Text-to-video prompts fail in a specific way: they describe a story when the model needs a shot. "A founder walks into an office and explains her product" is three events and a camera position; no current model holds all of it.
The fix is to write what a director writes — a shot list — and to accept that the sequence is assembled later.
The short answer
One shot per generation. Specify subject, action, camera, lighting, and duration. Do not describe plot, emotion, or dialogue; those belong in the edit and the voiceover. Then budget retries, because the first take is rarely the usable one.
The five fields of a shot
- Subject. Who or what is in frame, described physically rather than narratively. "A woman in a grey knit sweater at a wooden desk" beats "a thoughtful designer".
- Action. One motion, present tense. "She turns a page" is a shot; "she reviews the document and makes a decision" is three shots and an inference.
- Camera. Static, slow push in, pan left, handheld. Camera language is well represented in training data, so it works.
- Lighting. Window light, hard sun, overcast, practical lamps. Vague lighting produces flat output.
- Duration. Short. Five to eight seconds is where coherence holds; beyond that, drift is visible.
A brief that works
| Field | Weak | Strong |
|---|---|---|
| Subject | a professional | a man in his forties in a navy blazer, seated at a white desk |
| Action | working on something | typing on a laptop while glancing at a second screen |
| Camera | dynamic | slow push in from medium shot to close-up |
| Lighting | nice | overcast daylight from a window on the left |
| Duration | a while | six seconds |
The weak column is not wrong so much as undecided. Every undecided field is a decision the model makes for you, and it will not match the next shot.
Start from a still
The most reliable way to control a video generation is to control its first frame. Generate or photograph a still that looks the way you want the shot to start, then use image-to-video (Kling is built around this) or a reference-frame mode (Veo).
This changes the failure mode. Instead of asking "does the whole clip look right", you are asking "does the motion respect the frame I gave it", which is a narrower and more answerable question.
The retry arithmetic
If a usable clip takes six attempts, the real cost is six times the sticker price. That number should drive the shot list: fewer, simpler shots are cheaper than one ambitious shot that needs twenty takes.
Practically, change one field per retry and keep notes. If the first three attempts fail on the same element — hands, text, a specific object — the shot is beyond the model's current abilities and should be filmed or composited instead.
What to composite instead
Three things are worth doing outside the model every time:
- Legible text. Logos, UI, signage. Generated lettering is wrong in ways that are obvious to viewers.
- Hands interacting with objects. The classic failure. Shoot a plate of hands and composite, or design shots that avoid it.
- Precise product detail. A specific device, label, or packaging will be invented rather than reproduced.
Bad briefs, and what they have in common
The weak briefs share a structure: they describe an intention rather than a frame. "A sense of discovery" and "professional but warm" are notes to a human collaborator who can interpret them; a model treats them as noise and fills the gaps with its average.
The second common failure is stacking events. "She opens the laptop, reads the email, and looks worried" is three shots. A model asked to produce all three in one clip will compress them into something that looks like a glitch.
The third is describing the camera as an outcome rather than an instruction. "It feels cinematic" is not a camera move; "slow push in, shallow depth of field" is.
The two-sentence template
First sentence: subject, action, camera. Second sentence: lighting and duration. That is enough to control most generations, and short enough to iterate on.
"A woman in her thirties at a kitchen table turns a page of a printed report, medium shot, slow push in. Overcast morning light from the left, six seconds." Every field in that brief is a decision that would otherwise be made for you.
Frequently asked questions
How long should a generated shot be? Five to eight seconds. Longer shots drift in subject and lighting; the edit is where length is decided.
Can I generate a conversation? Not reliably. Generate each angle separately and cut between them; lip-sync is a separate problem with its own tools.
Which model should I start with? Whatever you can iterate cheaply on. Sora, Veo, and Kling all reward a well-written shot brief; the workflow matters more than the pick.
Last checked: 9 November 2026.
Related: Sora vs Veo vs Kling, the media asset pipeline, and video tools.