How to write AI video prompts that actually look cinematic

There is a moment most people hit a few days into using AI video, where the results are fine but never quite cinematic, and they assume they have hit the limit of the tool. Almost always, they have hit the limit of their prompt. A vague description gives a video model nothing to commit to, so it hedges, and hedging is what produces the flat, undirected clips that look like everyone else's. The good news is that prompting is a craft you can learn quickly, and these patterns show up in nearly every result that actually looks good, whichever model renders it.

Name the camera, not just the scene

This is the single highest-leverage change you can make. "A slow dolly in on a coffee cup" gives the model a job for the whole clip. "A coffee cup on a table" gives it nothing to move toward, so it either sits static or invents a motion you did not ask for. Name the move explicitly, a dolly, a pan, a handheld drift, a locked-off static, and you have handed the shot a spine.

Before: a woman in a cafe. After: a slow push in on a woman by a cafe window, she looks up from her book. The second one is a shot. The first is a description.

Describe the light before the mood

Mood words like moody, dramatic or cinematic are hard to render because they are feelings, not specifications. "Low, warm side light with a soft shadow falling to the left" is something a model can actually execute, and the mood you were reaching for usually falls out of it for free. Be concrete wherever lighting is concerned, and save the feeling words for genuine performance and emotion.

One beat per shot

Cramming two actions into one generation, "she picks up the phone and walks to the window and looks out," usually produces none of them cleanly, because the model has to compress three moments into one continuous take. Split it into separate shots and let the tool sequence them. You will get cleaner motion in each and far more control over pacing when you cut it together.

Give it a reference whenever you can

A product photo, a character reference, even a rough phone clip of the space will out-perform a written description almost every time, because it removes an entire layer of guesswork about what things actually look like. Use words to describe what should change or move, and use images to establish what things are. That division of labour is where a lot of professional-looking results quietly come from.

Say what the shot is for

"An establishing shot to open a thirty second ad" gets treated differently from "a hero shot for the thumbnail," even with identical subject matter, because the planning layer reads intent as much as content. Telling it the shot's job in the sequence, and not only what is in the frame, is often the difference between a generic result and one that cuts well against what comes before and after it.

Be specific about the subject, not just the vibe

Models fill in whatever you leave out, and they fill it in with the average of everything they have seen, which is exactly how you end up with a generic result. "A person" becomes a stock-looking stranger; "a woman in her sixties with short grey hair, in a linen apron" becomes someone specific. The same goes for objects and places. The more concrete the subject, the less the model has to guess, and guessing is where the blandness comes from. You do not need a paragraph, you need the two or three details that make this subject this subject and not any subject.

Match the format to where it will live

A prompt is not just what is in the frame, it is also the shape and length of the frame. A vertical nine-by-sixteen clip for Reels wants its subject composed for a tall frame and its action front-loaded into the first couple of seconds; a wide sixteen-by-nine establishing shot can breathe. State the aspect ratio and the intended length up front, because a shot composed for the wrong frame has to be awkwardly cropped later, and cropping throws away half the pixels you paid to generate.

It is also worth keeping shots short. Most strong AI video is built from clips of a few seconds each, cut together, rather than one long generation, because short clips are cheaper to iterate on and easier for a model to keep coherent. Think in shots, not scenes.

How long should a prompt actually be?

Long enough to be specific about subject, camera and light, and no longer. A common failure is the wall-of-adjectives prompt that piles on "cinematic, epic, hyper-detailed, award-winning, 8k" until the model cannot tell what actually matters. Those words feel like they should help and mostly cancel each other out. Two clear sentences that name the subject, the move and the light will beat a paragraph of adjectives almost every time. Write the brief, not the thesaurus.

The mistakes to stop making

Three habits quietly ruin prompts. Piling on adjectives until the model cannot tell what matters. Describing a static scene and then being surprised there is no motion. And rewriting the entire prompt when one thing is off, instead of changing the single variable that failed and regenerating. Fix those three and your hit rate climbs immediately, without changing tools or models at all.