Writing Prompts for AI Video Models: Describe Causes, Not Effects (and Stop Stacking '8K Masterpiece')

There is one master rule for video prompts: describe what causes the picture, not the picture itself. Light: write the source's position, quality and colour temperature, never "light hits her forehead". Motion: contact point, weight, end state, never "fluid". Dialogue: who speaks and for how many seconds (about 2.5 words per second), never "emotional". "8K, masterpiece, volumetric" are CLIP-era leftovers; not one of Kling's, Google's or Alibaba's official guides asks you to stack them. They give you sentence slots.

Three official formulas, one lesson

Kling: Subject + Subject Movement + Scene + (Camera Language + Lighting + Atmosphere). Google, for Veo 3.1: [Cinematography] + [Subject] + [Action] + [Context] + [Style & Ambiance]. Alibaba Model Studio, for Wan: Subject + Scene + Motion, plus Aesthetic Control + Stylisation in the advanced version. Every slot in all three is a fact you can state in one sentence. None has a slot called "quality words".

"masterpiece, best quality, 8K, ultra detailed" is an artefact of the Stable Diffusion 1.5 era: those tokens co-occurred with high-scoring images in training captions, so stacking them worked on a CLIP encoder. Modern video models read whole sentences through an LLM-style text encoder. The tokens add nearly nothing and crowd out slots that carry information; MiniMax H3 accepts 7,000 characters per text item, but attention is not spread evenly across them.

So the master rule is one line: describe the cause of the picture, never the result.

Lighting: write the source, not the landing spot

The most reproducible failure in production pipelines: the prompt says "warm yellow light brightens her forehead and nose bridge", and the output has a blob of yellow on the forehead and nose bridge, exactly where the prompt pointed, like a torch held to the face. The model executes a landing spot as "paint something here".

Write the cause instead: A warm-white pendant lamp casts soft light from the upper right of frame; cool window daylight fills from the left. Source position, quality (hard or soft) and colour temperature; where it lands is the model's arithmetic. Google's own Veo examples take this form: "harsh fluorescent overhead lights", "the green glow of the monochrome monitor". Sources and qualities, not landing spots.

Three companions. Never write the light/shadow boundary; falloff is a gradient, and naming a "boundary" draws a hard edge. Never use saturated colour names for temperature; 3200 K on skin is warm white, and "yellow" gets you yellow paint. Write volumetric light and god rays as a medium: "a layer of unsettled smoke hangs in the booth". The model knows what light does in smoke.

Motion: contact point, weight, ground relation, end state

"He angrily shoves the door open" leaves four things unspecified; the model guesses each.

• Contact point: palm at what height on the door, or a shoulder?
• Weight: solid oak or hollow core; does the body lean into it?
• Ground relation: feet inside or outside the threshold; does a step accompany the push?
• End state: door at what angle, body stopped where, hand still on the door or not?

End state matters most. A clip has an endpoint; leave it unstated and the model picks one, typically a return to the starting pose or a looping gesture. With several shots in one segment, each shot's end state is the next shot's start state; write it and the cuts stop jumping.

"Fluid", "natural", "powerful" are result words. Delete them.

Dialogue: speaker attribution and seconds

Say who speaks. Two people in frame and no attribution, and the lip-sync lands on the wrong face. Alibaba's Wan guide requires "Character says: 'line'"; Google's Veo example is "A woman says, 'We have to leave now.'"

Budget seconds by speech rate. English runs about 2.5 words per second, Mandarin about 5 characters: a 12-word line needs 4–5 seconds, plus roughly 0.75 s per terminal punctuation mark. Over-budget produces a very stable artefact: a 7-second shot with 4.5 seconds of dialogue and no closing action gets the line spoken twice to fill time. The fix is a closing beat at the end of every segment, a small action without displacement: setting down a cup, looking away. Under-budget, the back half of the line spills onto the next shot's face.

Wan's official voice formula: voice = spoken content + emotion + tone + pace + timbre + accent, example He says: "Study hard, make progress every day", relaxed tone, moderate pace, clear voice, American English. Sound effects = source material + action + ambient sound. Veo's convention: dialogue in quotes, sound effects prefixed SFX:, background prefixed Ambient noise:. Both agree: tone and pace are modifiers outside the quotes, never inside them.

Five things never to write

1. Negations. "No fog", "no extra people" put fog and people into the conditioning. Google's guidance for Veo is to describe what you wish to exclude, for example "a desolate landscape with no buildings or roads" rather than "no man-made structures". Write the positive opposite.
2. The wrong silence word. Wan's official control words are "No dialogue" (Chinese 无台词) and "No background music"; a paraphrase such as 无对白 reads as ordinary description and does nothing.
3. Performance notes inside quotation marks. Wan uses quotes as its dialogue delimiter, so "(sneers)" inside them gets spoken. Put delivery outside the quotes, in the tone and pace slots.
4. Camera moves on every shot. A locked-off camera is the default; write a push-in or a pan only with a narrative reason. Wan's control words are "static shot", "camera push-in", "camera pans left". Move on every shot and the cut looks like a commercial.
5. Facing along the depth axis. "She turns and walks into the depth of the frame" buys three seconds of the back of her head. Write facing as frame-left or frame-right; express depth as a positional end state: "walks to the window and stops".

The sixth was covered above: SD trigger words. No official guide from Kling, Google or Alibaba lists masterpiece, 8K, subsurface scattering or volumetric lighting.

Before / after

The left column states results, which the model can only paint; the right states causes, which it can compute.

Per-model quirks

The overlap outweighs the differences: slot formulas, natural language, references by index or name. The differences cluster around sound: Wan and Veo have explicit dialogue syntax, the others lean on reference audio. Switching models means switching dialogue syntax and reference index; subject, motion and light are written the same everywhere.

A shot template that survives contact with the model

1. Segment header: location; opening state (who is frame-left / frame-right, posture, what is in their hands); one lighting sentence (position + quality + colour temperature); static camera, or a move with a reason.
2. One line per shot: seconds / framing / who is in frame doing what (contact point, weight, end state) / dialogue (speaker + quoted line + tone and pace outside the quotes) / lighting only where it differs from the header.
3. Closing beat: one small action without displacement after the last line.
4. Self-check: delete result adjectives (fluid, stunning, cinematic, 8K); convert negations to positives; recompute dialogue seconds at 2.5 words/s; no facing along the depth axis; references in this model's index format.

Key figures and sources

Sources

• https://help.aliyun.com/zh/model-studio/text-to-video-prompt
• https://help.aliyun.com/zh/model-studio/wan-video-to-video-api-reference
• https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1
• https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide
• https://kling.ai/quickstart/klingai-video-o1-user-guide
• https://kling.ai/quickstart/klingai-element-library-3-user-guide
• https://platform.minimax.io/docs/api-reference/video-generation-v2-create
• https://www.volcengine.com/article/40840

Tools mentioned

Read next