How to Write AI Video Prompts: Lessons from 12,000 Examples

How to Write AI Video Prompts: Lessons from 12,000 Examples

Text-to-video went from a novelty to a real production tool in about a year. Wikiprompt now holds more than 12,000 working video prompts across Seedance, Veo 3, Hailuo, Kling, Sora, Grok Imagine, Runway and Wan. We read through the best of them to answer one question: what separates a prompt that produces a flat, drifting clip from one that produces something you would actually publish?

The patterns below are not theory. They are the habits that show up again and again in the prompts our community rates highest.

1. Describe the shot, not just the scene

The single biggest difference between amateur and professional video prompts is that the good ones direct the camera. A beginner writes "a knight fighting a dragon." A working prompt writes a shot: a continuous one-take, a match cut, a locked-off static frame, a triptych.

Look at Cinematic One-Shot: The Girl and the Stagehands or the Ultra-Cinematic Continuous One-Shot Momentum Chain. Both name the camera move as the star of the clip, and the model rewards them with coherent motion instead of a slideshow. The Cinematic Triptych Portrait with Car goes further and specifies the composition itself.

If you take one thing from this guide: write like a cinematographer, not like a caption.

2. Give it a duration and a cut list

Video models drift. The longer the clip, the more the subject wanders off-model. The prompts that hold together tell the model exactly how long the clip is and how many cuts it contains.

Cinematic fire-poi ritual video (4 cuts, 15s) is the template: a fixed length, a fixed number of shots, and a one-line description per cut. The One Table: A Whole Life 15s Emotional Brand Film uses the same structure to compress an entire narrative into fifteen seconds. Breaking a clip into numbered beats keeps the model on rails and gives you a storyboard you can edit shot by shot.

3. Name the camera and the era to lock a look

The fastest way to a distinctive aesthetic is to reference a real recording format. "Cinematic" is generic. "Early-2000s DV camcorder footage" is a look the model can actually reproduce, grain, interlacing, auto-exposure hunting and all.

The community leans on this hard: Early-2000s DV Camcorder Footage of the Great Pyramid Being Built, the Handheld MiniDV Backstage Vlog of a K-pop Idol, Early-2000s Korean Home-Video Camcorder Footage, and the Y2K Skater Video Dodge Sequence. Found-footage realism is one of the strongest genres in text-to-video right now precisely because the format does half the work.

4. Ask for audio on purpose

Newer models generate sound, but only if you request it. The Cinematic Sci-Fi Mystery Sequence with Native Stereo Audio spells out the soundscape as its own line. If you leave audio implicit, you get silence or generic filler; name the ambient sound, the foley, and the score and you get a finished clip instead of a moodboard.

5. Match the model to the job

Not every model is good at everything. Based on how our community actually uses them:

Great starting points by use case

Product and brand films: Carrot Pro: Apple-Style Product Launch Film, Golden Crispy Kingdom KFC Commercial, Ultra-Realistic Luxury Beauty Commercial (9:16).

Action and anime: Ultra-Cinematic Anime Swordsman, One Move to Defeat the Enemy, Rooftop Static 15s Anime Fight.

Food and macro: Ultra-Realistic Cinematic Food Video.

Image-to-video: Turn a Static Poster into a 15-Second Cinematic Teaser shows how to animate a still you already have.

A quick recipe you can copy

Put these five lines at the top of any video prompt and fill in the blanks:

Everything else, mood, color, lens, is refinement on top of that skeleton. Browse the full video prompt collection to see how the community fills it in, or run any of these directly in Create and iterate from there.