Seedance 2.0 prompt engineering: how to write AI video prompts that actually work

You type "a cat sitting on a rooftop at sunset" and expect a Miyazaki-level scene. What you get is a blurry blob on a flat rectangle with an orange gradient behind it. The gap between what you picture in your head and what AI actually generates almost always comes down to one thing: the prompt.

AI video generators have gotten remarkably good at turning text into motion. But they're not mind readers. The models respond to specific patterns of language, and if you don't speak that language, you'll burn credits re-rolling the same mediocre output over and over.

This guide covers the prompt patterns that actually produce good results across today's major generators, including Seedance 2.0, Runway, Sora, Pika, and Kling.

TL;DR

• Structure every prompt with four parts: subject, action, environment, and style
• Use real camera terminology ("slow dolly forward") instead of vague descriptions ("zoom in")
• Front-load the most important details in your first sentence
• Keep prompts to 2-4 sentences; longer is not better
• Pick one visual style and commit to it
• Limit your scene to 1-2 subjects maximum
• Each platform has quirks; adjust your prompting strategy accordingly

The anatomy of a good video prompt

Every strong video prompt has four building blocks. You don't need to hit all four every time, but the more you include, the more control you have over the output.

Subject - Who or what is in the scene. Be specific about appearance, clothing, position, and posture. "A woman" gives the model nothing to work with. "A tall woman in a long black coat, silver hair pulled back, standing at the edge of a bridge" gives it a clear target.

Action - What's happening. Use specific verbs and describe the pacing and direction of movement. "Walking" is weak. "Striding forward with purpose, coat billowing behind her" tells the model how the movement should feel.

Environment - Where it's happening. Describe lighting conditions, time of day, weather, and setting details. "A city" could be anything. "A rain-soaked bridge over a canal at dusk, streetlights reflecting off wet cobblestones" puts the model in a specific place.

Style - How it should look. This covers cinematography, color grading, artistic style, and overall mood. Without this, you're leaving the aesthetic entirely up to the model's defaults.

Here's a prompt that uses all four:

A tall woman in a long black coat walks across a rain-soaked stone bridge
at dusk. Streetlights reflect off wet cobblestones. Slow dolly forward
following the subject, shallow depth of field. Cinematic, desaturated
teal and amber color grading, 35mm film grain.

That prompt is three sentences. It gives the model a subject, an action, an environment, and a style. That's all you need.

Camera language that AI understands

Camera direction is where most people leave the biggest quality gains on the table. If you don't specify how the camera moves, the model picks a static medium shot by default. That's boring.

Here are the camera terms that reliably work across generators:

Movement:

• slow dolly forward - works much better than "zoom in" (dolly is physical camera movement, zoom is lens adjustment; models understand the difference)
• tracking shot following subject from left to right
• orbiting around subject at eye level
• drone shot ascending over city
• steady push-in toward subject
• handheld camera, slight shake - great for documentary or found-footage feels

Angle:

• low angle looking up at subject - makes subjects look powerful or imposing
• overhead establishing shot - good for showing spatial relationships
• eye-level medium shot - neutral, conversational framing
• extreme close-up on hands - directs attention to detail

Lens effects:

• shallow depth of field, subject in focus, background blurred
• rack focus from foreground object to subject
• anamorphic lens flare
• wide angle distortion at the edges

The key insight: use the language a cinematographer would use. These models were trained on video descriptions that use professional terminology, so professional terminology gets better results.

Style and mood descriptors that work

Style descriptors tell the model what visual world your video lives in. Think of them as filters, but ones that affect the entire generation process rather than just a post-processing layer.

Film stocks and formats:

• 35mm film grain - adds organic texture, reduces that AI "clean" look
• anamorphic widescreen - horizontal lens flares, cinematic aspect ratio feel
• 8mm home video - nostalgic, degraded quality
• 8K, hyperrealistic - pushes the model toward photorealism

Lighting:

• warm golden hour lighting - the 30 minutes before sunset, everything glows
• cool blue moonlight - nighttime scenes with visible detail
• harsh overhead fluorescent - uneasy, institutional feeling
• volumetric light through fog - dramatic shafts of light, adds depth

Artistic styles:

• Studio Ghibli style - soft colors, detailed backgrounds, animation
• cyberpunk neon aesthetic - dark with saturated color accents
• documentary style, natural lighting - handheld, observational
• noir, high contrast shadows - deep blacks, dramatic highlights
• watercolor painting style - soft edges, blended colors

Color grading:

• desaturated teal and orange - the Hollywood blockbuster look
• warm vintage tones, lifted blacks - film photography aesthetic
• monochrome with selective red accents - Sin City style
• pastel color palette - soft, dreamy

You can combine these. "35mm film grain, warm golden hour lighting, desaturated teal and orange color grading" is a valid and effective style block. Just don't mix styles that contradict each other.

Motion and action descriptions

Video is about movement. A prompt that describes a still image will generate a video that looks like a still image with slight ambient motion. You need to describe what's moving and how.

Speed and pacing:

• slow motion or 120fps slow motion - emphasizes dramatic moments
• time-lapse - compresses hours into seconds, good for clouds, crowds, construction
• normal speed, fluid motion - default realistic pacing
• quick, energetic cuts - faster pacing (works better on some platforms than others)

Subject motion:

• Describe the arc of movement, not just the action. "Running" is a start. "Running toward camera, gradually slowing to a stop" gives the model a beginning, middle, and end.
• Include secondary motion: hair blowing, fabric rippling, dust rising behind footsteps. These details sell the realism.
• Specify direction: "walking from left to right" prevents the model from guessing.

Environmental motion:

• leaves drifting across frame
• clouds moving rapidly overhead
• waves crashing against rocks in the foreground
• traffic flowing in the background, headlights streaking

Environmental motion adds life to scenes even when your main subject is relatively still. A person standing in a field is static. A person standing in a field with wind rippling through tall grass and clouds drifting overhead is alive.

Common mistakes and how to fix them

1. Too vague

Bad:  a person walking
Good: A woman in a red dress walking down a rainy Tokyo street at night,
      neon reflections shimmering on wet pavement, medium tracking shot,
      warm neon glow against cool shadows

The bad prompt gives the model almost nothing. The good prompt specifies the subject, action, environment, lighting, camera, and mood.

2. Too many subjects

Cramming five characters into one prompt overwhelms most generators. The model tries to render all of them and none look right. Stick to 1-2 subjects per generation. If you need a crowd, describe it as an environment element ("busy marketplace with people in the background") rather than trying to specify each individual.

3. No camera direction

If you don't tell the model how the camera should move, it defaults to a static shot or random slow drift. Always include at least one camera instruction.

Bad:  A spaceship flying through an asteroid field
Good: A spaceship flying through an asteroid field, camera tracking
      alongside the hull, asteroids tumbling past in foreground and
      background, dramatic side lighting, anamorphic lens

4. Conflicting styles

"Realistic anime" or "cinematic cartoon" sends mixed signals. The model tries to satisfy both and the result is neither. Pick one visual style and commit to it. If you want anime, go full anime. If you want photorealism, don't add illustration keywords.

5. Prompt too long

Some people write 500-word essays as prompts. Most generators have an effective attention window, and details buried in paragraph five get ignored. Keep prompts to 2-4 sentences. Front-load the most important information. Subject and camera go first, environment and style follow.

Platform-specific tips

Each generator has its own strengths and quirks. Here's what works best on each.

Seedance 2.0

The biggest advantage of Seedance 2.0 is the reference system. Instead of trying to describe everything in text, you can upload reference images for character appearance, composition, and visual style, plus reference videos for camera movement and motion patterns.

This means your text prompts can be shorter and focused on what the references don't cover. If you've uploaded a reference image of your character and a reference video with the camera move you want, your text prompt only needs to describe the action and mood.

For beat-sync mode, describe the energy and mood rather than specific cuts. "High energy, fast-paced, dynamic camera movements matching the beat" works better than trying to choreograph each transition in text.

Runway

Runway's Motion Brush is its standout feature. You can paint directly on the frame to indicate where movement should happen and in what direction. This means your text prompt can focus on the content and style rather than trying to describe complex motion paths.

Gen 4.5 responds well to concise prompts with strong style keywords. Keep text descriptions focused on the look and feel; let Motion Brush handle the movement.

Sora

Sora responds well to longer, more descriptive prompts compared to other generators. It handles cinematic language particularly well. Full sentences that read like a screenplay description tend to produce good results.

Sora is also strong at understanding spatial relationships, so describing where elements are relative to each other ("a cat perched on the left windowsill, city lights visible through the glass behind it") pays off here.

Pika

Short and punchy works best on Pika, especially in Turbo mode. One to two sentences. Front-load your subject and action. Pika's strength is speed, so lean into that by keeping prompts tight.

For best results, focus on a single clear action and let the model handle the rest. Overloading Pika prompts with style descriptors tends to muddy the output.

Kling

Kling excels at human subjects. If your scene includes people, spend extra prompt space on facial expressions and body language. "Smiling softly while tilting head slightly to the left, eyes narrowing with warmth" produces noticeably better results on Kling than generic descriptions.

For longer-form outputs (Kling supports up to 2 minutes), break your story into emotional beats and describe the progression: "begins with a serious expression, slowly breaking into a wide grin as confetti falls."

Prompt examples: before and after

Here are side-by-side comparisons showing how small changes in prompting produce dramatically different results.

Example 1: Animal scene

Bad:  a dog running

Good: Golden retriever running through shallow ocean waves at sunset,
      slow motion, water droplets catching golden light, wide angle
      tracking shot, warm amber tones

The bad prompt produces a generic dog jogging in an undefined space. The good prompt creates a specific cinematic moment with clear lighting, camera work, and atmosphere.

Example 2: Sci-fi environment

Bad:  futuristic city

Good: Aerial drone shot descending into a dense cyberpunk city at night,
      neon signs in Japanese, flying cars drifting in background layers,
      rain streaking across frame, blade runner color palette,
      anamorphic lens

"Futuristic city" could mean a clean utopia or a gritty dystopia. The good prompt picks a specific vision and commits to it with concrete visual details.

Example 3: Dance performance

Bad:  person dancing

Good: Young woman in a flowing white dress performing contemporary dance
      in an empty warehouse, dust particles visible in shafts of light
      from high windows, slow dolly orbiting subject, 35mm film grain,
      muted earth tones

The bad prompt doesn't specify the type of dance, the setting, the lighting, the camera, or the mood. The good prompt creates a complete scene that any generator can work with.

Example 4: Nature close-up

Bad:  flowers blooming

Good: Macro time-lapse of a pale pink peony unfurling its petals,
      water droplets on the edges catching light, dark background,
      shallow depth of field, soft natural side lighting

Specificity wins every time. Name the flower. Describe the camera technique. Set the lighting. These details are what separate scroll-past content from stop-and-stare content.

FAQ

How long should an AI video prompt be?

Two to four sentences. Most generators have an effective context window for prompts, and details beyond that get progressively less attention. Put your most important information first: subject, action, and camera. Style and environment details follow.

Does prompt engineering work the same across all AI video generators?

The fundamentals are the same everywhere. Subject, action, environment, and style are universal building blocks. But each platform has different strengths. Sora handles longer descriptions well; Pika prefers short prompts; Seedance 2.0 lets you offload visual descriptions to reference images. Adjust your strategy per tool.

Should I use negative prompts for video generation?

Some platforms support negative prompts (describing what you don't want). When available, use them sparingly for specific issues: "no text overlays, no watermarks, no sudden camera jumps." Don't fill negative prompts with long lists; focus on the problems you're actually seeing in your outputs.

How do I maintain character consistency across multiple video clips?

This is one of the hardest problems in AI video. Your options: use reference images of the same character for each generation (Seedance 2.0's multi-reference system is built for this), keep your character descriptions word-for-word identical across prompts, or generate all clips in one session and use video extension features to build longer sequences from a single starting point.

What's the most common beginner mistake?

Being too vague. Most first-time users write prompts like "a beautiful landscape" and wonder why the output is generic. The fix is specificity. Name the location, the time of day, the weather, the camera angle, and the color palette. You're directing a scene, not making a wish.

Start prompting better

Good prompt engineering is a skill, and like any skill, it gets better with practice. Start with the four-part framework: subject, action, environment, style. Add camera direction. Be specific. Cut what's unnecessary.

The generators are only getting better. Your prompts should too.

Try these techniques on your next generation at Seedance 2.0 and see the difference a well-structured prompt makes.