How to Write Good AI Video Prompts: A Practical Framework (2026)
A good AI video prompt describes seven things in a fixed order: subject, motion, camera, composition, lighting, style and timing. Write one line for each, put the most important for your model first, and keep the whole thing under about 80 words. That structure — not clever wording — is what separates cinematic output from the mush most people get.
Key takeaways
- Use the seven-element framework: subject → motion → camera → composition → lighting → style → timing.
- Order matters more than vocabulary. Models weight the opening of a prompt heavily.
- Show the camera, don’t just name the scene. “Slow push-in on…” outperforms “cinematic shot of…” every time.
- The same prompt produces different results across Runway, Kling and Sora — adapt per model.
- You don’t have to start from a blank page: extracting prompts from videos that already work is the fastest way to learn good structure.
The seven-element framework
Every video prompt answers the same seven questions. If one is missing, the model improvises — and its improvisation is usually generic.
| # | Element | The question | Example fragment |
|---|---|---|---|
| 1 | Subject | Who or what is in the shot? | a surfer in a yellow wetsuit |
| 2 | Motion | What does the subject do? | duck-diving under a breaking wave |
| 3 | Camera | How does the camera behave? | slow motion, tracking shot from the side |
| 4 | Composition | Where is the subject in frame? | low angle, subject against open sky |
| 5 | Lighting | What is the light doing? | backlit at golden hour, hard rim light |
| 6 | Style | What is the aesthetic? | 35mm film, slight grain, warm grade |
| 7 | Timing | How does it pace? | 4-second clip, single continuous take |
A complete prompt built from the table:
A surfer in a yellow wetsuit duck-dives under a breaking wave, tracking shot from the side in slow motion, low angle against open sky, backlit at golden hour with hard rim light, 35mm film with slight grain and warm grade, 4-second single continuous take.
Nineteen words of structure beat sixty words of adjectives. Notice what is not there: “cinematic masterpiece”, “ultra quality”, “8K” — filler that consumes your prompt budget without directing a single pixel.
Writing order vs. reading order
List elements in the order above and models do best. But position carries weight: if your video lives or dies on the camera move, put the camera clause at the very front (Slow push-in on a surfer…). The opening ~20 words are the strongest signal most models give.
Before and after
Weak prompt (what most people write):
Cinematic AI video of a surfer riding a wave, high quality, 4K, beautiful lighting, viral style.
The same intent, seven elements applied:
Surfer in a yellow wetsuit carving down a wave face, tracking shot from a boat parallel to the break, low horizon line with spray filling the lower frame, backlit golden-hour sun through the spray, 35mm film look with warm grade, 5-second clip cutting on the turn.
What changed: every adjective was replaced with a direction. “Beautiful lighting” told the model nothing; “backlit golden-hour sun through the spray” produced a specific image because it is a specific instruction.
Adapting to your model (this is where most prompts fail)
The framework is universal; the dialects differ. The same seven-element prompt goes into Runway, Kling and Sora and comes back as three different videos — because each model was trained to weight different vocabulary.
Practical differences you will hit:
- Runway rewards concise camera language and responds well to motion-bracket syntax (
camera: slow push-in). - Kling handles longer natural-language descriptions and is strong at subject motion — spend your words on action.
- Sora/Veo respond to narrative context — a sentence about why the scene happens can improve coherence.
The full side-by-side analysis is in why the same prompt differs across Runway, Kling and Sora. The rule of thumb: structure stays, vocabulary bends.
The fastest way to get better: learn from videos that already work
Blank-page practice is slow. Two habits that compress the learning curve:
- Extract, then study. Take videos you admire — viral clips, ads, transitions — and extract their prompts. Reading a hundred structurally correct prompts teaches the grammar faster than writing twenty. Start with how AI understands a video to see what the analyzer is actually reading.
- Keep a style library. Save extracted prompts tagged by mood and use-case instead of saving links. Our visual style library method shows the filing system.
There is a reason video prompt engineering has become its own discipline: prompt skill is the multiplier on every model subscription you pay for.
Seven mistakes that quietly ruin good prompts
- Stacking quality tags (
8K, masterpiece, best quality) — leftovers from image models; video models spend tokens on them instead of direction. - No camera clause — the single most common omission. The scene is described; the filming is not.
- Two subjects, no hierarchy — if both matter, say which moves and which holds.
- Burying the lede — your key instruction in word 60 gets less weight than in word 5.
- Negative prompts everywhere — useful for specific artifacts (text, watermark), counterproductive as a paragraph.
- Ignoring duration — a prompt that reads like a 30-second sequence in a 5-second clip produces rushed, cut-happy output.
- Never iterating — one generation is a draft. Change exactly one element per retry so you know what worked.
The ninth-tenths version of this list is 7 common mistakes in AI video production, including the one 90% of creators make.
A reusable template
Copy this and fill the blanks:
[Camera move] on/around [subject + action],
[composition], [lighting], [style/grade],
[duration + pacing].
Example filled in:
Slow orbital push-in around a chef plating dessert,
centered composition with steam rising, soft window
light from the left, warm 35mm film grade,
5-second continuous take.
Ten seconds to fill, seven elements guaranteed — that is the whole framework.
FAQ
How long should an AI video prompt be?
Aim for 40–80 words: long enough to cover all seven elements, short enough that no element gets diluted. If you’re past 100 words, you’re probably stacking adjectives — cut directionless ones.
What makes a video prompt “good” technically?
Three properties: every element present (no model improvisation), direction stated at the front (highest-weight position), and vocabulary that matches your target model’s dialect. Everything else is aesthetic preference.
Do negative prompts help in video?
Sparingly. Use them to suppress specific artifacts (no on-screen text, no watermark). A long negative paragraph competes with your positive description for attention.
Should I write prompts in English even for Chinese output?
For most models, yes — training data is English-heavy, so English prompts are more reliable even when the desired on-screen language is Chinese. Add the required on-screen language as an explicit instruction (dialogue in Mandarin, English subtitles).
Can AI write the prompt for me?
Yes — feeding a reference video to a video-to-prompt generator produces a structurally complete draft; your job shifts from writing to editing. See how to extract a prompt from any video.
How do I know my prompt is good?
Run it three times. Consistent, directed output with the camera behavior you asked for = good prompt. Wildly different results = the model is improvising, which means something in the seven elements is missing or buried.
Written by the SocialToPrompt Content Team. Updated October 2026. Practice the framework on socialtoprompt.com — extract prompts from real videos with free signup credits.
Share Article