MATCH THE CLOCK
Your prompt has a duration. The model has a duration setting. They are two different numbers in two different places, and nothing checks that they agree. When they disagree the shot fails — and it fails in a way that looks exactly like a bad prompt.
Why this one wastes so much time
Both failures send you back to the words. The shot feels rushed, or it feels padded and strange, and the obvious response is to rewrite the description. So people rewrite it three times, get three bad renders, and blame the model — while the number in the dropdown, which nobody looked at, stayed wrong the whole time.
Check the clock before you touch a single word.
THREE STATES, TWO OF THEM BROKEN
The action you described takes about as long as the model is set to render.
what you see- — The shot starts and finishes on the action you wrote
- — Nothing is rushed and nothing is held
- — The last beat lands with a little room after it
The model has exactly enough timeline for the thing you asked for, and none spare to invent with. This is the only state where the prompt is being tested on its own merits.
what to doNothing. Judge the shot now — any remaining fault really is the prompt.
If you have to be wrong, be wrong with the setting too long. A padded render still has a usable front section you can trim. A compressed one is wrong in every frame and trimming cannot save it.
HOW LONG IS YOUR PROMPT?
A prompt does not come with a duration printed on it, so you have to work it out. Count the beats — every discrete physical action, plus every camera move, because a move is a beat too. Then give each one the time it needs to read.
The audience misses it. It did not happen.
The working range. A beat lands and you feel it.
It drags, and the model starts inventing inside the hold.
Worked, on a real prompt
Beat 09 of the Roof Radar pack — the contract signing.
So the setting is six. Not five, which rushes the withdrawal, and not eight, which leaves two seconds of a clay hand doing something you never asked for over a document you cannot afford to have redrawn. The arithmetic takes fifteen seconds and it is the difference between a usable plate and a re-render.
TIMESTAMPS MAKE IT STRICTER, NOT SAFER
Writing beats against a clock — 0–2s hand enters · 2–4s signs · 4–6s withdraws — is good practice. It also states your duration out loud, which means the setting is no longer a judgement call. It must equal your last timestamp. Exactly.
Timestamps against a clock that does not exist are worse than no timestamps at all. The model tries to hit marks that will not fit in the time it has, so instead of one compressed action you get every beat individually wrong.
BEFORE YOU HIT GENERATE
- 1Count the beats in what you wrote.Actions and camera moves both. Multiply by two seconds.
- 2Set the model's seconds to that number.This is the control most people never look at twice.
- 3State the same number in the prompt text.Our house rule already: cap every render explicitly, or the model overruns.
- 4If you used timestamps, the last one equals the setting.No rounding. No nearly.
- 5Change one, change all three.This is how the number drifts out of agreement — a prompt edit that never reached the dropdown.
THE LAW UNDERNEATH
The same number living in two places, with nothing forcing them to agree, is one of the most reliable ways to lose a day in this work. The truck-door logo shot lost one to exactly this: a hard-coded frame range, a hard-coded end time, and a zone start time, three independent places each quietly discarding good footage, all reporting the same useless symptom — “it shows up late and disappears early.”
When something is wrong with timing, there is rarely one cause. Find every place the duration is written down, and make them all say the same thing.
Related: Test it in motion is how you find which input caused a failure — and the duration setting is an input, even though it is not part of the prompt.