The difference between a good AI video and a bad one is almost never the tool. It is the prompt. After generating thousands of clips for Fruit Love Island, we have a system that works across Grok, Kling, Veo, and Runway. Here it is.
The Five-Layer Prompt Structure
Every effective AI video prompt has five layers, in this order. Skip any layer and the output gets worse.
Layer 1: Shot Type
Tell the model what kind of shot you want before describing anything else. This is the single most impactful word in your prompt.
- Close-up — face and emotions, best for dialogue scenes
- Medium shot — waist up, good for conversations between two characters
- Wide shot — full body and environment, good for establishing scenes
- Over-the-shoulder — looking past one character at another, creates intimacy
- Tracking shot — camera follows a character, adds movement
Layer 2: Subject Description
Describe who or what is in the frame. Be specific about appearance but do not overload. Three to five physical details is the sweet spot.
Too vague:
A woman standing in a room
Too much:
A 28-year-old woman with auburn hair in a messy bun, green eyes with gold flecks, wearing a vintage 1940s silk blouse in dusty rose with pearl buttons, high-waisted navy trousers, brown oxford shoes, a silver locket around her neck...
Just right:
A young woman with red hair in a messy bun, wearing a pink silk blouse and navy trousers, standing in a sunlit kitchen
Layer 3: Action
What is happening in this shot? Use present tense. One action per prompt. If your character is doing two things, make two clips and cut them together.
Bad (too many actions):
She walks to the table, picks up a letter, reads it, looks shocked, and drops it
Good (one clear action):
She picks up a letter from the table and reads it, her expression shifting to shock
Layer 4: Environment and Lighting
Where is this happening and what does the light look like? Lighting cues have a massive impact on mood. The model responds well to cinematography terms.
- Golden hour lighting — warm, romantic
- Harsh overhead fluorescents — clinical, uncomfortable
- Candlelight — intimate, dramatic shadows
- Backlit silhouette — mysterious, cinematic
- Overcast diffused light — neutral, documentary feel
Layer 5: Style and Mood
End with the overall aesthetic. This is where you can reference film styles, color palettes, or emotional tones.
Example modifiers:
Cinematic, shallow depth of field, warm color grade, handheld camera movement, film grain
A Complete Prompt Example
Putting all five layers together:
Full prompt:
Close-up of a young woman with red hair in a messy bun, wearing a pink silk blouse, picking up a letter from a wooden kitchen table and reading it with a shocked expression. Sunlit kitchen, golden hour light streaming through a window. Cinematic, warm tones, shallow depth of field.
That is 45 words. It covers shot type, subject, action, environment, and style. It will produce a usable clip on any major AI video tool.
Platform-Specific Tips
Grok Imagine
Responds well to dialogue in prompts. You can include what a character is saying and the video will include speech. Keep dialogue short — one to two sentences max per clip. Longer dialogue gets garbled.
Kling 2
Excels at motion. If your scene involves walking, running, dancing, or physical action, Kling handles it better than competitors. Use their “Professional” mode for more precise prompt following.
Veo 3
Understands camera direction terminology best. Terms like “dolly in,” “crane shot,” and “rack focus” produce accurate results. Also the best at generating synchronized audio from prompt descriptions.
Runway Gen-4
Strong at maintaining visual consistency across multiple generations from similar prompts. Best for sequences where you need several shots of the same scene from different angles.
Common Mistakes
- Writing a screenplay, not a prompt: AI video models want visual descriptions, not narrative prose. “Maria felt a wave of nostalgia” means nothing to a video model. “A woman looks out a rain-streaked window, eyes glistening” works.
- Prompting for specific text: AI video models cannot reliably render text on signs, screens, or objects. If you need text in frame, add it in post-production.
- Ignoring aspect ratio: A 9:16 vertical video needs different composition than 16:9 horizontal. Mention orientation in your prompt or set it in the tool’s settings.
- Contradicting the reference image: If you upload a reference of a blonde character but prompt for “dark hair,” the model gets confused. Align your prompt with your reference.
- Forgetting camera movement: Static shots feel lifeless. Even a subtle “slow push-in” or “gentle handheld sway” adds life to the clip.
The Prompt Library Approach
The fastest way to improve is to build a library. Every time a prompt produces a great result, save it. Copy the exact wording into a document organized by shot type. Next time you need a similar shot, start from your library instead of from scratch.
We keep a spreadsheet with columns for: prompt text, tool used, quality rating (1–5), and a screenshot of the result. After 50 entries, patterns emerge. You learn which words each tool responds to best. That knowledge compounds.
The rule of thumb: If you are writing more than 60 words, your prompt is too long. If you are writing fewer than 20, it is too vague. The sweet spot is 30–50 words covering all five layers.