How to Write a Vertical AI Short Drama Script That Ends on a Cliffhanger

Vertical short dramas live or die on one moment: the last three seconds of each episode. A viewer who reaches the end with an unanswered question swipes to the next episode. A viewer who reaches a neat resolution closes the app.

Writing for AI video adds a second constraint. Every line of your script eventually becomes a prompt for a shot of a few seconds, so a script that reads well but cannot be broken into shots will fight you at every step. This guide shows how to write an episode that does both: it hooks, and it converts cleanly into prompts for Seedance 2.5, Wan 3.0 or Kling 4.0.

The shape of a one-minute episode

Vertical episodes are short, often somewhere around one to two minutes. Inside that, a dependable shape is:

Beat	Time	Job
Cold open	0-5 s	Pick up the last episode’s hook or show something strange. No setup.
Problem	5-20 s	One character wants one thing and something is in the way.
Escalation	20-45 s	The obstacle gets worse or a second character complicates it.
Turn	45-55 s	New information changes what the scene meant.
Cliffhanger	last 3-8 s	A reveal, a threat or a choice, left unresolved.

You do not need a three-act structure in sixty seconds. You need one want, one obstacle and one turn. Everything else is cut.

Five cliffhangers that work in vertical

A good cliffhanger raises a specific question the next episode can answer. Vague mystery (“something is wrong”) is weaker than a concrete one (“why is her name on the slip?”). Five patterns that hold up:

1. The reveal. A detail recontextualizes the scene: a signature, a photo, a name on a screen.
2. The arrival. Someone who should not be there walks in. Cut on their face, not on the reaction.
3. The choice. A character must pick between two bad options. Cut before they choose.
4. The overheard line. A character hears one sentence that changes everything. End on the listener.
5. The countdown. A deadline appears. End on the clock, the message or the closing door.

Each of these is visual. That matters, because the final shot is usually silent or nearly silent, and it has to land on a phone screen with the sound off.

Write every scene as shots

AI models generate shots of a few seconds, so write your scenes in units you can generate:

• One location per scene. Each scene heading becomes a location reference image.
• One action per shot. If a sentence contains “and then”, it is two shots.
• Name the speaker on every line. No “they argue”. Write the lines.
• Keep spoken lines short. At a natural pace, people speak roughly two to three words per second. A 6-second shot holds one line of about 12 words, plus a beat of reaction.
• Write sound. Rain, a door bell, a phone buzzing. Native-audio models use it.

Use a light screenplay layout: a scene heading (INT. or EXT., place, time of day), action lines in the present tense, character names in capitals, parentheticals for delivery, and dialogue under the name. Then check the runtime by hand: count the dialogue words, divide by about 2.5 for seconds of speech, and add the seconds of each action-only shot. If the total is far over your target, cut a scene rather than squeezing every shot. Consistent formatting is not cosmetic here: it lets you copy each block straight into a shot list.

Example: episode 1 of an original series

Here is a complete one-minute episode for Night Bloom, our example series about Mara, a night-shift florist, and Jun, a courier.

EP 01 - "WRONG ADDRESS"

INT. FLOWER SHOP - NIGHT

1. Rain on the window. Neon "BLOOM" sign flickers. (4s, no dialogue)
2. MARA (late 20s, copper bob, green apron) trims stems. Door bell rings. (5s)
3. JUN (30s, yellow rain jacket, delivery bag) steps in, soaked. (5s)
   JUN: Same address as last night. Fourteen Orchard Lane.
4. MARA doesn't look up. (5s)
   MARA: There is no Fourteen. The street ends at twelve.
5. JUN sets the empty delivery slip on the counter. (6s)
   JUN: Then who keeps signing for them?
6. MARA turns the slip toward herself. A beat. (4s)
7. INSERT - the slip. The signature reads: MARA. (4s, rain only)
8. MARA's face. She has never signed it. The bell rings again behind her. (4s)
   CUT TO BLACK.

That is 37 seconds of shots before titles, which leaves room for a title card and an opening recap. Notice the structure: the problem (impossible address), the escalation (someone signs), the turn (her name), and an arrival cliffhanger on the bell. We never see who walked in.

Turn the script into prompts

Each numbered shot becomes one prompt. A reliable layout, recommended on the Seedance 2.5 model page, is three layers: a one-sentence summary, a timed breakdown, then consistency notes.

Shot 5 on Seedance 2.5 Text-to-Video, with Mara as reference image 1, Jun as image 2 and the shop as image 3:

Jun from @Image 2 places a delivery slip on the counter of the flower shop from @Image 3 in front of Mara from @Image 1, night drama, medium two-shot, static camera.
0-2s: Jun lays the damp slip flat on the counter.
2-6s: He looks at Mara and says quietly, "Then who keeps signing for them?"
Warm shop light, rain on the window behind them, soft rain ambience, no music, no subtitles.

Settings: aspect_ratio 9:16, duration 6, resolution 480p for the draft, generate_audio true.

Shot 7, the insert, has no people and no dialogue, so it is a good candidate for the lower-cost Wan 3.0 Text-to-Video:

Extreme close-up of a damp paper delivery slip on a wooden counter. A handwritten signature reads "Mara". A drop of rain water runs across the paper. Very slow push-in, shallow depth of field, warm light. Only the sound of rain.

Settings: aspect_ratio 9:16, duration 4, resolution 720p, generate_audio true.

Readable handwriting in a generated insert is the most likely thing to fail. If the word comes out wrong, make the insert as a still image instead, approve it, and animate it with image-to-video, or simply keep the shot as a still with a slow zoom in the edit.

Keep the same three-layer structure in every prompt of the episode, so a change in the result comes from what you changed, not from how you phrased it. Before you generate anything, lay the shots out in order as a strip of rough thumbnails or a simple table with the duration of each. Reading the strip left to right shows the rhythm of the episode: a run of five medium shots with no close-up, or a turn that arrives too late, is easy to spot and cheap to fix at this stage.

The cliffhanger shot itself

Give the final shot more attention than any other:

• Hold it. Four to six seconds. A cliffhanger that cuts too fast feels like an error.
• End on a face or an object, not on action. The viewer should be looking at the question.
• Use sound as the hook. A door bell, a phone buzz or a single line off screen can carry the reveal while the picture stays on the reaction.
• Do not resolve it in the recap. Episode 2’s cold open should replay the last two seconds, then answer one part of the question and raise another.

For our shot 8 on Seedance 2.5 Image-to-Video, use an approved close-up of Mara as the start image so the face is exactly right, and let the prompt carry only the motion and the sound:

She stares at the slip, very still. Her eyes lift slowly toward the door behind the camera. The door bell rings once. Static close-up, the rain outside goes quiet.

Settings: duration 4, resolution 720p, generate_audio true. The output follows the start image’s shape, so make that image vertical.

Write episode 2’s cold open now

The cliffhanger is only half of the hook. The other half is the first five seconds of the next episode, and it is easiest to write both at once, while the question is fresh. A cold open does three things fast: it replays the last image, it answers one part of the question, and it opens a new one.

EP 02 - "SIGNED"

INT. FLOWER SHOP - NIGHT (continuous)

1. Replay: MARA's face, the bell ringing behind her. (2s)
2. She turns. The doorway is empty, the door swinging. (3s, rain louder)
3. On the counter: a fresh bouquet that was not there before, with a card. (4s)
4. MARA reads the card. (4s)
   MARA (V.O., reading): Fourteen Orchard Lane. Thank you for signing.

Episode 1 asked “who walked in?”. Episode 2 answers “nobody you can see” and replaces it with a sharper question: who left the bouquet, and why does the card thank her? The audience got something, so the swipe felt worth it, and they still have a reason to keep going.

Plan the season, not just the episode

A run of episodes needs a spine, or the cliffhangers start to feel arbitrary. Before writing episode 3, write one line per episode for the whole season:

• Episodes 1-3: the mystery appears and gets personal (the address, the signature, the bouquet).
• Middle episodes: each answers one question and raises a bigger one. Alternate reveal cliffhangers with choice and arrival cliffhangers so the pattern does not get predictable.
• Final episodes: the questions converge on one answer, and the last episode ends on an image that resolves the main mystery while leaving one thread for a possible second season.

Keeping that list next to the script also helps production. You can see which locations and characters recur, build their reference images once, and reuse them across every episode instead of recreating them per script.

Package the episode

Before publishing, add a 9:16 title card with the series title and episode number, and preview the episode on a phone to check that captions and the title are not hidden under the app’s buttons. The full delivery checklist is in vertical 9:16 AI video settings for Shorts and Reels, and the end-to-end production process is in how to make an AI short drama.

FAQ

How long should an AI short drama episode be? Short: often around one to two minutes. Write to the story, not the clock: an episode with one want, one obstacle and one turn usually fits in 8 to 14 shots.

How many words of dialogue fit in one AI video shot? Plan around two to three spoken words per second. A 6-second shot comfortably holds one line of around 12 words with a short reaction.

Should I write the script in screenplay format? A light version helps: scene headings, character names in capitals, dialogue under the name. The point is that each block maps to one shot and one prompt.

What makes a good cliffhanger? A specific, visual question that the next episode can answer. End on the face or object that carries the question, and hold the shot long enough for it to register.

Can AI write the script for me? You can draft with any writing assistant, but edit it into shots yourself. The structure, the turn and the final image are the decisions that make an episode work.

Next, generate episode one shot by shot in the video generator with Seedance 2.5, Wan 3.0 or Kling 4.0 preselected.