Prompt From Video / Scene intelligence
Start with an idea or a reference video, then build a structured prompt tailored to the AI video model you want to use.
Track a red rally car laterally through a sunlit desert bend. Keep the vehicle sharp while a broad dust plume builds behind it with natural motion blur…
01The output stays grounded in visible evidence: shots, timing, movement, lighting, and sound.
Start from a blank page
An original idea and a reference clip are different inputs. When you already know the scene you want, start with the idea builder below. It separates the concept from camera, lighting, audio, and model direction so you can see what each phrase controls.
The builder creates an editable draft in your browser. It does not pretend that one template is a finished film: review the output, remove contradictions, and adjust the visual decisions for the duration and aspect ratio you plan to generate.
Editable output
A production brief, not a bag of adjectives
Strong prompts make relationships explicit. The subject acts inside a defined space; the camera observes from a deliberate position; light reveals the important information; sound follows visible causes. Use the six-part checklist to find what is missing before you spend generation credits.
Identify the main subject, its defining traits, and the physical place before adding style.
Describe the opening state, the important change, and where the movement settles.
Choose framing, lens distance, focus behavior, and one primary camera move.
Name motivated light sources, palette, contrast, texture, and atmosphere.
Connect dialogue, effects, ambience, music, or silence to the visible timeline.
Preserve identity and continuity, then use syntax that suits the selected video model.
Copy, complete, and revise
Start with this order: subject and setting; action from opening to final state; framing and one camera move; motivated light and visual finish; sound; then continuity constraints. The labels are a drafting aid, not commands your model needs. Remove the labels after every field describes one compatible result.
For example: a florist closes a street stall at dusk; she folds the awning, lifts one last bucket, and exits frame; a medium-wide camera tracks slowly left; warm shop light meets cool rain; canvas snaps, buckets scrape, and distant traffic remains low; preserve her red coat and the wet street layout throughout.
Replace emotion labels with behavior the camera can record. Instead of “she feels nervous,” write that she checks the empty street twice, tightens her grip, and pauses before stepping into the rain. Replace “cinematic” with the framing, lens distance, light source, contrast, or movement that creates the intended look.
Keep the scene facts stable, then change emphasis. Kling needs legible physical motion. Veo benefits from sourced, timed audio. Seedance needs shot-to-shot continuity. Sora benefits from clear spatial and world rules. Runway usually works best when one action and one camera move remain concise.
If you begin from an image, treat the still as the opening frame. Do not spend half the prompt repeating what is already visible. State which details must stay fixed, what begins to move, how the environment responds, and whether the camera follows or stays locked.
Revise once for contradictions and once for timing. A five-second shot cannot comfortably contain several locations, four camera moves, long dialogue, and a complete transformation. Split that concept into separate shots, or keep the one change that carries the idea. A useful prompt is specific because its decisions agree, not because it contains the most words.
Keep a small version history when you test the prompt. In the first generation, confirm the subject and main action. In the second, change only camera behavior. In the third, refine light, texture, or audio without rewriting the action that already works. Record which model, duration, aspect ratio, source image, and seed produced each result. A video prompt generator can organize the written direction, but side-by-side tests reveal how a particular model interprets it. Controlled variants also make failures useful: you can tell whether a weak result came from an unclear verb, conflicting motion, overloaded timing, or a style instruction that changed the scene more than expected. Keep the version that solves the main visual problem during the next generation test, even when a longer draft sounds more impressive on the page.
Choose the right starting point
Start from an idea when you are inventing the subject, world, and action. Start from a video when timing, framing, edit rhythm, or sound in an existing reference is the part you need to understand. The reference analyzer above maps those decisions to timestamps and returns six formats without publishing the uploaded clip.
Neither route copies a finished video. The useful output is a set of production choices you can inspect and revise. A new generation will still vary with model, seed, duration, source image, and the controls available in that model’s interface.
Use concrete visible action instead of backstory or emotion labels.
Give one shot one main camera move; split conflicting moves into separate shots.
Keep identity, wardrobe, environment, and direction consistent across cuts.
Do not request silence and loud ambience, or locked framing and a moving camera, at once.
Generate a short test before expanding a concept into a longer sequence.
Before you generate
It turns a visual idea or the evidence in a reference clip into written direction for an AI video model. A useful prompt covers subject, action, camera, composition, lighting, timing, style, and audio instead of relying on a short mood phrase.
Yes. Use the idea builder to structure an original concept. The upload and URL workspace is for the different job of reverse-engineering a reference video shot by shot.
The idea builder includes portable, Kling, Veo 3, Seedance, Sora, and Runway direction. Reference analysis returns tailored versions for those five models plus a generic format.
Use enough detail to control visible decisions, but remove words that do not change the shot. A single shot may need one concise paragraph; a multi-shot sequence needs clear beats and continuity instructions.
Include them when they matter. Name one motivated camera move with speed and direction. For models that generate sound, identify dialogue, effects, ambience, music, or intentional silence and connect important sounds to visible actions.
Treat it as an editable production brief. Check that it describes one coherent result, remove conflicting instructions, and adapt duration, aspect ratio, or model-specific controls in the generation interface you use.
The idea builder creates an editable prompt in your browser at no charge. Reference-video analysis also includes limited free usage; check the workspace notice for the current guest or signed-in allowance.
Use the idea builder and describe the image as the starting subject, setting, and composition. Then add only the motion, camera response, sound, and changes you want the video model to create.
