How do you write AI prompts for over-the-shoulder shots that look natural?

A natural over-the-shoulder prompt names three depth layers — a soft foreground shoulder, a sharp subject at eye level, a lit background — plus a 50–85mm lens feel and a named light source, assembled in a fixed element order and ending with a negative prompt that excludes anatomy drift, extra figures, and flat lighting.

Assemble the prompt in a fixed element order and describe the depth layers explicitly — one documented production held a 9-element assembly order (camera spec, lens and aspect ratio, lighting source, palette, composition, atmosphere, mood register, film attribution, negative prompt) across every frame of the film. A labeled template you can adapt:

"Over-the-shoulder shot from behind [Character A] — [A]'s shoulder and back of head soft in the near foreground, [Character B] framed at eye level in sharp focus, shallow depth of field, 85mm portrait-lens feel, [named light source, e.g. warm practical lamp light from camera left], [palette and mood line], in your film's aspect ratio. Negative: no warped shoulder anatomy, no third character, no flat even lighting."

Each element does a specific realism job. The soft foreground shoulder is the depth cue that separates an OTS from a flat two-shot; eye-level framing reads as a listener's perspective; 50–85mm lens language compresses the space the way real dialogue coverage is shot. Name the light source instead of using an adjective — specifying "warm yellow from the lamps only, like all the refs" produces more accurate results than generic "warm lighting," a correction documented in one production's lighting passes. And keep the shot size medium or close: in an extreme wide the shoulder stops registering as a frame anchor, so the OTS composition collapses.

Attach character references to every OTS prompt, because the frame contains two characters and both must hold. Use multi-angle character sheets that include close-up panels so small details — scars, accessories — survive the angle change; one documented 70-second film kept two characters consistent across every scene this way, with no LoRA fine-tuning, by keeping the sheets loaded in the invideo agent's context so they ride along with every prompt automatically.

Once the OTS lands, you can build the matched reverse by changing only the camera line and keeping every other element fixed — coverage pairing is its own workflow, but the rule of thumb is one variable per re-prompt.

Before spending credits, run a quick naturalness check against the prompt itself: shoulder visible and soft in the foreground, subject sharp at eye level, light source named and matched to the scene's established lighting, lens feel in the 50–85mm range, shot size medium or closer, negative prompt excluding anatomy drift and extra figures.

Watch some of these to see what works for you:

the AI always needs to see what the character is exactly, right? Or else it'll kind of hallucinate and imagine something that's under the cap. So, we don't want to do that. We always want the character to be seen as we see it on the character sheet.

— invideo's creative team