Published October 2026 · 10 min read
You spend an hour tuning a prompt. It produces exactly the shot you wanted. Then you try the same prompt on a different model — cheaper, faster, or better suited to a different shot — and it collapses. The subject drifts, the camera ignores you, the light goes flat.
This is not a bug and it is not your fault. Prompts are not portable in the way people assume. They are closer to a dialect than a specification: the same words carry different weight depending on who is listening.
Understanding why makes migration predictable instead of a guessing game.
Each model learned “what a video looks like” from a different corpus. If one was trained heavily on cinematic film and another on social media clips, then an unspecified shot will default to the visual grammar of that source. Ask for “a person walking” with no other detail and one model gives you a film still, the other gives you something that looks like a phone video.
This is why a prompt that feels complete on one model feels under-specified on another — the second model's default is further from what you had in mind.
Models have different assumptions about how much things should move. Kling tends toward restraint; Seedance toward energy; Sora toward narrative continuation. The same phrase “she turns” produces a subtle head movement on one and a full body pivot on another.
When migrating, the first thing to check is whether the motion scale changed. It usually did.
Some models treat a prompt as a checklist and try to satisfy every clause. Others treat it as a mood board and render the overall impression. If you migrate from a literal model to an impressionistic one, clauses start getting dropped — not because the model failed, but because it prioritised differently.
On a literal model, adding detail improves precision. On an impressionistic model, adding detail dilutes.
Filters are not identical across providers, and they fire silently. A prompt with a word that is fine on one platform may trigger a soft refusal on another — the result is not an error message but a bland, evasive render that ignores the interesting part of your request. If a migrated prompt suddenly produces something generic, a filter is a real possibility.
These are tendencies we have observed, not official specifications. They are consistent enough to be useful when migrating.
| Model | Prefers | Weakness when migrating |
|---|---|---|
| Veo | Precise cinematography language: lens, light source, camera move | Underspecified prompts default to a generic commercial look |
| Kling | Detailed subject description and style references | Over-smooths faces; motion tends to be conservative |
| Sora | Longer, narrative, conversational prompts | Will extend the scene beyond what you asked |
| Seedance | Explicit motion description and energy | Punishes vague camera combinations; may drift from the subject |
A practical consequence: a prompt written for Veo and moved to Kling will often lose its lighting precision, because Kling weights subject description more heavily and treats the lighting clauses as secondary. The fix is not to add more lighting words — it is to move the lighting clauses earlier and make them more concrete.
When moving a working prompt to a new model, do this in order. Do not skip to step four.
Generate one attempt with the prompt exactly as it is. This gives you a baseline and tells you what kind of failure you are dealing with. Skipping this step is why people over-correct — they rewrite everything when only the motion scale was wrong.
Compare the output to your intent across the six dimensions: subject, action, setting, light, camera, texture. Usually exactly one has degraded. Common patterns:
Change one thing. Re-run. If the broken slot recovers, you are done. If it does not, the problem is elsewhere and you have learned something. Changing three things at once means you cannot tell which one worked — this is the most common process error in prompt tuning.
Counter-intuitively, migrated prompts usually need to get shorter, not longer. The original prompt accumulated clauses that compensated for the first model's specific weaknesses. On a different model, those clauses are noise competing for attention. Cut anything that is not load-bearing and see if quality improves — it often does.
Migration has a cost, and sometimes rewriting from scratch is faster. Rewrite when:
A useful rule: if the shot is simple, migrate. If the shot is complex and already working, keep it on the model it works on and route something else instead. The whole point of routing is that you do not have to make every shot work everywhere.
Concrete beats abstract. Here is a real migration of a shot from Veo to Kling, with the reasoning at each step.
The intent: a woman in her forties at a kitchen window, early morning, turning to look outside. Five seconds, medium shot.
Original prompt (worked well on Veo):
medium shot, 42-year-old woman at a kitchen window, soft directional morning light from camera right, cool shadows, slow head turn toward the window, static camera, shallow depth of field, 35mm film look
Result on Kling: the light went flat and the woman read as significantly younger. The head turn was smaller than intended.
Diagnosis. Three slots degraded, which is unusual — and the reason is that Kling weights the subject slot most heavily, so the trailing lighting clauses lost influence. Two things follow: the subject description needs to be stronger, and the lighting needs to move earlier.
Migrated prompt:
42-year-old woman, fine lines around the eyes, no makeup, at a kitchen window, soft morning light from camera right creating a shadow across the left side of her face, turning her head slowly to look outside, medium shot, static camera, shallow depth of field
What changed and why:
Net effect: same length, different order, two clauses added and one removed. That is typical of migration work. It is rarely a rewrite and usually a reordering plus one targeted addition.
Note also what did not change: the intent, the shot size, and the duration. Those three are portable across every model, because they describe the shot rather than the rendering. If you find yourself changing the intent during a migration, you are not migrating a prompt any more — you are redesigning the shot, and that is a different task with a different cost.
The practical answer to non-portability is to store prompts per model rather than per shot. Keep a note for each recurring shot type with one version per model you use. After a few months you have a translation table that makes migration nearly free.
Two things make this work. First, store the intent alongside the prompt — a one-line description of what the shot is meant to achieve, so future-you can rewrite rather than guess. Second, note the retry count each version needed. A prompt that works on the second attempt is worth keeping; one that needed seven is a warning sign that the model is wrong for the shot.
Model-specific guides on this site cover the vocabulary each engine responds to: Veo, Kling, Sora, Seedance. The six-slot framework gives you the structure that makes slot-level diagnosis possible in the first place — without a consistent structure, step two of the migration method has nothing to compare against.
And if you are choosing which model to migrate to, the model selection decision tree covers that separately from the prompt question.