Plan a one-shot AI video before generation

Write the shot contract before writing the prompt

A one-shot AI video plan is a decision document for one continuous clip, not a compressed storyboard. Start with the job the shot must perform: reveal a product, establish a location, demonstrate one action, or create a transition. Then specify the opening frame, the subject's single readable action, one camera behavior, the environment, and the final frame. If the concept needs a cut, a second location, or another beat, split it into another shot instead of hiding an edit inside one prompt.

Google's current video-generation prompt guide presents subject, action, camera movement, shot composition, lighting, style, and ambiance as prompt elements and keywords that can guide a generated video. Use those categories as a completeness check, not as a requirement to fill the prompt with adjectives. The guide distinguishes physical camera movement such as a dolly from lens movement such as a zoom. That distinction is useful at planning time: choose the visual effect you need, name one movement precisely, and remove any direction that competes with it.

Purpose

One sentence describing what the viewer should understandA visual objective is easier to review than a mood-only request.

Opening frame

Subject, framing, setting, light, and initial stateDescribe what must already be visible before motion begins.

Motion

One subject action plus one camera behaviorName direction, speed, and endpoint without inserting an implied cut.

Exit frame

A stable final composition with a clear review targetPlan a usable end state rather than letting the shot stop mid-action.

Set acceptance and rejection gates before spending

Convert the creative intention into observable checks. An acceptable clip should communicate the declared purpose without relying on a caption, keep the subject recognizable through the action, preserve essential geometry, complete the selected camera move, and settle into the intended exit frame. Add task-specific checks such as clear negative space for later copy or an uninterrupted view of a product feature. These are review criteria, not claims that the model will satisfy them.

Write hard rejects separately: an extra limb or object, changing product geometry, unreadable or invented text, a camera move that reverses direction, a scene cut, a subject that exits unintentionally, or a final frame that cannot be edited. Review the entire output frame by frame. A polished opening still does not rescue drift later in the clip, and a visually attractive result is still a rejection when it misses the shot's job.

Purpose test

Pass only if the intended idea reads from the clip itselfDo not use imagined voiceover or future captions to excuse an unclear shot.

Continuity test

Inspect identity, count, geometry, lighting, and backgroundRecord the first frame where a required element visibly changes.

Motion test

Confirm the action and camera move have one direction and endpointReject accidental cuts, reversals, or unfinished movement.

Editability test

Confirm the opening and exit frames can serve the planned editA usable shot needs a practical insertion point, not only an interesting middle.

Fit the plan to a current model path

Check the current model list and credit balance immediately before generation. OfflineCreator Studio's current public catalog lists Kling 2.6 Pro as a 42-credit text-to-video model for social video and Veo 3.1 Fast as a 96-credit text-to-video model for premium campaigns. Those labels and costs are a current catalog snapshot, not a quality ranking or a guarantee that either model fits every shot.

The disclosed fal endpoint for Kling 2.6 Pro accepts `16:9`, `9:16`, or `1:1` and exposes five- or ten-second output. The Veo 3.1 Fast endpoint accepts `16:9` or `9:16` and its current input schema exposes four-, six-, or eight-second choices. Plan inside the selected path's documented envelope. Provider options do not automatically become OfflineCreator MCP arguments, and longer duration is not evidence that a shot will be more coherent.

Example: plan a five-second tabletop reveal

Purpose: reveal the material and silhouette of an unbranded ceramic speaker in one restrained landscape product shot. Opening frame: `16:9` medium close-up, speaker centered on a dark walnut table, warm practical light behind it, soft blue dawn through a side window, no labels or lettering. Action: a thin status light turns on once while a faint vibration moves the nearby dust. Camera: one slow dolly in, stopping before the speaker fills more than half the frame. Exit: hold a stable three-quarter hero view for the final second.

Prompt draft: “Single continuous five-second shot, 16:9 medium close-up of an unbranded matte ceramic speaker centered on a dark walnut table at blue dawn, warm practical lamp in the background, a thin amber status light turns on once and nearby dust moves subtly, camera makes one slow straight dolly in and stops at a stable three-quarter hero composition, consistent speaker shape and surface, no cuts, no lettering, no additional objects entering frame.” The duration and ratio fit the cited Kling endpoint, but this is an editorial specification, not a tested output or performance claim.

Acceptance: the speaker's silhouette and openings remain stable, the status light changes only once, the dolly travels forward without panning or orbiting, no text appears, and the last second provides a steady frame. Reject if the object changes shape, the camera reverses, the lamp moves, extra controls appear, or the shot ends while the camera is still traveling. Preserve the prompt and rejection reason with the output so the next revision changes one cause rather than rewriting the whole concept.

Turn every rejection into one bounded revision

When a generated clip fails, compare it with the written shot contract before rewriting the prompt. Name the earliest observable break and map it to one planning field: opening frame, subject action, camera behavior, environment, timing, or exit frame. If the speaker changes shape before the light turns on, revise the subject-continuity direction. If the camera circles instead of moving straight forward, revise the camera line. Keep all other accepted instructions unchanged for the next attempt.

Store a compact revision note beside the generation record: planned behavior, observed behavior, first failing moment, rejection gate, and the single change proposed for the next run. Do not label an untested revision as a fix. If several requirements fail independently, rank them and test the most fundamental one first; changing subject, motion, lighting, duration, and composition together would make the next output difficult to diagnose. Stop the loop when the shot's purpose no longer fits one continuous clip and move the extra beat into a separate shot plan.

Choose the next action from the shot plan

Use the vertical-video guide when this shot must become a silent-first `9:16` concept with an editor handoff. Use the B-roll guide when the purpose is an insert for a longer YouTube edit. Return to the workflow directory when the plan exposes a different generation task, such as animating an approved still rather than generating from text.

Keep the guide inside its evidence boundary

This page owns one-shot planning before generation: purpose, opening frame, action, camera behavior, duration choice, exit frame, and review gates. It does not own multi-shot storyboards, editing, model benchmarking, or a promise that a provider will preserve identity, product geometry, typography, or exact motion. Recheck the live model list and tool schema before spending because this page has a monthly freshness requirement.

The page remains an unreviewed research draft. Current primary documentation supports the published input constraints and planning vocabulary, while recent community evidence is incomplete and anecdotal. A review-ready revision would need relevant full-source coverage and real shot-level artifacts: the exact plan and prompt, model response, unedited output, frame review, and documented pass or rejection decision.