Name each role once
Use stable labels such as Person A and Person B. State which uploaded portrait belongs to each role and keep those labels unchanged throughout the shot list.
Avoid asking both characters to transform, swap places, and change clothing in the same short clip. Each extra change makes continuity harder to evaluate.
Describe shots rather than adjectives
A usable sequence might be: Person A looks down an empty street; Person B appears at the far end; the camera moves closer as they recognize each other. This gives the scene a beginning, movement, and payoff.
Add lighting and mood after the action: overcast dusk, wind through loose paper, and a restrained handheld camera. Keep on-screen text out of the generation prompt if you plan to add it during editing.
Evaluate the output against the prompt
Check identity, role order, framing, duration and any requested sound independently. A visually dramatic clip can still fail the central story if the wrong person takes the wrong role.