details from the middle of long sentences. Put the most important element first, because most video models weight the opening words more heavily.

Camera Language That Actually Changes Output

Video models are trained partly on footage with embedded metadata and captions, so real cinematography vocabulary has measurable effects. Use these terms deliberately:

• Shot types: wide shot, medium shot, close-up, extreme close-up, over-the-shoulder, aerial view, macro.
• Camera movements: dolly in, dolly out, pan left/right, tilt up/down, tracking shot, orbit (360-degree), crane shot, handheld, static shot.
• Lens and focus: shallow depth of field, deep focus, telephoto compression, fisheye lens, rack focus from foreground to background.
• Frame rate terms: slow motion, 120fps, time-lapse, hyperlapse.

One movement per prompt is the rule. Combining "camera orbits while dollying in and tilting up" usually produces jittery, indecisive motion. If you need a complex move, generate the clip in one clean motion and plan a cut, not a compound camera instruction.

Model Differences: Same Prompt, Different Results

Models do not interpret prompts identically, and knowing the differences saves hours of trial and error.

Practical takeaway: write a base prompt, then keep a per-model variant file. When you switch tools, you usually only need to rewrite the camera and style blocks, not the subject and action blocks.

Image-to-Video vs. Text-to-Video Prompting

Text-to-video gives the model full control; image-to-video gives you control of composition and only asks the model for motion. Most professional workflows use image-to-video for anything that needs a specific look.

When prompting image-to-video, do not re-describe the image. The model can already see it. Instead, prompt only what should change: "the woman turns toward the camera and smiles, hair moves in the wind, camera pushes in slightly." Re-describing static elements often causes the model to regenerate and distort them.

When prompting text-to-video, you carry the full descriptive load, so all five blocks matter. Expect to iterate more: two to four generations per shot is normal even with a good prompt.

A Repeatable Prompting Workflow

Follow these steps to go from idea to usable clip without randomizing your way there:

1. Write the shot list first. Break your video into single-action shots of 3–8 seconds. One clip equals one action.
2. Draft each prompt in the five-block order (subject, action, camera, scene, style), under 60 words.
3. Generate a test batch at low settings if available. Generate 2–3 variations per prompt.
4. Diagnose failures by block. Wrong subject? Fix block one. No motion or weird motion? Fix block two or three. Bad look? Fix block five. Change one block at a time.
5. Lock the style block. Once you have a look that works, reuse the identical style wording across every shot for visual consistency.
6. Save winning prompts to a library. Reusable prompt skeletons are your most valuable asset across projects.

Common Prompting Mistakes to Avoid

• Negative-style instructions. Most video models do not handle "no blur, not shaky" well. State what you want, not what you don't.
• Too many simultaneous actions. "She runs, jumps, and waves" produces mush. One action per clip.
• Abstract mood words as the whole prompt. "Ethereal, dreamy, emotional" does nothing without a concrete subject and motion attached to it.
• Ignoring duration limits. Asking a 5-second model for a 20-second scene forces the model to slow or loop motion. Plan cuts instead.
• Copying image prompts verbatim. Image prompts optimize composition; video prompts must also specify time. Always add motion before reusing an image prompt.

Treat every prompt as a shot direction you would give a camera operator. If a human cinematographer could not film your sentence, the model probably cannot either.

Example Prompts You Can Adapt

Nature documentary look: "A red fox steps through deep snow, sniffing the ground, static camera at ground level, telephoto lens, shallow depth of field, golden hour light, snowy forest background, shot like a BBC nature documentary."

Product shot: "A matte black wristwatch rotating slowly on a reflective surface, macro lens, studio lighting, dark gradient background, camera orbits 90 degrees, high-end commercial style."

Narrative scene: "A man in a raincoat stops at a crosswalk and looks up at the sky, medium shot, camera pushes in slowly, rain falling, city street at dusk, moody teal-and-orange color grade, cinematic."

Key Takeaways

Effective AI video prompting is structural, not magical: subject, action, camera, scene, style, in that order, under 60 words, with exactly one motion per clip. Learn real camera vocabulary because models understand it, adapt your style block per model rather than rewriting everything, and iterate one block at a time so failures tell you exactly what to fix. Build a prompt library as you go, and within a dozen generations you will have reusable skeletons that make every future video faster and more predictable.