Every week we read a hundred AI-video prompts that open with the word "epic." Almost none of them read like films. They read like wish lists — a hero, a city, some fire, "in 4K." And then a client is surprised when the result is a postcard instead of a scene. The tools have changed, but filmmaking has not. A video prompt is not a description of a world; it is a brief for a shot. The difference is the whole craft, and it is learnable. This is the checklist our directors run before a single frame renders.

Start with the frame, not the scene

Models are excellent at worlds and unreliable about intention. So begin the way a DP begins — with the frame. Ask three questions before you type anything else: what are we looking at, from how far away, and through what lens? A wide establishing shot of a city is a different prompt entirely from an 85mm close-up of a hand on a doorknob, even if the city is behind both.

Naming a lens changes everything. Write 35mm, slightly wide and the model leans cinematic; write 50mm, eye level and it leans documentary and intimate. We use focal length as a mood dial more than a technical spec — short lenses stretch and exaggerate, long lenses compress and isolate.

Direct the light like a gaffer

The fastest way to tell a hobby prompt from a studio prompt is light direction. "Golden hour" is a default. What we want instead is a decision: key light from the left, practical lamp behind the subject, single overhead source, volumetric haze catching the beam.

Choose the source, the direction, and the mood of the falloff. A hard source with long shadows reads noir. A soft window light wrapping the face reads warmth. Backlight against deep shadow reads mystery and separates your subject from the frame. When clients say "make it cinematic," nine times out of ten what they mean is "give it a motivated light source." So give the model one.

Motion and pacing are half the story

Video models see two things at once: a frame and a movement through time. Most prompts only describe the first. Name the camera move and the pace explicitly. Is it a slow dolly-in, a handheld push, a locked-off static wide while everything moves through frame? Is the motion calm and drifting or agitated and quick?

Pacing works the same way it does on a real set. Short prompts with clipped sentences render as quick, punchy cuts. Long, flowing sentences invite slow pushes and dissolves. If a beat matters, slow the language around it. If a moment should snap, end the sentence where the cut lands. The rhythm of your words becomes the rhythm of the footage.

Structure the prompt shot by shot

One paragraph trying to do eight things produces mush. The most reliable structure we use at the studio is a mini-storyboard in text — each shot gets its own line, its own camera, its own light:

  • SHOT 1 — establishing: the setting, the light, the lens, what the audience should notice first.
  • SHOT 2 — medium: the subject in action, the direction they move, the key emotion.
  • SHOT 3 — close-up: the detail that carries the meaning — a hand, an eye, a telling object.
  • SHOT 4 — payoff: the reveal, the shift in light or motion, the final beat.

Each shot can be generated separately and cut together — that is how we work every day, and it gives you control no single mega-prompt can match. You fix one shot without re-rolling the whole sequence.

Describe versus direct

The single biggest failure we see is describing instead of directing. Describing tells the model what is; directing tells it what happens. "A detective stands by a window in the rain" is a photograph. "A detective stands by the window as rain streaks past; he doesn't turn around; the camera pushes in as his hand tightens around the glass" is a scene.

The model will happily paint you a million gorgeous stills. It will not choose which still matters. That choice is the directing — and it has to be in the prompt.

Verbs are worth more than adjectives. stops, glances, reaches, hesitates — these are the difference between a slideshow and a story. When we audit a weak prompt, it is almost always adjectives with no verbs attached. Delete the adjectives first, add a verb, and the scene comes alive.

A worked example: before and after

Here is a real brief a client brought us for a perfume launch — the "before":

"Epic cinematic perfume commercial, beautiful woman in a dark room, golden light, luxury, 4k"

Everything is a mood, nothing is a decision. The model produces something pretty and completely generic. Here is the "after" we built with them:

"A single beam of moonlight cuts across a dark room. 35mm, slow dolly-in. A woman's hands lift a glass bottle into the light — she pauses, not looking at it yet. Close-up, 85mm, rack focus from the bottle to her eye. Volumetric haze in the beam, long shadows. Final beat: she finally looks at it, and lets go a breath she was holding."

Same product. Same budget of words. The second version is a film because it makes decisions: a light source, a camera move, a sequence, an emotional arc. Nothing in it is decoration. Everything serves the one feeling we decided the audience should carry out — relief.

A woman under a tungsten streetlight in the rain, a cinematic noir still
One motivated light source, one camera move, one emotion — the whole checklist in a single frame. Still from our noir reel, "Smoke & Mirrors."

The five failure modes we fix every week

When a prompt comes back flat, it is almost always one of these five. Check yours before you render:

  • No camera. No lens, no distance, no move. Add the frame before the content.
  • No light decision. "Cinematic lighting" is not a choice. Name a source and a direction.
  • No verbs. Adjectives describe; verbs direct. Add action and sequence.
  • One wall of text. Break it into shots. Give each shot its own job and its own line.
  • Stacking moods. "Epic, beautiful, dramatic, moody" cancels itself out. Pick one feeling and defend it.

The checklist, in one breath

Before you hit render: one feeling, a lens and a distance, a motivated light, a named motion, and a shot-by-shot structure. If you can read your prompt and point to each of those, you are directing. If you can't, the model is — and it will choose safe every time.

The good news is that this is a craft, not a trick. We run this checklist on every project that comes through Donnyfy, and we've turned first-time prompters into confident directors in a single session. Start smaller than you think, decide harder than you're used to, and let the model follow the brief instead of the mood.