How to write an AI video prompt that actually works

You can usually tell a prompt is going to fail before the render finishes. It reads like a mood board: a pile of adjectives, a genre, maybe a director's name, and no statement of what is physically happening in front of the camera. The model returns something that matches the vibe and nothing else, and because it technically did what was asked, there is no obvious thing to fix.

The mismatch is in what a prompt is for. A text-to-image prompt describes a frame, and a frame tolerates ambiguity because the model only has to make one internally consistent picture. A video prompt describes a frame plus everything that happens to it over several seconds. Every unstated detail is a decision the model makes for you, and it makes a fresh one each time you re-roll.

So the practical definition worth adopting is narrow: a prompt is a shot description handed to a crew that will never ask a follow-up question. Everything a cinematographer would need to know, and nothing that a cinematographer would find meaningless.

That definition is worth adopting for what it implies about who can do this. Writing a good prompt is not a creative gift, and it is not a secret phrasing being withheld from you by people who post their results. It is the ordinary skill of saying precisely what you want, which most people already have in the parts of their life where the stakes made them practise it. The ones getting output you envy are rarely more imaginative than you are. They are more specific, and specificity is learnable in an afternoon.

The four things that break first

• Too many actions. A clip of a few seconds can hold one action. Ask for a character who walks in, sits down, opens a laptop and sighs, and the model will blend them into a single ambiguous motion or pick one at random.
• Adjectives standing in for nouns. "Cinematic, epic, beautiful, 8K" describes how you want to feel about the result. It does not describe the result. Replace each one with the concrete thing that would cause it: a lens, a time of day, a camera height.
• Unanchored subjects. "A man" is a different man in every generation. If the same person needs to appear in shot two, the prompt has to describe them in enough fixed detail to be reconstructible, or the shot has to start from a reference image.
• Contradictions the model has to resolve. "Handheld documentary feel, smooth cinematic dolly" asks for two mutually exclusive camera behaviours. The model will silently pick one, and which one it picks will change between runs.

Order matters more than wording. Most video models weight the beginning of the prompt more heavily, so the parts go from the thing that must be right to the thing you can afford to lose. Put the style tokens last, where they act as a finish rather than as the subject.

That is one sentence per part, declarative, no metaphor. It is longer than the prompt most people write and it is not more creative. It is less. The creativity moved out of the prompt and into the decisions about what to specify, which is where it survives being re-run.

Note what is absent. No word count padding, no resolution claim, no request for the model to try hard. Asking for 8K in a prompt does not change the output resolution, which is set by the model and the render settings; it just spends attention that the subject needed.

Write camera moves as one instruction, not two

Camera language is the part most people get wrong by trying hardest. A short clip supports one move: a push in, a pull back, a pan, an orbit, or a static hold. Static is underrated and reads as deliberate when everything else in a feed is drifting.

Combining moves is where prompts go abstract. "Dolly in while orbiting and craning up" describes something a crew could shoot and a model cannot, so it approximates, and the approximation looks like the wobble people mean when they say AI video looks like AI video.

This is the problem that separates a nice single clip from a watchable sequence, and prompt text alone does not solve it. Two generations from identical text are two different people who happen to match the same description, in the same way two casting calls with the same brief return two different actors.

There are three approaches, and they are worth choosing between deliberately rather than defaulting to the first one.

• Describe by role, not by name. Invented character names mean nothing to the model and quietly waste tokens. "The narrator" or "the same woman in the navy apron" carries information; "Sarah" carries none unless Sarah is a public figure the model has seen.
• Lock the frame, not the words. Generate one image you are happy with, then use it as the first frame or reference for each clip. This is the only method that holds a face still across a whole sequence, and it moves the effort from prompt-writing to image selection.
• Design around the problem. The most reliable fix is to need consistency less. Hands, objects, landscapes, over-the-shoulder framing and shots where the subject faces away all sidestep the failure. A lot of strong faceless content is built this way on purpose.

Generation cost per clip is the number that decides whether a format survives contact with a weekly schedule. We broke down what a faceless channel actually costs to run

Working from an existing text sidesteps most of this, because the source already describes the character in fixed words: filming a book without rewriting it

The spoken layer is written separately and it is written as direction, not as prose. The most common mistake is handing a text-to-speech model a paragraph built for the eye, full of long sentences and subordinate clauses, and being surprised that it sounds like someone reading aloud, because that is exactly what it is.

Script for the ear instead. Short sentences. One idea each. The rhythm you would use if you were explaining the thing to somebody standing next to you, which means contractions, and sentences that occasionally begin with "and".

What actually changes the delivery

• Punctuation. A comma is a short beat, a full stop is a longer one, and a line break is longer still. This is the most reliable pacing control you have, and it works on every engine.
• Sentence length. Emphasis comes from a short sentence after a long one. No engine adds that for you, and no amount of instructing it to sound excited replaces it.
• Spelled-out numbers and units. "Twenty twenty-six" and "four thousand hours" read correctly. "2026" and "4,000 hrs" are a gamble that differs by engine, and the failure is audible.
• Phonetic respelling for names. Any proper noun the engine is likely to mangle should be written the way it sounds. Fixing it afterwards means regenerating the whole take.

Instructions like "speak in an excited tone" are worth testing rather than assuming. Some engines accept a style directive; on others the words end up spoken aloud in the render, which is a failure you only catch by listening to the whole take.

Which engine you are writing for changes what these controls do, which we covered in how to choose an AI voice for short-form video

The instinct after a bad generation is to rewrite the whole prompt. That is the slowest possible method, because it changes many variables at once and teaches you nothing about which one was responsible. A prompt that got worse for an unknown reason is not recoverable, so people re-roll until something acceptable appears and never learn why.

Change one part at a time, against the six-part structure. The part to change is the one the failure points at.

Keep the ones that worked

Prompts that survive several rounds are assets, and most people throw them away. Keep them in a file with the model they were run against and the date, because the same text on a new model version behaves differently and you will want to know whether the prompt changed or the model did.

The version note matters more than it sounds. Providers update models continuously, and a prompt that mysteriously stopped working is very often a prompt that did not change at all.

That judgement is a separate skill from prompting, and we went into it in what actually holds attention in the first three seconds

Everything above assumes you are writing each prompt by hand. That is the right way to learn what the model responds to, and the wrong way to publish four videos a week for a year. At that cadence the prompt stops being something you write and becomes something a system generates from a script.

The shift is real and worth being clear-eyed about: you stop authoring prompts and start authoring the RULES that produce them. The six-part structure becomes a template, the script supplies the subject and action per scene, and the style block stays fixed across an entire series so episodes look like they belong together.

This is also where the honest limitation sits. A generated prompt is only as good as its template, and a template built from vague instructions produces vague prompts at scale, faster than anyone can review them. The hand-written practice is what tells you whether the template is any good.

• Fix the style block per series. One style description reused across every scene of every episode. This is what makes a channel look like a channel rather than a collection of unrelated experiments.
• Vary only subject and action. The script decides what happens in each scene. Camera, light and style should change rarely and deliberately.
• Keep a review step before anything publishes. Generation is cheap and wrong output is public. A review window in front of the upload is the cheapest insurance in the entire pipeline.

AutoVidGen builds the scene prompts from the script and holds a review window before the upload: how AutoVidGen turns one prompt into a finished video