AI Video Prompt Guide: Structure, Motion, and Model Choice

At minimum, a strong video prompt answers five questions:

• Who or what is the subject, and how do they look at the start?
• Where is the scene, including foreground and background cues that matter?
• What happens during the clip, in an order the duration can support?
• How does the camera see it: shot size, angle, and movement?
• What is the atmosphere: light, weather, mood, and, if the tool supports it, sound?

Most tools also need delivery choices outside the prose prompt: duration, aspect ratio, and whether you are starting from text alone or from an image. Those controls are model- and interface-dependent. Check the product or API settings for the clip you are generating instead of treating one tool's limits as universal.

No prompt structure guarantees realistic physics, perfect continuity, readable on-screen text, or a usable first take. Treat the first generation as a draft, then change one variable at a time.

This guide uses SCAAL: Subject, Camera, Action, Atmosphere, and Length. It is a practical checklist created for GetBetterPrompts, not an industry standard. Official docs use related ideas in different orders. Google's Veo prompting material groups cinematography, subject, action, context, and style.

Runway's prompting guides emphasize clear visual and motion descriptions. Kling's official text-to-video guide uses subject, movement, scene, and optional camera, lighting, and atmosphere. SCAAL is a portable starting point; each model and interface interprets the same words differently.

Subject: [who/what, appearance, starting position]
Camera: [shot size, angle, movement or locked-off]
Action: [primary motion, then any follow-on events]
Atmosphere: [light, weather, mood, optional sound]
Length: [duration target, pacing, delivery ratio]

Skip empty parts. If the tool already locks duration or aspect ratio in the UI, put those choices there and keep the text focused on what the setting cannot express.

Text-to-video must describe both appearance and motion. Runway's text-to-video guidance treats visual description and motion description as the two essential elements. Use it when you do not need a fixed starting frame, or when you want the model freer to invent composition.

Image-to-video starts from a still that already defines composition, subject, lighting, and style. The prompt should mostly describe what should move: subject action, environmental motion, camera motion, timing, and speed. Re-describing every visual detail can fight the image. Official Kling image-to-video guidance centers on subject plus movement for the same reason.

Choose image-to-video when identity, product look, or layout must stay close to a reference still. Choose text-to-video when you are exploring a shot and do not yet have a reliable first frame. Either path still benefits from SCAAL; only the balance between visual prose and motion prose changes.

Name the subject with concrete traits: age range or type, clothing, materials, color, count, and starting pose or position. "A woman in a red coat standing at the edge of a wooden pier" is easier to control than "a person near water."

Describe only the scene details that affect the shot. Foreground objects, surfaces, weather, and light sources matter when they change reflections, depth, or mood. Long inventories of props often compete for attention inside a short clip.

Keep subject text stable when you generate related angles of the same person or product. Change camera and action between takes. Stability helps, but it is not a guarantee of perfect identity match unless the tool's reference or identity features are available and you use them as documented.

State one primary action the duration can finish. Then add follow-on events only if the clip length and model can support them. Natural language sequences such as "X happens, then Y, then Z," or rough timestamps, are documented approaches in current Runway Gen-4.5 prompting materials. Complex choreography can work, but crowded simultaneous motion is harder to control.

When results feel chaotic, simplify before you invent more adjectives. Prefer readable cause-and-effect: a subject turns, wind moves fabric, rain hits pavement. Mechanical instructions like "rotate 45 degrees" can look robotic unless you want mechanical motion.

Image-to-video prompts should not contradict strong motion cues already in the still. Blur, mid-stride poses, or dust trails imply motion. Fighting those cues usually needs a cleaner source image, not a louder prompt.

Camera language separates a video prompt from a still-image prompt. Start with shot size and angle: extreme wide, wide, medium, close-up, low angle, high angle. Then add movement only if you want it: slow dolly in, tracking, orbit, crane, handheld, or locked-off tripod.

Simple motion is often easier to diagnose. If a shot fails, reduce to one clear camera move plus one subject action, then rebuild complexity. That is a control strategy, not a universal platform limit. Current Runway Gen-4.5 documentation explicitly supports complex sequenced instructions and detailed camera choreography inside a single prompt.

If you need a still camera, say so in natural language and still describe what moves inside the frame. Video models are biased toward motion; "static background" and "locked-off camera" help when the subject should carry the energy.

Duration is usually a product setting, not only a sentence in the prompt. Limits change by model and surface.

Examples from current official docs, not universal facts: Gemini API Veo 3.1 documentation describes 8-second outputs; Runway Gen-4.5 documents 2 to 10 seconds on that product surface; Kling's official text-to-video guide discusses 5-second and 10-second generations for the workflow it covers. Always read the control panel or API field for the tool you are using.

Match the action to the clock. A walk across a room needs more time than a head turn. If you ask for several beats in a short clip, expect compression, cuts, or missed steps. Longer durations help sequenced actions; shorter durations favor one beat.

Name pacing when it matters: real-time, slow motion, quick cut energy, lingering hold. Pair pacing words with concrete motion so the model is not guessing what "cinematic" means for timing alone.

Light and weather do double duty: they set mood and they create visible motion opportunities, such as shifting sun breaks, fog drift, or neon reflections on wet streets. Prefer concrete cues over lone adjectives. "Eerie fog in an abandoned hospital corridor, flickering fluorescent lights" is more actionable than "scary mood."

Style anchors help when the wrong era or genre appears: film stock, decade, documentary handheld, clean digital cinema, commercial product lighting. Put the highest-priority mood early if mood is critical, then keep the rest of the prompt consistent with that choice.

Where a model supports native audio, you can direct ambience, effects, and quoted dialogue in the same brief. Support and quality vary by model and by product versus API access. Do not assume every video tool can speak, lip-sync, or score the shot.

Continuity fails for ordinary reasons: rewritten subject lines, conflicting wardrobe details, different aspect ratios, or prompts that invent new props between takes. Copy the subject block exactly when you only mean to change camera or action.

Reference images, ingredients, start and end frames, and extension features are model-dependent. On the Gemini API, current Veo 3.1 documentation describes image-based direction with up to three reference images, first-and-last-frame workflows, and extension of previously generated Veo clips.

Other tools offer their own reference and frame controls. Product apps and APIs do not always expose the same options.

References improve the odds of consistency. They do not guarantee identical faces, logos, or product geometry across every take. For critical brand work, plan to select, crop, and composite in an editor.

Set aspect ratio before you generate. The model composes for the frame you request. Cropping afterward throws away intended edges.

Common delivery targets:

• 16:9 for landscape web, presentations, and many desktop players
• 9:16 for vertical social placements
• 1:1 for square feeds and centered product or talking-head clips

Supported ratios and resolutions differ by tool. Current Veo 3.1 Gemini API docs list landscape 16:9 and portrait 9:16 among generation options. Runway Gen-4.5 documents multiple ratios on its product surface, including 16:9, 9:16, and 1:1.

Kling's official text-to-video guide lists 16:9, 9:16, and 1:1 for the workflow it describes. Confirm the live settings for your account rather than memorizing any one list as permanent.

Pick a workflow by the job, not by a quality leaderboard. Unsupported rankings hide the real differences: duration, audio, references, extension, frame control, editing loop, and whether you are in a consumer app or an API.

Google video models (Gemini API guidance): Google's current video-generation documentation recommends Gemini Omni Flash as the default for video generation, highlighting multi-input reasoning, conversational editing, and related strengths. It points to Veo 3.1 when you need capabilities such as scene extension, last-frame control, or legacy-pipeline integration. Treat that as Google's documented split for its own family, not as a claim that either model beats every other vendor.

Runway Gen-4.5: Official product docs describe text-to-video and image-to-video, sequenced instructions, camera choreography, and duration and aspect-ratio controls on that surface. It is a strong fit when you want an editing-oriented generation loop with explicit motion direction.

Kling: Official Kling prompting guides teach structured subject, movement, and scene writing for text-to-video, and subject-plus-movement writing for image-to-video. Capability details change across Kling versions and surfaces, so verify duration, audio, and reference features in the product you actually use. This guide does not rank Kling against other tools.

Sora (historical): OpenAI discontinued the Sora web and app experiences on April 26, 2026, and states that the Sora API will be discontinued on September 24, 2026. If you are migrating an old Sora workflow, rebuild the brief with SCAAL and test on the tool you choose now. Do not assume any one model is a drop-in replacement.

App UIs and APIs can differ even inside one vendor. A feature visible in a studio product may be missing, preview-only, or named differently in the API. Read the surface you are calling.

For Gemini-specific access and product-family detail, use the site's Gemini guide. Keep this page focused on portable video-prompt craft.

Weak text-to-video: "A cool video of someone on a pier at sunset."

Improved text-to-video: "Medium shot of a woman in a red wool coat standing at the edge of a wooden pier. She turns toward camera as wind lifts her hair. Slow dolly in. Overcast late afternoon, soft diffused light, grey water behind her. Real-time pacing, melancholic mood. 16:9."

What changed: a specific subject, a single readable action, one camera move, atmosphere that implies lighting, and a delivery ratio. The model has fewer gaps to invent.

Weak image-to-video (given a product still of a matte black earbud case on marble): "Make it cinematic and premium."

Improved image-to-video: "Keep the earbud case fixed on the marble. Slow 180-degree orbit at eye level. Soft studio key light from above, subtle reflection crawl on the matte surface. No new props. Locked product scale. 1:1."

What changed: the prompt stops redesigning the product and instead directs motion, camera, and light behavior that the still does not already encode.

Flicker, jitter, or morphing limbs: Too many simultaneous motions. Reduce to one subject action and one camera move. Say which elements stay still.

Ignored camera move: Reinforce the angle or move in plain language and remove conflicting moves. Iterate rather than stacking synonyms.

Identity drift across clips: Reuse the exact subject paragraph, keep ratio and style anchors identical, and use reference or frame controls when the tool provides them.

Unwanted cuts or rushed beats: The duration may be too short for the sequence, or the prompt implies an edit. Lengthen the clip, simplify the beat list, or add a continuous-shot instruction if you need one take.

Wrong genre or era: Add stronger style anchors and remove mixed style words that fight each other.

Garbled text or logos: Most video models are weak at readable typography. Add titles and logos in post whenever the message must be legible.

Judge success by the visible clip against your checklist. Do not ask the model to reveal hidden chain-of-thought or internal reasoning as a verification method.

Before you hit generate:

• Subject appearance and start position are specific
• Scene details are limited to what the shot needs
• Primary action fits the selected duration
• Camera size, angle, and movement are explicit or intentionally open
• Atmosphere and style do not contradict each other
• Text-to-video includes visuals plus motion; image-to-video focuses on motion
• Aspect ratio matches the destination
• Reference or frame controls are set if continuity matters
• Audio directions are included only when the chosen model supports them
• You know which one variable you will change if the first take misses

Write video prompts like short shot briefs: subject, camera, action, atmosphere, and length. Choose text-to-video or image-to-video based on whether you need a fixed first frame. Select models by workflow needs such as editing loop, extension, references, or audio, not by unsupported rankings. Compare the output to your checklist, change one variable, and repeat.

When you want help turning a rough idea into a clearer video brief, use the free video prompt tool on GetBetterPrompts.