The complete prompt engineering guide for AI video, 2026.

A senior creator's prompt looks almost nothing like a beginner's. It is longer, more specific, written in the vocabulary of cinema rather than the vocabulary of search. Most importantly, it works the first time. This guide is a transfer of that vocabulary — the structures, words, and patterns we've found that hold up across Veo 3.1, Sora 2, Kling v3, Wan 2.7 and Seedance 2.0.

The universal prompt structure

Every video prompt that consistently lands across our pipeline shares the same skeleton. Memorize this structure and you'll outperform 80% of generations regardless of model.

[Shot type, lens] of [subject] [action]. [Lighting]. [Camera move]. [Style reference].

That's it. Five parts. In order. Worked example:

"Medium close-up, 50mm anamorphic, of a young chef plating a dish in a copper-lit kitchen. Practical key from a window, deep shadows. Slow handheld push-in. A24 aesthetic."

Part 1: shot type and lens

Lead with how the shot is framed. The vocabulary of cinematography is far more precise than the vocabulary of "video":

• Wide shot — establishes location, subject is small in frame.
• Medium shot — subject from waist up.
• Medium close-up — subject from chest up.
• Close-up — subject's face and shoulders.
• Extreme close-up — eyes, hands, single details.
• Over-the-shoulder — view from behind one character looking at another.
• Low angle / high angle — vertical position of camera.

Lens vocabulary tells the model what optical look you want:

• 14mm — extreme wide, distortion
• 24mm — wide, naturalistic
• 35mm — slightly wide, documentary feel
• 50mm — natural perspective, the "human eye"
• 85mm — portrait lens, compressed background
• 135mm+ — long lens, heavy compression
• Anamorphic — wide aspect ratio with horizontal flares

Part 2: subject and action

Be specific about who and what. "A man" is weak; "a delivery courier in his 30s" is strong. "Walks" is weak; "weaves through pedestrian traffic" is strong. The model rewards detail because detail is the only signal it has about your intent.

Part 3: lighting

Lighting is the single most underrated part of a prompt. It is what separates "AI-looking" output from "filmed" output. Vocabulary that consistently works:

• Practical key — light from a real source in the scene (lamp, window, sign)
• Tungsten — warm, incandescent
• Halogen — neutral, slightly cool
• Sodium / neon — heavily colored, urban
• Golden hour / blue hour — natural time-of-day light
• Overcast — soft, even, no shadows
• Hard / soft — quality of shadow
• Backlit / rim-lit / silhouette — direction of light relative to subject

Part 4: camera move

Specify the move explicitly. "The camera does X." Vocabulary:

• Holds / locked — no movement.
• Push-in / pull-out — toward or away from subject.
• Pan left/right — horizontal rotation.
• Tilt up/down — vertical rotation.
• Dolly — physical movement of the camera.
• Tracking shot — camera moves with the subject.
• Crane / jib — vertical movement up or down.
• Handheld — small organic motion.
• Steadicam — smooth movement that follows the subject.

Also useful: speed and timing. "Slow push-in," "snap pan," "two-second dolly," "sustained tracking shot."

Part 5: style reference

End with a reference that tells the model what aesthetic universe the shot lives in. References that work consistently across models:

• Cinematic — generic but reliable
• Documentary — handheld, naturalistic
• Commercial — clean, polished, brand-safe
• Editorial photography — magazine-shoot energy
• Film noir — high contrast, hard shadows
• A24 — restrained, intimate, color-graded
• 1970s American cinema — warm, slow, lived-in
• Anamorphic — wide aspect, lens flares
• Music video — heightened, kinetic

"The job of a prompt is to give the model exactly enough information to make one specific shot. Most prompts are too short and too vague. The fix is almost always to add more direction, not less." — AI Studio Editorial

The negative space rule

What you leave out matters as much as what you put in. If you don't mention "fast cuts," you won't get fast cuts. If you don't mention "music," there will be no music. Conversely, if you mention "no music" you're spending tokens on something the model wasn't going to add anyway.

Put your tokens into the things that are active choices: lighting, composition, camera, style. Skip the absences.

Model-specific tweaks

The universal structure works everywhere, but each model has its quirks:

• Veo 3.1 rewards camera language. Be explicit about lens and move.
• Sora 2 rewards organic-feeling prompts. Don't over-direct; trust the model with looseness.
• Kling v3 rewards scene-first thinking. Lead with the scene, then specify shot coverage.
• Wan 2.7 rewards multi-shot prompts. Treat the prompt as a sequence, not a single shot.
• Seedance 2.0 rewards specificity. Adjective density helps.

The bottom line

Prompt engineering is just shot description with a bigger vocabulary. Once you internalize the five-part structure and the vocabulary of cinematography, your iteration count drops and your output gets noticeably better. Open AI Studio, copy one of the prompts above, and run it through the Roster. The right way to learn is to ship.