Directs AI-generated video by decomposing a brief into single-intention shots - one camera move plus one subject action per 5 to 8 second clip - then serializing the shot list to the target model's multi-shot mechanism with timestamp-exact audio cues and reference-locked identity, restricted to non-UI plates because AI video never carries legible product UI. Use when someone asks "write a prompt for an AI product video", "my generated video morphs and breaks physics halfway through", "how do I make a 30 second launch trailer with AI video", "how do I keep the same character across video shots", or when a single video prompt stacks several actions and camera moves. Do NOT use for trailer narrative structure - use launch-trailer-director and video-storyboard instead; for beats showing product UI, use marketing-ui-animation; for still imagery, use photoreal-prompt-craft.
---
name: ai-video-direction
description: Directs AI-generated video by decomposing a brief into single-intention shots - one camera move plus one subject action per 5 to 8 second clip - then serializing the shot list to the target model's multi-shot mechanism with timestamp-exact audio cues and reference-locked identity, restricted to non-UI plates because AI video never carries legible product UI. Use when someone asks "write a prompt for an AI product video", "my generated video morphs and breaks physics halfway through", "how do I make a 30 second launch trailer with AI video", "how do I keep the same character across video shots", or when a single video prompt stacks several actions and camera moves. Do NOT use for trailer narrative structure - use launch-trailer-director and video-storyboard instead; for beats showing product UI, use marketing-ui-animation; for still imagery, use photoreal-prompt-craft.
---
# ai-video-direction
Video models allocate temporal attention across the clip: stack two actions and a camera move into one prompt and the subject morphs, beats drop, and physics breaks - premium-tier credits burned on unusable footage. Real cinematography is also built from single-intention shots, so the constraint matches the model's failure mode and professional grammar at once. The second mistake this skill prevents is worse than the morph: no model renders pixel-accurate real interfaces, so any generated frame carrying legible product UI ships hallucinated interface text, and one impossible screen poisons every true claim in the film. One clip, one move, one action, no legible UI - the law of the pack.
## Operating procedure
Run the steps in order. The shot list is decomposed before any serialization because the multi-shot mechanism is chosen per shot count; identity locks attach before generation because they cannot be added retroactively; audio is cued last because cues must match final timestamps.
### Step 1: Elicit the brief
Gather exactly these inputs before decomposing anything. If any input is a guess, label it a guess and move on.
- **Charter**: the client's brand charter block. Paste the charter block if one exists. If the client has a charter but the foundations pack is not installed, elicit the needed values directly and label them provisional. If no charter exists at all, halt and route to premium-design-foundations - never substitute worked-example values.
- **Runtime and beats**: total runtime and beat structure. For a trailer, launch-trailer-director owns the beat grid (hook, then 3-5 vignettes, then an accelerating montage, then a 2-3s end card); if that pack is absent, elicit the grid directly.
- **UI/non-UI split**: every beat showing legible product UI routes to marketing-ui-animation (Remotion). This skill takes only the non-UI plates - device, environment, and atmosphere shots with screens dark, off-axis, or below legibility.
- **Identity subjects**: the subjects that must stay identical across shots, with reference images.
- **Set block**: the lane's registered set-block serialization from brand-consistency-prompting.
- **Audio**: dialogue or SFX requirements.
- **Ratios**: delivery aspect ratios.
- **Grade**: grading deliverable yes or no; yes routes to structured-control-pipeline for high-bit-depth export.