Begin with the behavior
Describe what the model should observe or predict. “Kitchen video” leaves substantial room for interpretation; a sequence such as finding a utensil, opening a container, and returning an item gives a sourcing conversation a concrete starting point. List essential activities separately from useful extras.
Define the participant’s view
Specify where the camera sits and what must be visible. Head and chest positions can frame hands and objects differently. Ask for representative footage before committing to a capture specification. Include tolerance for occlusion, motion, pauses, and changes in lighting.
Plan variation intentionally
State which differences matter: participant, location, tools, object arrangement, or task order. Multiple recordings of one setup may be useful, but they should not be described as equivalent to broad environment coverage. Identify which conditions belong in evaluation and which belong in training.
Agree on a review unit
Decide whether acceptance is based on full sessions, complete tasks, or individual clips. Request the metadata needed to connect each unit to its activity and capture conditions. Keep technical review, permission review, and commercial terms explicit before a delivery is approved.