Prompt Engineering for AI Video: Real Examples That Work in Runway, Pika, and Sora

Prompt engineering for video differs from prompting text or image models in an important way: you’re not just describing what something looks like, you’re describing an event that unfolds through time. A good text prompt might describe a scene; a good video prompt describes a shot. That shift in thinking changes what details you include and how you structure them.

This guide is built around real example prompts and the specific reasoning behind why they work. Copy, adapt, and test these directly in Runway Gen-3, Pika 1.5, or Sora.

The Core Framework: Subject + Action + Environment + Camera + Style

Every effective video prompt contains most of these elements, roughly in this order:

[Camera shot/movement], [subject description and action], [environment and context], [lighting], [technical/style]

The order matters more than you might expect. Most video models give higher weight to concepts earlier in the prompt. Camera and shot type first means you get a coherent framing. Subject after that anchors the motion. Style elements at the end are treated as modifiers rather than instructions.

Category 1: Nature and Environmental Scenes

These are generally the most reliably successful category because AI video models are trained on enormous amounts of landscape, weather, and nature footage.

Working prompt:

Aerial drone shot, dense redwood forest stretches to the horizon, morning fog fills the valleys between trees, slow forward dolly, golden morning light from the right, cinematic, Arri Alexa

Why it works: “Aerial drone shot” immediately constrains the viewport and explains why we’re seeing the tops of trees. “Morning fog fills the valleys” gives the model a specific motion task (fog movement) that it handles well. “Slow forward dolly” prevents the jittery random motion that appears when you don’t specify. Arri Alexa signals professional exposure characteristics.

Variation for different mood:

Ground-level shot looking up through pine forest canopy, shafts of light break through treetops, slow tilt down, blue hour, realistic, 8K

Category 2: Urban and Street Scenes

Dynamic urban content — city streets, neon reflections, crowd movement — is a strength of Runway Gen-3 in particular.

Working prompt:

Eye-level tracking shot follows two people walking away from camera along a narrow alley, lanterns glow overhead, rain-wet cobblestones, evening, neon signs in the distance, Tokyo, shallow depth of field, anamorphic

Why it works: “Tracking shot follows” gives the model a stable motion instruction. “Walking away from camera” is easier for the model than walking toward the camera (face detail is the hardest part to maintain). The specific environmental details — lanterns, cobblestones, neon signs — give the model rich material to render.

For Pika specifically, try adding the style reference last:

Eye-level tracking shot, woman walks away through a narrow alley in rain, neon reflections on wet ground, lanterns overhead, Tokyo street, cinematic, moody color grade — negative: blurry, distorted, low quality

Category 3: Product and Commercial Shots

This is where image-to-video mode (available in Runway, Pika, and Luma Dream Machine) excels. Start with a high-quality product image from Midjourney or Stable Diffusion, then animate.

For a product float/rotate:

(upload product image) slow orbit around product, clean white studio background, soft diffused lighting with gentle rim light, smooth 360 rotation, professional product photography

For a lifestyle shot:

(upload product image) product in outdoor kitchen setting, natural morning light, gentle steam rises from coffee mug, slow push-in toward product, cinematic

The key insight for product work: keep the motion instruction minimal. The more you ask the model to move things, the more likely it is to distort your product. “Gentle steam” and “slow push-in” are safe; “camera spins dramatically while product floats” will produce chaos.

Category 4: Abstract and Artistic Motion

Abstract content is an underexplored category where models perform well because there’s no “correct” answer — failures are less visible.

Working prompt (Sora):

Macro shot, vivid ink drops bloom in slow motion through crystal-clear water, deep cobalt and gold pigments, black background, extreme slow motion, photorealistic, 8K

Working prompt (Luma Dream Machine):

Time-lapse of storm clouds forming over ocean, lightning flashes illuminate the undersides of dark cumulus clouds, wide shot, dramatic, high contrast

Working prompt (Runway):

Abstract fluid simulation, iridescent liquid metal surface in motion, ripples expand and interfere, overhead camera looking straight down, seamless loop, clean black background

Category 5: Character and Portrait Work

This is the hardest category. Be conservative in your expectations and specific in your instructions.

For characters in motion (medium shot, not close-up):

Medium shot, a man in a grey coat stands at a train platform at night, looks left then right as a train passes behind him, motion blur on train, still camera, moody street lighting, cinematic

What makes this work: “Medium shot” keeps faces small enough that flaws aren’t obvious. “Looks left then right” is a simple, slow motion the model can execute. The train passing behind is good background interest that doesn’t require the model to render detailed faces on multiple people.

For portrait-forward content: Use HeyGen instead of Runway or Pika. HeyGen is a specialised AI avatar platform that builds on the same underlying models but is trained specifically for face detail, lip sync, and talking-head video. It starts at $29/month and produces face quality that general-purpose video generators can’t match.

Prompt Structures for Specific Use Cases

YouTube intro/b-roll:

Cinematic b-roll, [action/scene relevant to topic], smooth motion, professional lighting, warm tones, slow motion, 4K

Social media (TikTok/Reels — vertical format):

Vertical 9:16 format, [subject and action], close-up perspective, bright and saturated, smooth camera movement, trending aesthetic

Background loop for presentations:

Seamless loop, [abstract or environmental scene], slow and minimal motion, no cuts, clean and professional

Atmospheric opener:

Establishing shot, [location], [time of day], [weather], slow drone pullback reveals scale, no people, cinematic, [colour palette]

What to Do When Prompts Fail

The result is completely wrong: Shorten the prompt radically. Remove everything except subject and one motion descriptor. Once you get something vaguely correct, add detail back iteratively.

The motion is too fast or jerky: Add these exact phrases: slow motion, smooth camera movement, gentle. Remove any word that implies speed: “rushing”, “dynamic”, “fast”.

Objects disappear mid-clip: This is temporal inconsistency from too long a generation or too complex a scene. Reduce to 4–5 seconds, simplify the scene, and try again.

The style is wrong: Runway responds well to direct film stock references (Kodak Vision3 film, Fujifilm Velvia, 16mm film). Pika responds better to aesthetic labels (dark academia, cottagecore, cyberpunk). Sora is most responsive to descriptive language about lighting and physics.

Building a Prompt Library

The most efficient approach to prompt engineering is iterative: when a prompt produces excellent output, save it immediately with the exact wording. Over time, you’ll build a set of 15–20 base prompts across categories that reliably produce good results, and you can adapt these for new projects rather than starting from scratch each time.

Tools like Notion or even a simple text file work well for this. Organise by category (nature, urban, product, abstract) and note which platform the prompt was tested on — a prompt that excels in Runway may produce mediocre results in Pika due to differences in how the models interpret stylistic language.