How to write a prompt for AI video: say it in plain words, like to a crew

By Deep Bhalerao, founder · Published

An army operations room full of uniformed operators at their consoles, a giant grey creature on the big screen at the front.

You do not need a special prompt language to make AI video in Agent Social. You describe the story in plain words, the way you would brief a film crew, and the studio does the rest. I made every episode of OVERCLOCK that way, from a chat.

How do you write a prompt for AI video?

In Agent Social, you write it as a short description of the story, in plain words, with no technical syntax. Say who it is about, where it happens, what happens and how it ends, and the studio plans the film as chapters, writes the dialogue and shows you everything before a single clip is filmed. One sentence is enough to start, and a synopsis or a finished script works too. Then you read what comes back and change it by talking: "make the ending darker", "cut the café scene", "put Nora on his left". The request is the start of a conversation, not a one-shot spell that has to be perfect. This is what makes AI story to video work for a whole film rather than one clip.

Why is plain language better than a technical prompt?

Because the studio does the technical work, and you are the director. In Agent Social, you talk to the studio the way you would talk to a film crew: "make the rain heavier", "redo the last shot from lower down". It writes each clip's screenplay for you to read, and nothing is filmed until you ask for it. A long list of technical terms tends to bury the one thing that matters, which is the story. A clear sentence about people and what they want tells the studio far more than a paragraph of keywords. It also means you can judge the result as a viewer, which is the job that matters. In OVERCLOCK Episode 3, the screenplays went through about 350 rounds of changes, every one asked for in plain words in the studio chat.

The same story, a weak request and a strong one
Instead of

A weak request

  • A cool scene with a monster, epic, 4K, cinematic.
  • No people, no place and no story
  • Nothing to carry into the next clip
  • You get a pretty picture that goes nowhere
Do this

A strong request

  • Singapore, sunrise. A small boy on a crowded beach is the first to see something rise from the sea.
  • A giant clawed creature walks ashore and ignores the people running from it.
  • It ends when the police pass the problem to the army.
  • You get a plan with people, a place and an ending

Which details matter most in a story request?

Four details matter most: who it is about, where it happens, what happens, and how it ends. In Agent Social, those four give the studio everything it needs to plan the chapters. "Who" is one or two people with a want. "Where" is one place you can picture, and it can be a street or a lighthouse. "What happens" is the trouble, in a sentence. "How it ends" is the turn, even if it is a small one. Add the tone and the look if you care ("photoreal and tense", "a warm cartoon"), and the length if you have one ("a three-minute film"). Leave out everything else. The studio fills in the gaps and you can fix them at the plan. You can even start from your own photos, and the same faces appear in every shot.

A request in four lines
  1. Who

    A lighthouse keeper named Mara, in her sixties, who lives alone.

  2. Where

    A lighthouse on a grey northern coast, in a winter storm.

  3. What happens

    Tonight the lamp goes out and a ship answers in the dark.

  4. How it ends

    She relights the lamp and the ship turns away from the rocks.

How do you describe a character or a place in a prompt?

Describe a character by what a stranger would notice first, and a place by what you could stand in. In Agent Social, each recurring character and place is made once as a picture, and you check it before any clip is filmed, so a few clear words are enough: "Mara, sixties, weathered face, a grey wool coat" or "a lighthouse kitchen with one lamp and a kettle". You can also use your own photos to play a character yourself, which is how I play Deep. People who appear only once, like a voice on a radio, are simply described in their clip. There is more in the guide to making AI characters and places.

A woman in a green dress walks down a supermarket aisle between shelves of tins and snacks, people gathered at the glass doors behind her.
OVERCLOCK Episode 4, Double Scoop: a place you can picture, a supermarket aisle with people at the glass doors.

How do you change an AI video request after the first try?

Change the plan or the one clip that missed, in plain words, and leave the rest alone. In Agent Social, you read the plan before anything is filmed, so ask for changes there first: a different ending, a shorter scene, a line that sounds more like your character. After a clip is filmed, name what was wrong as a direction, not a complaint: "she should look at him before she answers", "he leaves through the door on the left". Only that clip is filmed again. Some clips land at once and some take several tries. In OVERCLOCK Episode 5, the supermarket chapter took six takes, while the jaws coming out of the smoke were a single 30-second take, used whole on the first try. See writing a screenplay for AI video for more.

What does it cost to try a prompt?

Reading and changing the plan comes first, and a film costs what its finished minutes cost: from $25 per finished minute, up to 1080p. A one-minute film starts from $25, and a five-minute film from $125. In Agent Social, it is pay as you go, with no subscription and no free plan, and the credits you do not use never expire. The draft is the default and is enough to judge every clip, so you can check the story, the faces and the sound before you spend more. Upscale the finished film to 1080p once you are happy with it. A clear request means fewer clips filmed again, so read the plan before you ask for the first clip.

Check yours

Is your request ready to send?

  1. Does it say who the story is about?
  2. Does it name one place you can picture?
  3. Does it say what happens and how it ends?
  4. Is it a list of style words with no story?

Answer each question for your verdict.

Sources

Related: AI story to video — a whole film from your text

Try it

Start with one line.

You type

“A fisherman on a Goa beach finds a bottle with a map in it. He follows it at dawn and finds his grandfather's boat. Warm and quiet, three minutes.”

What happens

The studio writes the story as chapters with dialogue and shows you the plan before anything is filmed.

You type

“Keep my dialogue exactly as I wrote it.”

What happens

Your lines stay where you wrote them, and the studio fills in only what is missing.

You type

“Make the ending darker and cut the café scene.”

What happens

The plan changes before a single shot is filmed.

Make your own film in Agent Social.

Describe it in one sentence. The studio writes it, casts it and films it.

How to write a prompt for AI video — Agent Social