Write AI video prompts that assign every reference
Separate identity, action and scene in a reference-video prompt, and know when a fixed template already handles the instructions.
A useful reference prompt explains where the people come from and which parts of the video guide the result. Adding more adjectives does not resolve a conflict between a portrait, a motion clip and an unrelated scene request. Give each input one clear job first.
Check whether the instructions are editable
CastRemix has fixed templates and an editable composer. In a fixed template, the reference clip and performance instructions are supplied for you. Prepare the labeled photos and review the output settings instead of looking for a hidden prompt field.
Use the editable composer when you are bringing your own material and need to describe its roles. The examples below belong to that workflow. They are starting instructions, not exact-result guarantees.
Assign identity before describing style
Add your photo and video, then use the material labels or @ menu to insert the identifiers actually shown in the draft. For one performer, start with:
Use @Image1 for the performer's face, hairstyle and clothing. Recreate the movement and framing shown in @Video1 with that person in one continuous shot.
For a two-person clip, identify stable roles rather than whichever side someone occupies after a crossing. “The person who starts nearest the camera” can be clearer than a left/right instruction when the performers change places. Match the wording to what is visible in your reference.
Decide which source supplies the scene
An image often contains a background as well as a person. If the intention is to keep the video setting, say that the photo provides the character and the video guides the scene. If you want to keep the photo's setting instead, make that a deliberate alternative brief.
Avoid asking to preserve both incompatible locations. “Keep the outdoor garden from @Image1 and the enclosed studio from @Video1 unchanged” does not identify a coherent result. Make one scene choice before adding light or color instructions.
Describe the important action in order
Name two or three observable beats: raise the hand, turn toward the partner, then lower the hand. Use the video reference for their timing. Instructions about a hand touching a microphone or a foot landing on the floor are more testable than “perfect cinematic quality.”
Keep the first brief narrow. A new costume, an added person and a moving camera each introduce a separate change. If they are essential, add them deliberately in later tests after you understand the simpler result.
Turn the prompt into a review checklist
After generation, inspect the named identity, action and setting separately. If the person looks right but the hand misses the prop, rewrite only that interaction or try a more compatible reference. If several changes fail together, simplify the brief before repeating it.
For fixed duets, follow the role-assignment guide. For a result that changes appearance mid-clip, use the character-drift diagnosis to isolate the first visible problem.