How to Write AI Video Prompts That Actually Work (2026)

Why Most AI Video Prompts Fail

You typed "a cool video of a dog in a park." The model gave you a dog, sort of, in a park, sort of — drifting across the frame while the camera lurched in three directions at once. The problem was not the model. It was the prompt.

A video model must fill in every gap you leave: what the subject looks like, what moves, where the camera goes, how the light falls, what happens first. A few seconds of footage contains hundreds of those decisions. Leave them blank and the model guesses — differently every time. This guide shows you the formula professionals use, the five rules that fix most bad prompts, and how to adapt your prompting for Veo 3.1, Kling, and Runway.

The 7-Layer Prompt Formula

Every strong video prompt answers the same seven questions, in roughly this order. If your clip looks wrong, a missing layer is usually why.

1. Shot and framing. Wide establishing shot, medium shot, close-up, aerial. Name it in film terms.
2. Subject. Who or what is on screen, described concretely. "A woman in her 30s with curly hair, wearing a red raincoat" beats "a woman."
3. Action. One visible movement, with a strong verb. She "strides across the street" — not "moves energetically."
4. Setting. Where and when, with spatial detail. "A narrow neon-lit street in Tokyo at night" beats "a city."
5. Camera. One camera move, with a speed. Slow push-in, locked-off tripod, handheld tracking shot.
6. Light and look. Lighting direction, palette, genre. Golden-hour backlight, high-contrast noir, soft studio light.
7. Audio. On models that generate sound, name what you hear and what you do not: dialogue in quotes, ambient sound, "no music" when you mean it.

Google's Veo 3.1 prompting guidance covers the same ground in five parts — cinematography, subject, action, context, and style and ambiance — with audio as a separate direction.

5 Rules That Fix Most Bad Prompts

1. One action per clip

Most AI clips last only a few seconds. Cramming a character walking, sitting, waving, and laughing into one prompt guarantees mush. Give the model one visible event per generation. If your scene needs more, generate separate shots and stitch them together, or use a multi-shot mode.

2. One camera move per clip

"Zoom in while panning left and tilting down" reads like ambition. To the model, it is three conflicting instructions. Pick the single most important move and add a speed modifier: "slow dolly in," "fast handheld pan right." Without a speed word, the model picks an arbitrary default.

3. Use concrete verbs, not adjectives

Adjectives like "beautiful" and "dynamic" give the model nothing to render. Verbs do. "Pours," "sprints," "unfolds," "orbits," "flickers" each describe something visible. "Turns and lifts the cup" beats "looks happy."

4. Put timing in the prompt

When something must happen in sequence, say when: "For the first three seconds the mug sits still; then steam rises and a hand reaches in from frame left." Short timestamp cues such as [00:01] and [00:03] are documented in Runway's prompting guidance.

5. Say what you do NOT want

Negative constraints prevent common artifacts: warped hands, morphing faces, stray text, watermarks. End prompts with a short line such as "No text, no logos, no watermark." Keep it short and lead with the positive instruction — Runway's guidance favors positive descriptions of motion over negative lists for image-to-video.

Text-to-Video vs Image-to-Video: Prompt Differently

Text-to-video prompts must describe the whole scene — subject, setting, lighting, camera — because there is no starting image. This is the mode for exploring a new shot from scratch.

Image-to-video prompts should describe only what changes. The image already defines composition, lighting, style, and subject. Repeating what is visible wastes the prompt and can reduce movement. If your start frame shows a woman in a red jacket on a beach, do not re-describe the jacket — write "she slowly turns her head toward the camera, hair drifting in the wind." Runway's Gen-4 guidance is explicit: describe motion, not the picture. Image-to-video wins whenever a specific face, product, or location must stay consistent.

Model Cheat Sheets: Veo 3.1, Kling, and Runway

Veo 3.1 (Google)

Veo 3.1 is the realism and audio leader: it generates synchronized dialogue, sound effects, and ambient audio natively, which few competitors match in one pass. Write dialogue in quotes, name ambient sounds, and keep short clips to one scene — Google's guidance warns against chaining several distinct events into one prompt. Use reference images ("Ingredients to Video") to lock a character or product across shots.

Kling (Kuaishou)

Kling is strongest on smooth human motion and dialogue-driven clips. Its official VIDEO 3.0 guidance documents 3 to 15 seconds per generation, so write one self-contained action rather than a whole story. If you need more than one framing change, use Multi-Shot or Custom Multi-Shot instead of stuffing five camera directions into one prompt. Kling's newer releases add keyframe and reference systems — see our breakdown of Kling 4.0's 30-second clips and keyframes.

Runway (Gen-4 / Gen-4.5)

Runway rewards clarity over length: one clear scene-and-motion prompt, one camera beat per generation, and short proof runs before a long render. For image-to-video, describe subject motion, camera motion, scene response, and motion style — never restate visible details. Its documented duration range for Gen-4.5 is roughly 2 to 10 seconds. For the wider landscape, browse our directory of the 90+ best AI video generator tools in 2026.

3 Copy-Paste Example Prompts

Steal these structures. Swap in your own subject and setting; keep the camera, lighting, and timing parts.

Example 1: Product ad (Veo 3.1 or Kling)

"Close-up of a matte-black insulated water bottle on a wet stone table at dawn. Condensation rolls down the bottle. A hand reaches in from frame left, lifts the bottle, sets it back down. Camera: slow push-in. Light: cool morning side light. Look: high-contrast product photography. Sound: faint birdsong, soft clink of bottle on stone, no music. 6 seconds, 16:9. No text, no logos."

Example 2: Cinematic scene (Runway or Veo)

"Wide establishing shot of a lighthouse on a rocky cliff at dusk, waves breaking below. A lone figure in a yellow raincoat walks slowly along the cliff path toward the lighthouse. Camera: slow aerial drift left to right. Light: stormy golden-hour sky, deep blue shadows. Look: cinematic, anamorphic lens, film grain. 8 seconds, 16:9. No text."

Example 3: Talking head (Veo 3.1 or Kling)

"Medium shot of a friendly chef in her 40s, apron over a denim shirt, in a bright home kitchen. She looks into the camera and smiles. She says: 'The secret is cold butter — always cold butter.' Camera: locked-off tripod. Light: soft window light from the left. Sound: her voice, faint pan sizzle off-screen, no music. 8 seconds, 16:9."

All three share one action, one camera move, concrete verbs, a duration, and a short exclusion line. That is the whole system.

Debugging: When the Output Looks Wrong

Do not rewrite the whole prompt when one thing is off. Change one element per iteration — the prompt, the source image, or the model — so you know what fixed it. Start cheap: test at lower resolution or a short duration before spending credits on a full-quality render.

Symptom	Likely cause	Fix
Subject morphs or changes face	No visual anchor; too many actions	Use image-to-video or a reference image; cut to one action
Camera jerks or spins	Multiple or vague camera moves	Name one move + speed: "slow push-in"
Motion freezes or feels dead	Abstract action ("looks happy")	Use a visible verb ("turns and smiles")
Weird hands, text, artifacts	No constraints; too many elements	Simplify the scene; add a short "No…" line
Wrong mood, flat lighting	Style words overpowering scene	Move style after the action; name light direction

Save every prompt that works. Build a swipe file of winning prompts with the model, settings, and output attached. Next time you need a similar shot, you edit a proven prompt — not a blank box. For turning clips into finished content, see our guide to Google Flow, Google's AI video studio, and Vidu Q4's photo-to-cinema workflow.

Verdict: Prompt Like a Director, Not a Wisher

AI video models are not mind readers — they are literal directors working from your shot list. The creators getting clean, usable clips in 2026 are not using magic words. They write one action, one camera move, concrete verbs, timing, and a short exclusion line, then iterate one change at a time. That discipline transfers across every model.

FAQ

How long should an AI video prompt be?

Long enough to cover shot, subject, action, setting, camera, light, and audio — usually 60 to 150 words. Clarity beats word count.

Do I need different prompts for Veo, Kling, and Runway?

The core structure is the same, but each model has quirks. Veo 3.1 wants dialogue in quotes and audio direction. Kling wants one self-contained action per 3–15 second clip, with Multi-Shot for multiple framings. Runway wants motion-only prompts for image-to-video.

Why does my character's face change between shots?

Text alone does not lock identity. Use image-to-video from a reference still, or your tool's reference/Elements system, and keep the same character description word-for-word across shots. Check identity across the full clip, not just the first frame.

Can I use negative prompts in AI video tools?

Yes, in most tools — a short "No text, no logos, no watermark" line helps. But lead with positive instructions, especially in Runway image-to-video, where official guidance favors describing the motion you want.