Timestamped prompts: one take, directed by the second
Most AI video prompts describe a scene. This one directs it: every beat has a timecode, every line of dialogue has an owner, and the camera knows exactly when to pan. The result is above — a 25-second homage to the most famous police-desk scene in action cinema, generated as one continuous shot with sound. Below is the entire prompt, unedited, and the rules that make the technique work.
The idea: prompt like a shot list
Long takes are where AI video usually falls apart — the model wanders, the pacing mushes, the punchline lands nowhere. Timestamps fix that. Instead of one paragraph describing “a tense scene at a police station”, you hand the model a shot list: what happens at 0–5, what happens at 5–8, who speaks, where the camera goes. The model stops improvising structure and starts executing yours.
The full prompt
This is the exact prompt behind the film above — one scene, 25 seconds, Seedance 2.5 with native sound. Two characters are addressed as Hero 1 and Hero 2 throughout, so the model never confuses who does what:
| Time | Direction |
|---|---|
| 0–5 | Hero 1 — a man in dark sunglasses and a black leather jacket stands at a counter behind the armor glass, looking at a Hero 2 police officer. The police officer is seated behind the counter, writing on a form. |
| 5–8 | Hero 1 speaks to Hero 2, asking about Sarah Connor. Hero 2 is writing on a form. |
| 8–10 | Hero 2 looks up from his writing and speaks to Hero 1, refusing to let him see Sarah Connor. Hero 2 then resumes writing. |
| 10–12 | Hero 1 asks Hero 2 where Sarah Connor is. Hero 2 looks up and speaks to Hero 1. |
| 12–16 | Hero 2 tells Hero 1 that it may take a while and he can wait on a bench. Hero 1 turns his head slightly to the right. |
| 16–19 | Hero 1 looks directly at the camera and says, “I'll be back.” |
| 19–22 | Hero 1 walks away from the counter, his back to the camera. He walks out of frame. The camera pans to the right, showing the outside of the building at night. |
| 22–25 | Hero 2 is writing on a form with a pencil. The camera zooms in on his hand as he writes. |
| 25–30 | Hero 2 looks at the camera. Suddenly, Hero 1 in the car smashes through the counter by car, destroying it. Hero 2 is thrown back. Papers fly everywhere. |
Why this works
- Timecodes carry the rhythm. The pause before “I'll be back” exists because 12–16 is four unhurried seconds of bureaucracy. Tension is scheduled, not hoped for.
- Named roles prevent identity soup. “Hero 1” and “Hero 2” are used in every single beat — the model never swaps the jacket and the uniform.
- One action per beat. Each interval contains exactly one thing to play: a question, a refusal, a look. The model can't rush what it isn't given.
- Camera directions are text. “The camera pans to the right”, “zooms in on his hand” — written like stage directions, executed like a dolly grip (more moves in the camera prompts cheat sheet).
- The payoff is placed last — and off-screen until it isn't. Beats 22–25 deliberately calm the scene down so the crash at 25–30 detonates.
How to write your own timestamped prompt
One prompt, one scene, one generation — at 25 seconds on Seedance 2.5 that's 400 tokens, cheaper than a coffee run for a stunt that would total a real precinct set. And because it's a single take, there are no cuts to hide continuity mistakes — the technique lives or dies on the shot list. Write a good one, and the model keeps the promise.
Describe a scene and you get footage. Direct it by the second and you get a movie.
Frequently asked questions
Does the model really respect the timecodes?
Closely, though not to the frame. Beats land in order and roughly on schedule — the four-second bureaucratic pause reads as four seconds. Keep beats at least two seconds long; sub-second choreography is asking for luck.
Which models support this?
Any model with long takes benefits: Seedance 2.5 handles up to 30 seconds with sound and exact spoken lines; Wan 3.0 adds up to 10 reference images to a 30-second take. Short-take models (5–10 s) can still use mini shot lists — two or three beats.
Can I use my own cast in a scene like this?
Yes — cast saved actors as Hero 1 and Hero 2 and their faces anchor the whole take: synthetic characters, or verified real people with consent. The prompt structure stays exactly the same.