AI video generation: plan shots before spending credits

Video needs preproduction

Video generation adds motion and continuity checks to the review work required for a still image. A good workflow starts with shots, not with one giant prompt. Google's video generation prompt guide recommends breaking ideas into components such as subject, action, scene and style. Google's video generation overview describes text-to-video, image-to-video and editing-style capabilities for Gemini Omni Flash and Veo on its platform. The practical takeaway is to plan the sequence before generating.

Write a shot list

For a 20-second product explainer, do not ask for a finished commercial. Write five shots:

1. Establishing shot: desk, laptop, morning light, calm workspace.
2. Close shot: cursor selecting a secure login option.
3. Cutaway: abstract lock animation, no real brand marks.
4. User reaction: person nods, natural expression, no spoken dialogue.
5. End frame: clean space for title added later in editing software.

Each shot gets its own prompt, duration target, aspect ratio, motion instruction and review note. This makes failures cheaper. If shot three fails, regenerate shot three, not the whole story.

Prompt anatomy

A strong video prompt includes subject, action, environment, camera movement, lighting, style, duration and constraints. Example: Close-up of a laptop keyboard as a user selects a passkey sign-in option on a simple fictional interface, slow push-in camera, soft daylight, realistic office, no readable brand logos, no extra fingers, 5 seconds. If exact UI matters, generate from a reference frame or composite real UI later. Video models may produce plausible but inaccurate details, especially when asked for small text or precise interface states.

Use review gates

Review each shot for four things: continuity, artifacts, rights risk and editability. Continuity asks whether lighting, subject and style match the surrounding shots. Artifacts include warped hands, flickering objects, strange physics or unreadable text. Rights risk covers accidental brands, celebrity likenesses or recognizable copyrighted elements. Editability asks whether the shot has enough lead-in and tail frames to cut cleanly.

Cost controls

Before generating, set a credit budget per concept. Decide how many attempts each shot gets and what causes a stop. For example: two generations per shot, one edit attempt for the selected shot, then rewrite the shot description if both fail. Keep a prompt log with thumbnail, settings, output ID and reason for acceptance or rejection. This turns a creative session into a learning loop instead of a pile of abandoned clips.

Problems to check

• One-prompt movie: the prompt asks for too many scene changes, and the model blends them.
• Unplanned text: generated signs or UI copy look wrong and require manual repair.
• Continuity breaks: the same person, object or room changes between shots.
• No edit margin: clips start and end too abruptly for a clean cut.
• Budget drift: retries continue because there is no stop rule.

Next steps

Write a five-shot script for one simple idea. Generate only the first shot. If it fails, improve the shot prompt before moving on. If it succeeds, use it as a style reference for the next shot. Add music, titles, legal disclaimers and exact UI in your editing tool, not by hoping the model gets every frame perfect. Keep the accepted shot list and rejection notes with the edit so another person can reproduce the decisions.

Post-production plan

Decide what belongs in video generation and what belongs in editing software. Use generation for motion, atmosphere, rough staging and stylized shots. Use editing tools for captions, brand typography, exact UI, legal text, audio mixing and final timing. This division saves credits because you stop asking the model to solve tasks that deterministic tools handle better.

Storyboard cheaply

Before generating video, sketch the sequence as still frames or text cards. A simple storyboard reveals whether the story actually fits the duration. It also gives reviewers something cheap to approve. If stakeholders disagree at the storyboard stage, video generation would only make the disagreement more expensive.

Audio last

Add music, narration and sound effects after the edit has stable timing. Regenerating visuals after audio is locked creates rework. Temporary audio is fine for pacing, but final audio should follow the approved cut.

Inspect motion frame by frame where it matters

Play each candidate at normal speed first. Then inspect the part of the shot that carries the message: the hand approaching the laptop, the object moving into frame, or the fictional screen changing state. A clean first frame does not establish that the rest of the clip remains usable. Check the transition into the next shot as well. For the product explainer, reject a clip that changes the laptop's shape halfway through even if its lighting fits the brief. Keep a short rejection reason with each candidate so the next prompt addresses the observed problem rather than changing the whole scene.

Sources and further reading

Documentation checked October 4, 2026. Product features and availability can change.

• Google video generation prompt guide
• Google video generation overview