Every AI video creator has written this script: “A brave knight journeys across the kingdom, battles a dragon, and returns home a hero.” Then they type it into a generator — which has fifteen seconds and no memory of any kingdom — and get back a man in armour wobbling near something vaguely winged while the camera has a small panic attack. The idea was fine. The script asked a camera to do a novelist’s job.
Writing for AI video is a translation skill. You take a story-shaped thought and rewrite it into camera-shaped instructions: visible things, in one place, doing one thing at a time. This article teaches that translation for the format you’ll actually make — the 15-second short — with a shape simple enough to use tonight.
A Generator Is A Camera, Not A Screenwriter
Start here, because everything else follows. A screenwriter’s script contains intentions, backstory and meaning: “she hesitates, remembering their last argument.” A camera cannot film a memory. Or an argument that happened off-screen. It can film a woman pausing at a door, her hand stopping an inch from the handle. The hesitation is the same story — but only the second version can be generated.
When you write for a generator, audit every line with one rude question: could a stranger see this, with the sound off? If the answer needs backstory, the line isn’t filmable yet. Rewrite it until the meaning lives in something a body, a face or an object can do on screen.
- “He feels lonely in the big city” becomes: a man at a café window, an untouched second coffee across from him, crowds blurring past outside.
- “The shop is old and full of memories” becomes: dust in a shaft of light, a hand brushing a row of worn book spines.
- “She finally decides to leave” becomes: she puts a letter on the table, picks up a small suitcase, and opens the door to morning light.
And this is why the working creators I know think in frames before they think in plots. The storyboard-first workflow isn’t an organisational quirk — it’s the direct consequence of the generator being a camera. If you can’t picture the frame, the generator can’t either. And it will picture something for you. It has opinions. They aren’t yours.
One more consequence — and it’s good news: because the generator is a camera, your script doesn’t need dialogue to be a story. A look. A turn. An object changing hands. That’s narrative at 15 seconds. Dialogue comes later in your development, when your faces hold up under speech. For now, the camera rule is a gift. It lets you write stories your current skills can actually shoot.
The Three-Beat Shape: Setup, Turn, Land
Fifteen seconds is roughly three shots of five seconds, or one shot with three moments inside it. Either way, the story needs the same three beats. I teach this shape to every beginner — because it’s nearly impossible to make boring and very hard to break.
Setup (first third). Establish where we are and what normal looks like. A barista wipes a counter in an empty morning café. A kid sits with a kite on a hill, no wind. Nothing is “happening” yet — and that’s the point. The viewer is loading the scene into their head, and you’re giving them a stable world so the next beat has something to push against. In generation terms, the setup is your calmest shot: still camera, low motion, clean frame. It is also, not coincidentally, the easiest shot to generate well.
Turn (middle third). One thing changes. Not two. The café door opens and a soaked stranger shakes off rain. A gust finally lifts the kite — and the string snaps. The turn is where beginners overload, adding the change and a reaction and a camera move, and the shot smears. Give the turn a single event and let the subject’s movement be simple: a look up, a stand, a step toward. If you need inspiration for how much story one event can carry, study short sequences built shot-by-shot in the multi-shot storytelling guide and scale the thinking down to three beats.
Land (final third). The story arrives somewhere the viewer can feel. The barista slides the stranger a coffee and gets the day’s first smile. The kid laughs and runs downhill holding the broken string like a trophy. A landing is not a twist — it’s a small emotional resolution. Warmth, surprise, a quiet joke, a moment of beauty that recontextualises the setup. The kite story landed on joy instead of disappointment. That choice is the entire creative act. And no generator makes it for you.
Write your three beats as three plain sentences before you touch a prompt. If the middle sentence contains “and then also,” cut something. If the last sentence could swap with the first, you have a loop, not a story — fine for ambience, but know that that’s what you made.
One Location, One Character, One Action
This is the constraint that makes 15-second scripts generatable, and I want to argue for it properly rather than just handing it down as a rule. Every location change forces a new frame, new light, new background consistency. Every added character doubles your face-consistency problem — and introduces the hardest thing in AI video: two bodies interacting believably. Every simultaneous action splits the model’s attention and your viewer’s.
So: one location, because your setup shot can then serve the whole piece. One character — or one character plus a presence implied at the edge of frame. A hand. A shadow. A voice’s effect. And one action per beat. Watch how much story survives inside the constraint:
A woman alone in a rooftop greenhouse at golden hour (setup). She lifts a watering can, then stops — a single sunflower, taller than her, is turning slowly toward the light (turn). She laughs under her breath and waters its roots first (land). One location. One character. Three simple actions. And it has wonder in it — because the constraint pushed the writing toward one strong image instead of five weak ones.
The constraint is also your budget’s best friend. One location means your starting frame can be developed once, carefully, then reused across beats with different motion — the same image-to-video discipline described in text-to-video vs image-to-video. Beginners burn credits re-rolling backgrounds they never needed. A one-location script spends those credits on takes of the moment that matters.
When do you graduate from the constraint? When your one-character pieces hold together reliably and your character stops changing faces between beats — the system for that lives in character consistency and reference images for consistent characters. Add the second character after the first one behaves. You’ll feel the difference in your takes before any tutorial confirms it.
Writing Lines A Face Can Say
Sooner or later a concept will genuinely need a spoken line. A punchline, a greeting, a reveal. When it does, write for the face and the tools you actually have — not for a film actor. A few hard-won rules from generating more talking shots than I care to count.
Short lines win. Five to eight words is the comfort zone for a generated talking shot at this length. “You came back. I knew you would” carries more than a 20-word explanation ever survives. Every extra syllable is another chance for the mouth to wander off the words.
Front or near-front framing. Faces hold speech best when the camera can see the mouth plainly and the head stays mostly still. Write lines for a character who is seated, standing, facing us — not running, turning, or in profile shouting over a shoulder. Design the shot around the sentence, not against it.
One speaker, one line, then cut. Back-and-forth dialogue multiplies every problem. If two characters must speak, give each their own shot — which your three-beat structure happily accommodates — and let the edit create the conversation.
Emotionally flat is safer than emotionally huge. A quiet line survives generation. A screaming or sobbing line usually takes the whole face with it. Write the feeling into the words and the situation, and let the delivery stay small. The viewer will supply the emotion — viewers are generous that way.
And know when to skip speech entirely. A line that works better as an on-screen caption, timed in your edit, is not a failure of ambition. It’s a directing choice. And at 15 seconds it’s very often the right one. You control captions completely in Premiere. You merely negotiate with a generated mouth.
Turn The Script Into Three Frames Before You Generate
Here’s where the script becomes a shoot. Take your three beat-sentences and turn each into a frame description — the still image that beat opens on. Write them as image prompts, not video prompts: subject, place, light, composition. “A woman in a greenhouse among tall plants, golden-hour light through glass, sunflower head above hers, medium shot.” Generate the stills. Sit with them.
This step, which feels slow, is the fastest thing in the whole process. A bad beat reveals itself instantly as a still: the composition is muddy, the turn isn’t visible, the landing frame looks identical to the setup. Fixing that costs an image generation. Discovering it after the video generation costs five times more — and teaches you less clearly, because motion muddies your diagnosis.
Approve your three frames the way a director approves a storyboard. Because that is literally what they are. The deeper method lives in the storyboard-first workflow and in learning how a prompt is actually built in the AI video prompting system — but the beginner version fits on an index card:
- Three beats, three sentences, one action each.
- Three frames, generated as stills, checked at full size.
- Three motion prompts, each asking for that beat’s single action.
- Edit in order, trim hard, and let the landing frame breathe a half-second longer than feels natural.
Worked example, full pipeline — one of my favourite teaching scripts. Beat one: “Night. A food-truck cook works alone, fairy lights, empty street.” Frame: warm light, cook at the grill, steam. Motion: he flips something once, glances down the empty street. Beat two: “Headlights sweep in — an old bus stops, and tired passengers spill out.” Frame: bus door open, light spilling. Motion: doors open, first passenger steps down. Beat three: “The cook, grinning, lines up cups — the night just turned.” Frame: cook smiling, cups in a row, bus glow behind. Motion: he sets the last cup down and looks up. Three frames. Three prompts. Zero lines of dialogue. And it has a beginning, a change and a feeling. That is a film, at the exact scale your tools can shoot today.
Frequently Asked Questions
Can a 15-second video really tell a story?
Yes — if you accept its scale. One change, felt once. Setup, turn, land. The mistake is measuring it against a film scene. Measure it against a good joke or a photo series. Small forms with real endings.
How many words should the script be?
Three sentences is the working script. If you’ve written a paragraph of plot, you’ve written a summary of a longer video. Cut until each beat is one visible event, then stop cutting — the frames need something to hold.
What if my idea needs two locations?
Ask what the second location adds that a frame change in the edit can’t suggest. A doorway, a window, a cut from interior face to exterior light often buys the feeling of a move without the consistency cost. If the idea truly dies without it, save it for your second month. Ideas keep.
Write the story a camera can see, in three beats, in one place, around one person doing one thing at a time. It’ll feel like you’re writing smaller than your imagination — you are. You’re writing to the size of the tool in your hands today. And the strange truth of this craft is that the constraint is where the good ideas show up. The kingdom-and-dragon script can wait. It’ll still be there when your hands are ready for it. And by then you’ll know how to break it into frames.