A reference video is useful because it contains decisions that are difficult to communicate with a style label. It shows when a cut happens, how fast a camera moves, what crosses the foreground, when music changes, and how a subject remains recognizable between shots. A video-to-prompt workflow turns that evidence into written direction you can inspect and adapt.
The task is not to claim how the original was made. A finished clip cannot prove which camera body, private prompt, lens brand, or production conversation created it. Good analysis stays grounded in what is visible and audible, then separates those observations from creative rewriting.
Choose a focused reference
Start with the shortest clip that contains the behavior you want to understand. A thirty-second product edit is often more useful than a three-minute compilation with several unrelated visual languages. Public links and direct uploads should point to the exact media, not a playlist, profile, or feed.
Check source quality before analysis. Heavy compression can erase texture, motion blur, and small focus changes. Baked-in captions or interface controls may be interpreted as elements of the frame. Private, login-only, region-blocked, live, or DRM-protected sources may not be available to a server-side downloader.
Define your question. Are you studying camera motion, subject choreography, lighting, transitions, dialogue, or the full structure? A clear question helps you decide which evidence must survive in the final prompt.
Build a timeline before writing prose
Divide the reference at meaningful changes: cuts, major reframes, new actions, lighting changes, or sound transitions. Each segment should have a start and end time and one primary visual purpose.
For every shot, record:
- framing and camera distance;
- subject position and direction of travel;
- visible action in chronological order;
- camera movement, speed, and direction;
- motivated light, palette, contrast, and atmosphere;
- transition into and out of the shot;
- dialogue, effects, ambience, music, or silence.
Do not force every category to contain an answer. A locked shot has no camera travel. A silent reference has no dialogue. Recording absence is more honest than inventing detail.
For videos longer than two minutes, keep sequential segments instead of collapsing the ending into one summary. Long references often change visual grammar halfway through; preserving the timeline prevents an attractive opening shot from dominating the full prompt.
Separate observation from inference
“The subject becomes larger in frame while perspective changes gradually” is an observation. “The operator used a 35mm lens on a motorized dolly” is an inference unless the equipment appears on screen. The prompt needs the visible result, not the unsupported production story.
Use the same discipline for emotion. Instead of “she feels uncertain,” record the evidence: she pauses at the doorway, looks away, tightens her grip, and enters only after another person speaks. Observable behavior gives an AI video model something it can stage.
Style should also be described through properties. Record color relationships, contrast, texture, grain, depth of field, flare, haze, and movement. A list of artist or director names hides the decisions and makes the result harder to control.
Find continuity anchors
Multi-shot references work because some information changes while other information remains stable. Identify the anchors: a coat color, hairstyle, prop, direction of movement, time of day, background geography, or sound motif.
Then record how the edit carries motion across the cut. A subject exits frame right and continues in the same screen direction. A rotating wheel match-cuts to a circular bowl. A foreground object wipes the frame and reveals a new location. These relationships are more useful than simply naming “fast editing.”
When adapting the reference, protect the structural anchor but change the creative identity. You might keep a three-shot directional sequence while replacing the cyclist, city, wardrobe, and palette. That uses the reference as a grammar rather than trying to duplicate its content.
Translate evidence for the target model
The generic analysis is a factual record. Model-ready prompts reorganize it.
For Kling, foreground physical motion, contact with surfaces, stable anatomy, and one legible camera move. For Veo 3, identify speakers and connect effects, ambience, and music to the timeline. For Seedance, group connected shots, repeat compact identity anchors, and name motivated transitions. For Sora, preserve world rules, spatial depth, composition, and narrative change. For Runway, reduce the direction to one strong visible action and camera behavior when generating a single shot.
Do not add model syntax that contradicts the reference. If the original is a locked wide composition, a dramatic orbit may create a new idea, but it no longer explains why the source works. Mark creative changes explicitly so collaborators can distinguish analysis from adaptation.
Review the generated prompt
Compare the written result with the timeline, not just with your memory of the clip.
- Does every shot appear in the correct order?
- Is the ending covered?
- Are camera and subject motion described separately?
- Do transitions connect compatible shapes, directions, or actions?
- Are light sources and time consistent?
- Does dialogue fit the available duration?
- Are unsupported claims removed?
- Is the prompt describing a new production rather than promising an exact clone?
Copy one model format only after this review. Exporting the complete analysis as JSON or Markdown is useful when the timeline, audio cues, and alternative formats need to travel together.
Test one production decision at a time
A prompt is a hypothesis about what made the reference effective. Generate a short version, compare the result with the evidence, and revise the instruction the model missed. If motion is weak, clarify direction, speed, and contact. If the frame drifts, request a locked camera or one explicit move. If identity changes, shorten the sequence and strengthen the continuity anchor.
The best video-to-prompt workflow does not end with a large block of text. It creates a traceable path from source evidence to an editable brief, then to a test you can evaluate. That makes the reference useful even when the new generation looks different—because you can explain which production decisions survived and which need another pass.