Every impressive AI video you’ve seen started as a text prompt. And behind most of those prompts is the same hidden structure — a way of describing shots that gives the model everything it needs and nothing it doesn’t. This guide dissects that structure piece by piece.
The six questions every prompt must answer
A complete video prompt answers six questions. Miss one, and the model invents the answer — sometimes brilliantly, often badly.
1. Subject — who or what is on screen?
Name the visual facts that define the subject: “a street vendor in a white apron” beats “a man.” Include only details that matter — every extra adjective is a constraint the model must satisfy, and over-constrained prompts produce stiff, confused results. Two to four defining details is the sweet spot.
2. Action — what is happening, right now?
One clear verb in the present tense: “flipping skewers over charcoal,” “turning toward the camera.” AI video handles a single decisive action far better than sequences. “Walks to the window, opens it, and waves” will usually collapse into a mush of all three. One action per shot — if the story needs more, that’s more shots. (This is why shot planning matters.)
3. Camera — angle, movement, framing
This is the most underused and highest-leverage element. Real camera vocabulary transforms generic footage into directed footage: Framing: extreme close-up, close-up, medium shot, wide shot, aerial. Angle: eye-level, low angle (power), high angle (vulnerability), bird’s-eye, Dutch angle (unease). Movement: static, slow push-in, pull-back, pan, tracking shot, crane up, handheld sway.
“Static wide shot” and “slow crane-up from a close-up” generate completely different emotional experiences from identical subjects. If your footage looks amateur, the missing ingredient is almost always camera language.
4. Setting — concrete nouns, not adjectives
Models render nouns better than adjectives. “Night market, hanging lanterns, steam rising from food stalls” gives the model things to draw; “beautiful atmospheric street” gives it a mood to guess at. Three concrete nouns beat ten adjectives.
5. Lighting — name the source
Don’t write “good lighting” — describe light like a cinematographer: “warm lantern light from the left, cool moonlight fill from the right.” Golden hour, neon practicals, overcast softbox sky, candlelight — named light sources produce coherent, motivated lighting. Generic requests produce generic flatness.
6. Style — two words, at the end
Finish with minimal style direction: “photorealistic, cinematic.” Style words are seasoning — they work when the shot is already fully described, and they can’t rescue a prompt that’s missing the first five answers.
Putting it together: three examples
Weak: “beautiful girl walking city night cinematic 8k”
Better: “Medium tracking shot of a young woman in a red coat walking through a rain-soaked Tokyo side street at night, neon signs reflecting in puddles, magenta and cyan practical lighting, shallow depth of field, photorealistic, moody.”
Directed: “Slow dolly-in from medium to close-up: a young woman in a red coat pauses mid-step on a rain-soaked Tokyo side street, turning toward camera as neon signs flicker; wet asphalt mirrors magenta and cyan light; shallow depth of field isolates her face; photorealistic, tense, cinematic.”
Each version adds structure, not just words. The directed version specifies the camera move, the precise action beat, and the lighting behavior — so the model has decisions to execute, not decisions to make.
Common structural mistakes
- Front-loading style words. “Cinematic 8k masterpiece” at the start doesn’t improve anything; put craft first, style last.
- Resolution numbers. “8k, 16k, ultra HD” don’t change output resolution. Skip them.
- Shot lists in one prompt. “First she walks, then it cuts to the market, then a drone shot” — that’s three prompts, not one.
- Emotional adjectives without visual causes. “Sad, lonely, nostalgic” mean nothing visually. Translate: “empty street, single streetlamp, slow push-in” — that’s what lonely looks like.
- Text and logos. Unless your tool specifically handles typography, keep words out of the frame. Models still garble text regularly.
Adapting the structure per tool
The six questions are universal; the syntax isn’t. Some tools want comma-separated phrases, others parse natural sentences better. A few accept reference tags or camera presets. Learn your tool’s preferred format, but keep the underlying structure — subject, action, camera, setting, lighting, style — identical. The structure is the skill; syntax is just translation.
Practice drill
Take one weak prompt and rewrite it three times, changing only the camera element each time: static wide, slow dolly-in, handheld close-up. Generate all three. You’ll learn more about prompting from this single exercise than from any list of “magic words” — because you’ll see, concretely, that structure directs and adjectives decorate.
For the full reusable system built on this anatomy, see AI video prompting: a practical system that actually works.