Debug an AI Video Prompt with a Repeatable Revision Test

01 / DIAGNOSIS

Describe the miss before rewriting the prompt

A disappointing clip can fail in several different ways: the subject does the wrong action, the camera moves unexpectedly, an object changes, or the ending never arrives. Name the visible miss in one sentence before editing the prompt.

• Action: what should happen, and what did you see instead?
• Continuity: which person, object, clothing detail, or location changed?
• Camera: what was the starting view and where did the view end?
• Timing: which part of the action was too early, late, missing, or repeated?

“It ignored the prompt” is too broad to guide a useful revision. “The hand released the compass before it reached the center of the desk” identifies something you can test.

02 / TEST RECORD

Save a baseline you can compare

Keep the original prompt, input asset, model label and visible settings together with the output and review note. Record the date and the one acceptance check you are testing. Do not rely on memory to reconstruct what changed.

• Prompt and reference: keep the exact text and source file version.
• Run details: note the service, model/version shown, settings, and date.
• Acceptance check: state what a passing clip must visibly do.
• Observation: describe only what happened in this output, not what you think the model intended.

03 / CONTROLLED REVISION

Change one likely cause and rerun the same brief

Use an original fictional scene to illustrate the method: an adult museum conservator slides one brass compass from the left side of a wooden desk to its center, releases it, and leaves the camera fixed. A passing draft would show one compass move once and stop at the center.

If a draft instead stops short, keep the scene, input, model, and settings unchanged. Revise only the destination instruction: say that the compass reaches the center of the desk and stops there. Then compare the same acceptance check. This is a proposed experiment, not a report of a model result.

• If the action is missing, make the subject, verb, and visible end state explicit.
• If instructions conflict, remove one competing direction rather than adding more adjectives.
• If a reference image disagrees with the requested starting state, update the brief or choose a compatible reference.
• If several people or actions drift at once, reduce the scene to one actor and one event before adding complexity.

04 / REVIEW

Compare the changed variable, not your memory

Review the baseline and revised output against the same acceptance check. If results vary between runs, note that variability and avoid treating a single attempt as proof that a wording change caused the difference. When practical, repeat the baseline as well as the revision so you can see whether outputs vary without a prompt edit.

• What exact words changed?
• What stayed constant?
• Did the visible miss improve, stay the same, or change into a different miss?
• What is the next single hypothesis to test?

Stop revising when the output meets the project requirement or when the tool cannot reliably deliver it. More prompt text is not a substitute for a supported control or a different production method.