Write camera-motion prompts for short AI video

Recover a short shot by simplifying the movement

Start with one shot, one camera move, one subject action, and at most one environmental action. Google’s current video guidance recommends dedicating a short-video prompt to a single focused moment instead of chaining several distinct events. For image-to-video, it also recommends describing motion rather than repeating the subject, scene, and style already present in the source frame. A practical first draft is: “Slow dolly in toward the ceramic cup. Steam curls upward. The camera remains level in one continuous shot.”

If the result drifts, recover in a fixed order. First remove any second camera instruction. Then replace vague language such as “dynamic cinematic movement” with a named direction and pace. Next separate camera, subject, and environment into short sentences. Finally regenerate while changing only that motion clause. Runway describes iteration as reviewing a generation and reinforcing a missing component in clearer natural language; this procedure makes the changed variable visible instead of rewriting the entire prompt after every miss.

Know what the motion prompt sends to a cloud model

A camera-motion prompt is generation input, not private local editing metadata. OfflineCreator’s current MCP and CLI page says those interfaces use the cloud Studio service and that generation inputs go to the disclosed cloud provider selected for the job. Do not put client names, unreleased product claims, credentials, private filenames, or hidden campaign facts into a prompt merely to make the shot easier to recognize. Use neutral subject labels and keep the production brief outside the generation text.

This boundary also affects review notes. Record the model, selected duration, aspect ratio, exact prompt, and observable defects, but avoid copying confidential context into a shared comparison sheet. The claim here is limited to the documented cloud path. It does not assert a retention period, training policy, or provider privacy guarantee, and it does not describe a generation performed during this research.

Write the prompt as camera, subject, environment, and continuity

Use a four-part motion line: “[camera move and pace]. [subject action]. [environmental action]. [continuity constraint].” For example: “Slow truck right, parallel to the cyclist. The cyclist maintains a steady pace. Rain streaks diagonally through the streetlights. One continuous shot.” Google distinguishes a pan, where the camera rotates horizontally from a fixed position, from a truck, where the camera physically moves sideways. It likewise distinguishes a dolly, where the camera moves closer or farther away, from a zoom, where focal length changes while the camera does not move.

Those distinctions matter because “pan right beside the cyclist” mixes a fixed-position rotation with a parallel traveling move. Choose the physical behavior you mean. Use “slow pan right across the shelves” for a stationary reveal, “truck right beside the cyclist” for parallel travel, “dolly in toward the label” for a changing camera-to-subject distance, or “slow zoom in on the label” for optical magnification. Google’s image-to-video guidance separately identifies camera motion, subject animation, and environmental animation, with examples such as a slow dolly, fluttering clothing, and rolling fog.

Compare matched prompts before judging a motion term

Keep the scene and all non-camera words identical when comparing two moves. Matched prompt A: “Medium shot of a ceramic cup on a café table. Slow dolly in toward the cup. Steam curls upward. One continuous shot.” Matched prompt B changes only the camera clause: “Medium shot of a ceramic cup on a café table. Slow zoom in on the cup. Steam curls upward. One continuous shot.” This pair tests physical approach against optical magnification without also changing weather, subject action, style, or framing.

Review both outputs against the same contact sheet: opening frame, midpoint, closing frame, requested camera path, subject scale, geometry stability, unintended cuts, unwanted camera shake, and whether the steam motion competes with the move. Runway’s camera reference pairs named terms and prompts with Gen-4.5 outputs, so it is a useful external visual reference rather than proof that another model will interpret the same term identically. No authorized OfflineCreator generation was run for this page, and the two prompts above are a reproducible test design, not a claim about actual output quality.

Fit the shot to the selected model duration

Choose duration before adding beats. Google’s current Vertex AI documentation accepts 4, 6, or 8 seconds for Veo 3 models and identifies 8 seconds as the default. Runway documents 5- and 10-second Gen-4 clips and recommends treating each generation as a single scene. These are model-specific examples, not a universal duration rule and not evidence that every interface exposes every upstream option. Confirm the choices in the exact model and interface you will use.

For a five-second test, budget roughly one readable move: establish the composition, execute the dolly or truck, and let the final framing settle. For an eight-second test, a restrained reveal can have a clearer beginning and ending, but extra time is not permission to add a cut, a second location, and three camera paths. Submit the baseline prompt, wait for the complete file, inspect it frame by frame, and record pass or fail before changing one clause. If the move finishes too early, adjust pace; if it never resolves, simplify the action or use a supported longer duration.

Take the next step from the unresolved part of the shot

Return to the creative-workflow directory when the model, transport, or deliverable is still undecided. Use the disclosure guide when the clip will appear as concept media and needs accurate labeling. Use the image-prompt iteration guide only when the unresolved task is creating or revising a source still; camera-motion language for the resulting clip remains owned here. These links are chosen by the next production decision rather than by repeating the camera-motion query.

Editorial ownership and evidence boundary

This route owns the narrow task of writing and reviewing camera-motion prompts for one short shot. It provides named motion distinctions, two matched prompt variants, current examples of model-specific durations, and an output-review matrix. The matched prompts were not generated during research, so they must not be presented as a model comparison or a curated OfflineCreator output result. Runway’s external prompt-and-output reference is cited only as a first-party example library for its own model.

The required recent-evidence run returned current camera-prompt material, but it was shallow, promotional, off-task, or unverifiable; the detailed camera guides were outside the 30-day window, and some requested source coverage remained unavailable. Accordingly, this draft makes no claim about current practitioner consensus, prompt success rates, best-performing camera terms, or customer outcomes. If editorial policy requires completed matched OfflineCreator outputs rather than a matched test design and external first-party examples, the overlap guard requires consolidation into the creative-workflow directory instead of publication. Research does not activate that consolidation or this route.