How Long Should an AI Video Prompt Be?

Three prompt cards of increasing length flowing into an AI-generated coffee shop video scene

It’s the question every beginner asks and nobody answers straight: how many words should this prompt be? Ten? A hundred? A paragraph? The honest answer is there’s no magic number — but there are useful ranges, and they depend on what kind of shot you’re asking for. Here’s how I judge prompt length, with the same shot written three ways so you can see what the extra words actually buy you.

The Short Answer

For a single clear action, 20–40 words is plenty. For a directed shot with camera, light, and sound, aim for 60–100. Around 150 words, most models start quietly dropping your instructions — and past that, you’re usually better off splitting the idea into two shots or using prompt chaining instead of writing one enormous prompt.

Length isn’t about effort. It’s about how many separate decisions you’re handing the model at once.

One Action: 20–40 Words

When the shot does one thing, keep the prompt lean. Image-to-video especially: the starting frame already carries the subject, the outfit, the style, and the setting, so your prompt only owes the model motion, camera, and mood.

“A woman in a red coat walks into a small coffee shop, shaking rain off her umbrella, warm light, quiet morning mood, medium shot, slow push-in.”

That’s roughly 25 words and it’s enough. Subject, one action, setting, light, camera. Nothing in it fights anything else.

A Directed Shot: 60–100 Words

When you care about the details — what happens at the counter, what the room sounds like — spend more words, but spend them on specifics, not adjectives:

“A woman in her thirties wearing a red wool coat steps in from heavy rain into a small neighborhood coffee shop, folding a wet umbrella as the door bell rings. Steam rises from an espresso machine behind the counter; the barista looks up and smiles. She crosses to the window seat, coat dripping slightly, city lights smearing through the wet glass outside. Medium shot, slow push-in. Warm practical light, quiet morning mood. Audio: rain on glass, door bell, low cafe murmur.”

Roughly 80 words. Notice what the extra length bought: a second beat (the barista), sound design, and precise light. It did not buy “beautiful, stunning, cinematic, breathtaking” — four words that tell the model nothing.

Multi-Beat Scenes: Around 150, Then Split

Once a prompt needs her to enter, order, sit, open a notebook, and react to something outside, you’re writing a scene, not a shot. You can push toward 140–150 words and a strong model will hold most of it — but “most” is the problem. The detail it drops is random: the notebook, the line of dialogue, the weather. You find out which one only after paying for the generation.

That’s the moment to split. Two generations of 60–80 words each will beat one 150-word gamble almost every time, and the cut between them usually improves the scene anyway. Long prompt, or chained prompts — that trade-off is the real skill here.

What Longer Prompts Actually Do

Two patterns show up again and again in practice. First, instructions placed late in a long prompt get ignored more often than ones placed early — so front-load what matters: subject, main action, camera. Second, added words only help when they add decisions (a specific sound, a specific light source). Added adjectives just dilute the decisions around them. If your prompt is getting longer and your results aren’t getting better, the prompt isn’t too short. It’s too crowded.

What to Cut First

When a prompt feels flabby, cut in this order: stacked adjectives (“gorgeous, stunning, ultra-detailed”), style pile-ups (three genres in one sentence), and anything the reference image already shows. Keep the subject, the single main action, the camera move, and the sound. That skeleton is what the structure in anatomy of a great prompt is built on, and it survives at any length.

A Simple Way to Find Your Number

Stop guessing. Take one shot you care about, write it at 25, 80, and 140 words, and generate each once with everything else held steady. Compare what survived. You’ll learn more about your model’s appetite from those three generations than from any rule — including mine. Then save the version that worked as a starting point; a personal prompt template you’ve actually tested beats a word count you read somewhere, every single time.

Scroll to Top