Most AI video prompts describe a scene. A POV prompt has to describe a camera β and that is a harder problem than it looks.
The usual failure mode is not a bad scene. It is drift. The first five seconds are perfect first person. Then, halfway through, the model cuts to a shot of the dragon from behind, or a face appears in a reflection, and the clip quietly stops being POV. The scene was fine. The camera did not survive.
So the prompt needs to treat the camera as a set of invariants, not a vibe β rules the model cannot negotiate with, stated plainly, every time.
This is the structure I use. Every line earns its place by preventing a specific failure:
role: >
An invented part for the viewer, in one line β a dragon in flight,
a market trader, a rider on a water slide.
Invented only, never a real person.
setting: >
One invented place, one time of day, one kind of light.
No real place names.
opening_action: >
One clear action in the first two seconds. A wing beats,
a shutter lifts, water splashes over the lens.
camera:
- eye level, one continuous take, no cuts, no jumps
- about 4 seconds per beat, slow moves
- hands, wings or tools visible at the frame edges
- never: a face, a mirror, a third-person cut
audio:
- close sounds first: breath, footsteps, a wingbeat, a splash
- then the world opens up: chatter, wind, the roar of the slide
- no music, no narration, no dialogue
ending: >
One wide, slow reveal in the last five seconds.
Awe, not dread. Never a disaster.
output: 30 seconds, vertical 9:16, photorealistic 3D animation
disclosure: >
"AI-generated 3D animation" on screen in the first 2 seconds
and in the caption, every episode.
A few notes on why it is shaped this way:
The role is one line, not a paragraph. A long character description gives the model room to invent a face. One role, one line, and the viewer only ever appears as hands, wings or tools at the frame edges.
The setting is one place and one light. Two locations in one prompt is an invitation to cut between them. One place, one time of day, one light.
The camera rules are a deny-list. "Stay in first person" is a suggestion; "never a face, never a mirror, never a third-person cut" is a constraint. When a generation breaks it anyway, the deny-list tells you exactly what went wrong.
Sound is specified as a sequence, not a mood. Close sounds first, wide sounds second. Music and narration are explicitly excluded β a voiceover turns a POV clip into a documentary, and the format's whole point is that the viewer is inside it.
The cheapest fix for drift is to stop generating motion from text alone. Generate three stills first β the wings at the frame edge, the hands at the stall, the top curve of the slide β check them against the deny-list, and attach them as references to every generation after that.
If a still comes back in third person, fix the still: a bad still never becomes a good video, and a still is far cheaper to regenerate than thirty seconds of footage.
Keep a short ledger per episode β role, setting, light, the three anchor frames, the one close sound. Two minutes of checking per episode is what turns ten clips into a series.
| What you see | What to change |
|---|---|
| Cut to third person mid-clip | Slow the camera down, fewer moves, longer beats |
| Face or mirror appears | Restate the deny-list, regenerate the anchor stills |
| Wings or hands morph | Attach the same reference stills to every generation |
| Sound drifts into music | Name the one close sound in the prompt, mute anything else |
| Opening feels empty | Put the opening action in the first two seconds, no slow pan |
The prompt is written for a tool that takes reference images and generates sound with the picture β that is what holds a first-person camera steady across thirty seconds. Seedance 2.5 (released July 31, 2026) runs up to 30 seconds in one pass with sound, up to 50 reference assets, up to 1080p, no 4K. Seedance 2.0 tops out at 15 seconds per take, so use it as the cheap test tier.
I will not quote credit prices here β they change by the month and by region. Check inside your own account before spending on the full take.
First, this prompt has not been run end to end and verified frame by frame. Treat episode one as a test: stills first, a short clip, check the camera holds.
Second, the honesty rules are part of the prompt's spec, not an afterthought. The "you" is always invented. No disaster scenes β never a plague, a reactor accident, or a war. And the disclosure goes on screen and in the caption, every episode.
Get the camera rules right and the format does the rest.
More on how we build these at https://editlogic.io/