AIEH Writing a Short-Drama Script an AI Video Tool Can Actually Shoot
The generation step is not the bottleneck. Once you have run a first clip, producing another is five minutes of work. The bottleneck is knowing what to generate — and a script written the ordinary way will not tell you.
A conventional screenplay assumes a director, a camera operator and an editor will interpret it. An AI video tool interprets nothing. It shoots exactly the shot you describe, one at a time, a few seconds long. So the script you need is not prose: it is a shot list, and writing one is the skill that makes everything downstream mechanical.
The constraint that shapes everything
One property of the technology drives every rule on this page:
A generated clip is a few seconds long and holds one clear action.
This is not a limit a higher plan removes. It is how the models work. Ask a single clip for a sequence of events — she walks in, sees the letter, and drops her keys — and you generally get none of the three cleanly.
So a two-minute episode is not a two-minute generation. It is fifteen to forty separate shots, each with one action, joined in an editor. Once you accept that, the writing problem becomes clear: your job is to break a story into single actions.
Step 1: The beat sheet
Before any shots, write the story as a list of beats. A beat is a moment where something changes. Short-form drama runs on a compressed structure — roughly this, for a sixty-to-ninety second episode:
| Beat | Seconds | What it does |
|---|---|---|
| Hook | 0–3 | A striking image or an unanswered question |
| Setup | 3–15 | Who, where, what is at stake |
| Turn | 15–40 | The thing that changes |
| Escalation | 40–70 | It gets worse or stranger |
| Cliff | 70–90 | An unresolved question that pulls to the next episode |
Two rules about the hook, because they decide whether anything else you did mattered:
- It is three seconds, not thirty. Viewers on short-form platforms decide almost immediately. A slow establishing shot is a slow way to lose the audience you already had.
- It is an image, not an explanation. A woman burning a letter in a sink is a hook. A caption saying “she found out the truth” is not.
Write the beat sheet in plain sentences. It should fit on half a page. If it does not fit, the episode is too long for the format.
Step 2: Break each beat into single actions
Now convert each beat into shots. The test for whether something is one shot is simple: can you describe it as one continuous action, with one camera behaviour? If you need the word “then”, it is two shots.
| One shot | Two shots |
|---|---|
| She lights a match | She lights a match, then drops it in the sink |
| He looks up from the phone | He looks up, walks to the window, and opens it |
| Rain hits the window | Rain hits the window as the lights go out and she turns |
Being forced into single actions is not a handicap. It is how cutting already works in real drama — most film shots are only a few seconds long too. The format is pushing you toward correct editing rather than away from it.
Step 3: Write the shot list
The shot list is the actual working document. One row per shot, and every row carries what the generator needs:
| # | Beat | Shot description | Camera | Notes |
|---|---|---|---|---|
| 1 | Hook | Hands strike a match over a steel sink, letter beneath | Static, close | Hook — must be striking |
| 2 | Hook | Flame catches the corner of the paper, smoke rises | Slow push in | |
| 3 | Setup | She watches it burn, face lit from below | Static, medium | Same character as 4–12 |
| 4 | Setup | Wide of a small kitchen at night, one lamp on | Static, wide | Establishes location |
Five columns is enough. What matters is that each row is directly translatable into a generation prompt with no further thinking — because that is exactly what you will do next, thirty times in a row.
Write the camera column for every row. If you leave it blank, the model picks, and it usually picks more movement than you wanted. “Static” is a legitimate and frequently correct answer.
Step 4: Mark your continuity anchors
Before generating, go through the list and mark anything that must match across shots:
- Character — which shots contain your lead. These all need the same avatar; see same face every episode.
- Wardrobe — the red coat has to be the red coat in all eleven shots. Put it in every prompt that contains her.
- Location — the kitchen has to be one kitchen. Generate one approved still of it and animate from that still repeatedly.
- Time of day and weather — night stays night. Models will happily give you afternoon in shot 12 if nothing stops them.
This marking pass takes ten minutes and prevents the failure that ruins first attempts: thirty technically good clips that cannot be cut together because they are not obviously the same story.
What to write around, not into
Some things are worth designing out of the script entirely, because generators are unreliable at them:
- Readable text in shot. Signs, letters, phone screens and logos come out as convincing-looking nonsense. Write the shot so the text is not legible — an angle, a blur, a hand partly covering it — and add real text in the edit.
- Exact counts. “Seven candles” gives you some candles. If a number matters to the plot, show it a different way.
- Precise spatial arrangements. “The cup exactly left of the book, behind the lamp” gets approximated. Describe the feeling of the arrangement, not the coordinates.
- Hands doing fiddly things. A hand holding something simple works. A hand threading a needle is asking for trouble.
- A specific real person’s face. Unreliable, and frequently not permitted — see rights, likeness and consent.
Dialogue in a format that does not speak
Generated clips are picture, not performance. Plan for that from the start rather than discovering it in the edit. Four approaches, in rough order of how well they hold up:
- Voiceover. Record or synthesize narration and lay it over the cut. The most reliable, and it suits the confessional tone short drama already uses.
- On-screen text. Captions carrying the dialogue. Native to the format; most viewers watch muted anyway.
- Implication. Write scenes that play without words. A look, an object, a door closing. Usually the strongest choice, and always the cheapest.
- Lip-synced speech. Possible on some tools, and the most fragile. Do not build an episode that collapses if it does not work.
Reference audio can be supplied to the video tools — up to three clips — which is worth knowing when you want motion timed to a track.
A worked fragment
Beat: the turn — she realises the flat is not empty.
Broken into shots:
| # | Shot description | Camera |
|---|---|---|
| 14 | She freezes mid-step in a dark hallway, head tilted, listening | Static, medium |
| 15 | A door at the end of the hallway stands slightly open, light behind it | Slow push in |
| 16 | Her hand tightens on the strap of her bag | Static, close |
| 17 | She turns her head slowly toward the door | Static, close |
Four shots, four generations, roughly eight seconds of finished screen time — and a beat that reads as directed rather than assembled. Shot 16 is doing the most work for the least money: a close-up of a hand costs the same as anything else and carries tension better than a wide.
Once the list is written
The rest of the pipeline runs off this document without further decisions:
- Prompting for video — turning a shot-list row into a prompt that produces it.
- Same face every episode — locking the character before you generate thirty shots of her.
- What an episode actually costs — what a forty-shot list bills at, and how to cut that.
- From clips to a finished episode — putting it together.
Last reviewed: · Editorial policy · Report an error