AI video prompt guide: what actually worked
We ran 11 one-sentence prompts through Kling 3.0, Veo 3.1 Fast, PixVerse V6 and Runway Gen-4.5 and scored every clip blind. Here's each prompt, the winning clip, and the lesson behind it.
The 6 rules our data supports
- Lead with shot size and camera. "Close-up", "drone glide" and "camera pulls back" were the most reliably followed words.
- Name the one key moment. "Catches it mid-air" beats "playing with a frisbee". Models default to the easy version.
- One or two beats per clip. Three-step physics failed on every model.
- Repeat atmosphere you care about. Rain and confetti were the details most often dropped.
- Quote dialogue and say who speaks. Then pick a model with native audio (which ones).
- For vertical video, say "full subject in frame". 9:16 crops tight.
| # | Prompt | Best | Score | Lowest |
|---|---|---|---|---|
| 1 | Café laugh in the rain | Veo 3.1 Fast | 8.3 | Runway Gen-4.5 7.3 |
| 2 | Dog catches frisbee | Kling 3.0 & PixVerse V6 (tie) | 7.8 | Veo 3.1 Fast 5.8 |
| 3 | Flaming wok | Kling 3.0 & Veo 3.1 Fast & PixVerse V6 (tie) | 7.8 | Runway Gen-4.5 7.0 |
| 4 | Cyberpunk drone shot | Veo 3.1 Fast | 7.8 | PixVerse V6 5.8 |
| 5 | Park bench dialogue | Veo 3.1 Fast | 7.5 | Kling 3.0 6.3 |
| 6 | Milk into coffee | Runway Gen-4.5 | 8.3 | Veo 3.1 Fast 6.5 |
| 7 | Kickflip down stairs | Runway Gen-4.5 | 6.8 | Kling 3.0 5.5 |
| 8 | Origami crane takes flight | Kling 3.0 | 7.5 | Runway Gen-4.5 6.3 |
| 9 | Sneaker product spin (9:16) | Veo 3.1 Fast | 8.0 | PixVerse V6 7.0 |
| 11 | Knight & dragon reveal | Runway Gen-4.5 & PixVerse V6 (tie) | 7.3 | Kling 3.0 6.5 |
| 12 | Concert crowd & confetti | Veo 3.1 Fast | 7.5 | Runway Gen-4.5 6.8 |
Café laugh in the rain
Close-up of a woman laughing at a café window, rain streaking the glass, shallow depth of field, handheld.
Lesson: Shot size words work. "Close-up" was followed by Kling and Veo, while Runway went medium-wide with an extra person. If framing matters, state it first.
Veo 8.3 · Kling 7.8 · PixVerse 7.5 · Runway 7.3
adherence 9motion 7quality 9artifacts 8
Tight close-up, rain-streaked glass in foreground, natural laugh; subtle handheld drift.
Dog catches frisbee
A golden retriever catching a frisbee mid-air on a beach at sunset, slow motion.
Lesson: Name the exact moment ("catches it mid-air"). Veo and Runway showed the dog with the frisbee but skipped the catch. Only Kling and PixVerse showed it.
Kling 7.8 · PixVerse 7.8 · Runway 6.0 · Veo 5.8
adherence 9motion 8quality 7artifacts 7
Frisbee flies in, dog leaps and catches it; slightly flat colour.
Flaming wok
Chef flipping vegetables in a flaming wok, sparks and steam, kitchen ambient sound.
Lesson: Action verbs get softened. "Flipping" became stirring on Veo and a flame burst with no flip on Runway. Add what should be airborne ("vegetables tossed into the air").
Kling 7.8 · Veo 7.8 · PixVerse 7.8 · Runway 7.0
adherence 9motion 8quality 8artifacts 6
Real toss with vegetables airborne and flames; some morphing between beats.
Cyberpunk drone shot
Drone shot gliding over a neon cyberpunk city at night, flying cars, light rain.
Lesson: Weather details get dropped. Only Veo showed visible rain. Repeat important atmosphere ("heavy rain streaks, wet reflections").
Veo 7.8 · Kling 7.5 · Runway 6.3 · PixVerse 5.8
adherence 9motion 7quality 8artifacts 7
Rainy neon city, drone glide, flying vehicle; slightly repetitive.
Park bench dialogue
Two people having a conversation on a park bench; she says "I told you it would work.".
Lesson: For dialogue, ask for a close two-shot and quote the line. Only Veo clearly showed speech.
Veo 7.5 · PixVerse 7.0 · Runway 6.5 · Kling 6.3
adherence 8motion 6quality 8artifacts 8
Close two-shot, she clearly speaks; natural expressions.
Milk into coffee
A hand pouring milk into black coffee, the swirl in macro, top-down.
Lesson: Camera angle words ("top-down") were only partly followed. Runway’s swirl was the most beautiful but angled. Accept a near miss or regenerate.
Runway 8.3 · Kling 8.0 · PixVerse 8.0 · Veo 6.5
adherence 8motion 8quality 9artifacts 8
Beautiful swirl in a mug; angled rather than top-down.
Kickflip down stairs
Skateboarder doing a kickflip down stairs, tracking shot.
Lesson: Multi-step physics (flip + stairs + landing) failed everywhere. Split it into shorter shots or simplify the action.
Runway 6.8 · Veo 6.0 · PixVerse 6.0 · Kling 5.5
adherence 7motion 6quality 8artifacts 6
Board flips, landing on flat ground; stairs barely part of the trick.
Origami crane takes flight
Paper origami crane unfolding itself and flying away, stop-motion style.
Lesson: Style words like "stop-motion" mostly stuck on Kling. Veo and PixVerse produced a clean finished crane instead of the unfolding. Describe the process, not just the end state.
Kling 7.5 · Veo 6.8 · PixVerse 6.8 · Runway 6.3
adherence 8motion 7quality 8artifacts 7
Unfolds from flat into a crane and takes off; stop-motion feel.
Sneaker product spin (9:16)
A TikTok-style product shot: a sneaker rotating on a pedestal, studio lighting, 9:16.
Lesson: Vertical product shots work well. Add "full product in frame". Runway’s rotation was smooth but cropped the shoe at the edges.
Veo 8.0 · Runway 7.8 · Kling 7.3 · PixVerse 7.0
adherence 8motion 7quality 9artifacts 8
Premium studio lighting; minor shape drift during rotation.
Knight & dragon reveal
Medieval knight walking through a snowy forest, then the camera pulls back to reveal a dragon (multi-beat shot).
Lesson: Multi-beat shots need explicit camera language ("camera pulls back to reveal"). Runway did a true pull-back, Veo and PixVerse cut to a wide shot, and Kling showed the dragon from the start.
Runway 7.3 · PixVerse 7.3 · Veo 7.0 · Kling 6.5
adherence 9motion 7quality 7artifacts 6
Knight walks toward camera, camera pulls back, dragon reveal; dragon a bit cartoonish.
Concert crowd & confetti
A crowd at a concert, hands up, confetti falling, lead singer on stage.
Lesson: Crowds and hands are now mostly coherent, but small requested details (confetti) can arrive late or not at all. Put them early in the prompt.
Veo 7.5 · Kling 7.3 · PixVerse 7.3 · Runway 6.8
adherence 9motion 7quality 8artifacts 6
Hands up, heavy confetti, singer on stage; hands mostly coherent.
All prompts word for word: prompt library. Related: Physics test · Lip-sync test · Best AI video generator · Methodology
FAQ
How do you write a good AI video prompt?
In our tests, prompts worked best when they stated shot size and camera move first, named the single key moment of action, quoted any dialogue exactly, and kept to one or two beats for a 4–5 second clip. Atmosphere details like rain were often dropped unless repeated.
How long should an AI video prompt be?
Our 11 test prompts were one sentence each (about 15–25 words), and the best results scored above 8/10. Longer is fine (Runway Gen-4.5 accepts up to 1,000 characters), but every extra beat competes for a 4–10 second clip.
Can I copy these prompts?
Yes. All 20 prompts are in our prompt library, word for word, so you can rerun them on any tool and compare against our clips.
Which AI video generator follows prompts best?
Veo 3.1 Fast had the highest average prompt-adherence score in October, ahead of Kling 3.0, PixVerse V6 and Runway Gen-4.5, but each model won some prompts. See the per-prompt table on this page.