What to write, in what order, and which of it the model will actually act on.
Long prompt limits tell you the model expects structure, not volume. Where the useful length actually sits and what to cut first.
Generative video APIs tend to advertise a generous prompt limit. Several thousand characters, sometimes more. And because the number is printed in the documentation next to the field, it reads like a target.
It is not a target. It is the point at which the request is rejected. The point at which the prompt starts hurting the output arrives long before it, and the distance between those two points is where most disappointing clips are made.
Three things go wrong as a prompt grows, and they go wrong in a specific order.
Dilution. A model attends across the whole prompt. Every sentence you add reduces the relative weight of every sentence already there. The instruction you care most about — the one thing that makes this shot this shot — arrives at sentence two and is competing with forty others by the time you submit.
Contradiction. This is the big one, and it is almost always unintentional. You write "warm afternoon light" in the lighting paragraph and "cool, clinical atmosphere" in the mood paragraph. Both are reasonable. Together they are a coin flip, and the model resolves them by averaging rather than by choosing. Long prompts contain more pairs of sentences, and the number of pairs grows quadratically.
Specification of the unimportant. Long prompts are usually long because the writer described everything at the same level of detail. The colour of a background chair gets the same sentence budget as the subject's action. To a model, that ratio is the instruction about what matters.
The useful reframe: a prompt that has grown to a thousand words is rarely too detailed. It is usually unstructured, and length is how that shows up.
Most video systems want roughly the same three things, in roughly this order:
That is a complete brief for the overwhelming majority of shots, and it fits in well under a hundred and fifty words. If you cannot say what you want in that structure, adding more prose will not fix it — you have either got two shots pretending to be one, or you have not decided something yet.
The structural version also fails better. When a shot comes back wrong, a three-part prompt tells you which part to change. A five-paragraph prompt gives you no isolation at all, and the usual response — regenerate with a different seed — is how people burn a budget without learning anything.
Seeing the skeleton laid out helps more than reading about it; a prompt generator that prints the three-field skeleton as you type makes the shape visible, and the moment the fields are separate on screen it becomes obvious how much of a long prompt was the same claim stated three ways.
When a prompt is too long, cut in this order. It is close to a strict ranking of least-to-most useful content.
What survives is short, and it is the prompt you should have written.
There is an exception, and it is worth stating precisely so it does not get used as a licence.
Dialogue. If the clip contains spoken lines, the lines themselves are content, not description, and they can be as long as the delivery time allows. They do not dilute in the same way, because they are not competing to describe the same thing — they are the thing.
Similarly, a negative list — an explicit short list of what must not appear — earns its characters, because exclusions are the one instruction type that cannot be inferred from anything else in the brief.
Everything else: shorter.
Write the prompt. Then delete half of it, keeping the half that a stranger would need in order to draw the frame. Generate both. Nine times out of ten the short one is closer to what you meant, and it is cheaper to iterate on.
You can run that experiment in about ten minutes with any hosted tool — minimax-h3ai.video is one that shows the structure while you type, which makes the deletion pass considerably less painful than staring at a plain textarea.