Video-to-prompt reverse engineering guide
When you see an AI video clip you love, the fastest way to recreate the look is to break it into the same components a prompt is built from: camera movement, subject, lighting, motion type, and style. This guide gives you a structured checklist and assembles your observations into a clean, recreatable prompt you can paste into Runway, Kling, Sora, or Luma and iterate on.
Why reverse order matters
Most people who try to recreate a reference clip start by describing what the clip is about — “a woman in a café”. That alone produces hundreds of possible outputs. The professional approach starts with how the camera sees it, because camera movement is the single biggest determinant of feel, and it is the element most prompts skip.
Camera movement defines:
- The viewer’s emotional distance from the subject (push-in = intimacy, wide orbit = spectacle)
- The implied budget and style (handheld = indie gritty, locked-off = commercial clean, smooth dolly = cinematic)
- Whether the subject or the environment is the story
When your recreation feels wrong, it is almost always the camera move that diverged. Getting that right first lets everything else fall into place.
How it works
Reverse-engineering is observation in a fixed order. Camera movement comes first because it frames everything — a slow push-in feels nothing like a handheld orbit. Next, the subject and action: one clear sentence of what’s there and what it does. Then lighting (quality, direction, color temperature, contrast), which carries most of the mood. Motion type distinguishes ambient drift from deliberate action. Finally style — film stock, grade, lens, era. The builder joins these in the order video models parse best, so the result reads like a purpose-written prompt rather than a list of guesses.
Observation checklist
Work through these in order on your reference clip:
- Camera move — static, push-in, pull-back, pan, tilt, orbit, handheld drift, aerial descent, cut (these are separate clips)?
- Subject — what is it, where is it in frame (foreground, midground, background)?
- Action — what is the subject doing, and how fast?
- Lighting — quality (hard/soft), direction (front-lit, side-lit, backlit), color temperature (warm/neutral/cool), time of day?
- Ambient motion — wind in leaves, moving water, smoke, particles, crowd?
- Depth — is the background sharp or blurred? Is there visible depth of field?
- Style — film grain, color grade, lens aberration, era (contemporary, vintage, sci-fi)?
Describe each answer in a short phrase. The tool assembles them in the order models respond to best.
Tips for closing the gap
- Camera first, always. If the recreation feels wrong, it is usually the camera move, not the subject.
- Be specific about light. “Soft warm window light from the left, low contrast” beats “nice lighting” by a mile.
- Separate ambient from deliberate motion. Drifting smoke is ambient; a person turning is deliberate — name which you see.
- Iterate on one variable. Once you have a draft prompt, change a single element per generation so you can tell what moved you closer.
- Label approximation. You are describing what you see, not what the original prompt said — some gap is normal. Use “approximately” or “similar to” as a reminder to yourself when iterating.