How to Write Prompts for AI Video (Sora, Veo, Runway & More)

AI video is magical when it works — and baffling when it doesn’t. The difference is almost always the prompt. Video prompting has its own grammar: you’re directing motion, camera, and time, not just describing a still image. Here’s how to write prompts for AI video that actually work.

The Video Prompt Formula

[Subject + action] + [Scene/environment] + [Camera movement] + [Lighting/mood] + [Style] + [Technical specs]

Weak: “a car driving”
Strong: “A vintage red convertible driving along a coastal highway at sunset, ocean on the left, camera tracking alongside the car at road level, golden-hour light with lens flare, cinematic film grain, 24fps, photorealistic, 5 seconds”

The strong version directs like a filmmaker: what’s happening, where, how the camera moves, what it feels like. Every unspecified element is a dice roll.

The Layers That Matter Most

1. Subject + Action (Be Verb-Specific)

“Walking” is vague. “Striding confidently,” “strolling,” “sprinting” — verbs carry motion information the model needs. Describe how things move, not just that they move. Multiple subjects? Specify each one’s action, or the model will improvise.

2. Camera Movement (The Video Superpower)

This is what separates video prompts from image prompts. Use film language:

• Static: “locked-off tripod shot” — stable, reliable.
• Tracking/dolly: “camera glides alongside,” “slow push-in.”
• Aerial: “drone shot ascending,” “bird’s-eye view descending.”
• Handheld: “subtle handheld shake” for realism.

Full guide: controlling camera movements in AI video prompts.

3. Time and Pacing

Mention duration and pace: “slow-motion,” “time-lapse,” “gradual transition from day to night.” Without pacing cues, the model picks arbitrarily — and “arbitrary” often means “weird.”

4. Style and Realism Level

“Photorealistic,” “cinematic,” “anime,” “vintage film.” Also specify what to avoid: “no morphing faces, no extra limbs, smooth motion” — video models still struggle with these.

Tool-Specific Tips

• Sora (OpenAI): excellent at following detailed narrative prompts; write like a shot description.
• Veo (Google): strong realism; conversational refinements work well via Gemini.
• Runway/Pika: great image-to-video — start from a still you like, then describe the motion. Often more controllable than pure text-to-video.

For choosing between tools: which tools generate video from text.

Common Failures (and Fixes)

• Morphing/warping: keep prompts simpler; complex multi-action scenes break coherence. One clear action per shot.
• Weird physics: “natural movement,” “realistic physics” help marginally. Shorter clips fail less.
• Text in video: avoid requesting text/signage — it garbles.
• Inconsistent characters: use reference images (image-to-video) instead of describing the same person twice.

The Shot-List Approach (Pro Workflow)

Don’t try to generate a 60-second masterpiece in one prompt. Plan 4-6 shots (5 seconds each), generate them separately with consistent style words, then edit together. Consistent style descriptors across shots (“cinematic, golden hour, 35mm film”) sell the illusion that it’s one continuous piece. AI video right now is a shot generator, not a film generator — work with that.

Genre-Specific Prompting Tips

• Cinematic/narrative: “cinematic, 24fps film look, shallow depth of field” + emotional direction (“a sense of longing”). Think like a director describing the shot to a cinematographer.
• Commercial/product: “studio lighting, clean background, slow orbit around product, premium feel.” Simplicity sells — one product, one move, perfect light.
• Social/UGC style: “smartphone footage aesthetic, natural lighting, authentic feel.” Ironically, prompting for less polish works for platforms that reward authenticity.
• Abstract/artistic: “flowing particles, morphing colors, meditative pace.” Abstract subjects hide AI’s weaknesses (no faces, no physics to get wrong).
• Nature/wildlife: “golden hour, telephoto lens compression, slow motion.” Nature is AI video’s comfort zone — fewer hard edges to warp.

Sound Design: The Missing Half

AI video is silent, and silence reads as “unfinished.” Fix it:

• Music: match energy to visuals — AI music tools (see music prompting) can score your clips.
• Sound effects: freesound.org and similar libraries — footsteps, wind, water transform perceived realism.
• Voiceover: AI narration over AI visuals — the full synthetic production pipeline.
• The rule: viewers forgive visual imperfections with good audio; they never forgive bad audio with good visuals. Spend your effort accordingly.

Building a Video Prompt Library

Save what works: the camera phrases that deliver, the style combos that sing, the negative prompts that prevent disasters. Video prompting is young enough that shared knowledge is gold — and your personal library of proven prompts becomes a genuine competitive edge as the tools mature. The directors of the AI era won’t be the ones with the best tools; they’ll be the ones with the best prompts.

From Prompt to Final Video (Full Workflow)

Writing the prompt is one step in a pipeline. 1. Concept: decide the story beat — one clear idea per clip. 2. Prompt draft: write it using the techniques in this guide. 3. Generate 3-5 variations: never settle for the first output; variation is where quality hides. 4. Select and refine: pick the best, note what to fix, adjust the prompt, regenerate. 5. Edit: assemble clips in an editor, add sound, color grade. AI video is raw material, not finished product. 6. Sound design: the most underrated step — good audio makes mediocre video watchable; bad audio ruins great video. Budget equal time for sound and visuals. The prompt gets you 40% of the way; selection, editing, and sound get you the rest. For camera-specific prompting, see controlling camera movements.

If you want to go deeper, OpenAI’s prompt engineering guide is the definitive free resource — written by the people who built the models.

Bottom Line

How to write prompts for AI video: direct like a filmmaker — specific action verbs, explicit camera movement, pacing, style, and what to avoid. Keep shots short and simple, use image-to-video for control, and assemble multiple shots for longer pieces. The models are improving fast; the prompting fundamentals aren’t changing.