Write Like a Director: A Prompt Pattern That Holds a Shot

Most weak AI video prompts fail in the same way. They describe a mood and hope the model will invent a scene. "Cinematic product reveal, ultra detailed, epic lighting, slow motion, 8k" is a pile of adjectives. It does not say who is on screen, what happens, where the camera stands, or when anyone speaks. The result is usually a pretty five-second loop that you cannot cut into an ad, a tutorial, or a story.

A better prompt reads like a shot list. This article is a pattern I reuse, with examples and a short list of failure modes.

The five-part shot card

Before I type a prompt, I fill in five fields. If a field is blank, I do not generate yet.

1. Subject. A person, a product, or both, described as nouns you could point at.
2. Action. One main verb. A hand places a bottle on a counter. A rider turns a corner. A presenter looks up and delivers a single line.
3. Camera. Move, lens, and distance: slow dolly-in, 35mm, mid shot. Or locked-off wide, 24mm. Or handheld tracking shot at chest height.
4. Light and setting. Window light from the left, rainy street at night, softbox over a seamless backdrop. Time of day belongs here.
5. Sound. Dialogue as a complete sentence, plus room tone or a cue: footsteps on gravel, a kettle click, crowd wash under the line.

Here is the same idea written two ways. The first is the adjective pile. The second is the shot card flattened into a paragraph.

A stunning cinematic commercial of a coffee brand, highly detailed, dramatic, professional, masterpiece.

Close-up of a ceramic cup on a walnut counter as a thin stream of espresso fills it. The camera starts tight on the crema and eases into a short dolly-out to a mid shot of the cup and a folded linen napkin. Warm morning light from a window on the right, shallow depth of field, no logo animation. Soft cafe room tone, a distant saucer clink at the end of the pour.

If the pour happens and the camera eases back, the prompt worked. If a barista appears who was never mentioned, you know which clause to tighten.

One action per render

The most common way to break a shot is to ask for a montage. "She walks into the kitchen, makes coffee, sits down, opens a laptop, and smiles at the camera" is five shots. In a single generation, the model will usually rush them, smear the cuts, or change the room halfway through.

Give the render one job. If the story needs a walk-in, a pour, and a line to camera, plan three clips, or use a longer take and mark the beats by timestamp:

0–8s: mid shot, she enters the frame and sets the cup down. 8–16s: the camera eases in as the pour begins. 16–24s: she looks up and says, "This is the only recipe I use on recording days." Locked 35mm, kitchen window light, consistent wardrobe. Quiet interior, cup on wood at 8s, line at 16s.

Timestamped beats only help if the actions fit inside one unbroken camera move. They are a way to keep a setup, a turn, and a payoff in the same take, not a substitute for an edit.

Tag every reference, or the model will guess

When the exact face, product, or location matters, attach a file and name its job in the prompt. "Use @Image1 as the character's face, @Image2 as the kitchen, @Image3 as the color grade" is clearer than dropping three pictures into a slot and hoping. Audio can set rhythm the same way: a scratch voice track or a bed of room tone tells the model when the line should land.

I lean on Seedance 3.0 when a prompt has to carry mixed references — stills, short clips, and audio in one generation — because you can tag each file for the job it controls, and the voices and effects are rendered with the picture. The official advice matches the pattern above: name who is on screen, what they do, the camera move, the lens, and the light, and stick to one main action per render.

A small vocabulary that pulls more weight than adjectives

You do not need a film-school glossary. A short, consistent set of terms is enough:

• Distance: extreme close-up, close-up, mid shot, wide, very wide.
• Height: low angle, eye level, high angle, top-down.
• Move: locked off, slow dolly-in, dolly-out, pan left, tilt up, tracking shot, handheld, crane down, orbit.
• Lens feel: 24mm wide and slightly distorted, 35mm natural, 50mm tight, 85mm portrait, shallow or deep depth of field.
• Light: soft window light, hard noon sun, practical lamps only, overcast, neon spill, golden hour, tungsten interior.

Pick one from each line. Mixing "handheld and locked off" or "24mm and 85mm" in the same sentence is how you get a camera that cannot decide what it is.

Failure modes I now check for

Stacked style words, no verbs. If you can delete half the prompt and lose no action, rewrite it.

Unspoken dialogue. "She talks about the product" will wander. Write the sentence.

Conflicting physics. A slow-motion pour inside a snappy handheld walk-and-talk usually collapses. Choose a tempo.

Identity without a lock. If the same person or product must return in the next shot, do not rely on the prose description alone. Reattach the reference.

Revising by regeneration. When a take is almost right, describe only the change — weather, time of day, a swapped prop — and keep the framing. That is a prompt skill too: the second prompt should be a diff, not a sequel.

Write the shot card first, then the sentence. If a stranger could storyboard from your prompt, the model has a fair chance of doing the same.