Video Look Sound

---
name: video-look-sound
description: Lock the visual world and complete audio direction before narration or footage generation. Use when deciding locations, backgrounds, wardrobe, lighting, palette, narrator or dialogue voice, music, ambience, sound effects, captions, and whether speech should be separate TTS or native model audio.
---

# Direct the look and sound

Create one production bible that later narration, shot planning, generation,
assembly, and QA stages can execute without reinterpreting taste or audio intent.
Do this after the script and recurring-character references are settled, but
before generating narration or visual clips.

## Lock the visual world

Record decisions that must remain stable across shots:

- locations and backgrounds per beat, including time of day and practical set anchors;
- recurring wardrobe, hair, props, product markings, and other identity anchors;
- lighting direction, contrast, color temperature, palette, texture, and realism level;
- camera grammar: framing range, lens feel, movement, depth of field, and transitions;
- aspect ratio, composition safe areas, and caption-safe regions.

Separate invariants from intentional changes. A character may move to a new
location, but their approved identity and wardrobe do not drift unless the
script explicitly calls for that change.

## Lock the sound world

Record all audible and readable layers:

- speech mode per beat: off-camera voiceover, visible native dialogue, or explicitly no speech;
- voice provider/model or recorded source, voice ID, language, accent,
  warmth, energy, pacing, pronunciation notes, and performance references;
- music role, genre, tempo range, emotional arc, instrumentation, and where it yields;
- ambience and room tone per location;
- motivated sound effects and transition sounds;
- caption wording source, style, placement, line length, and timing policy.

Do not leave a generic instruction such as "warm voice" or "cinematic music."
Write enough specific direction that a replacement segment still sounds like
the same production.

## Choose speech deliberately

When a piece contains people or characters and spoken content, do not infer
whether they speak. Before selecting a video model, show the user three bounded
speech-design routes:

1. **Visible native dialogue** — lip-synced performance generated with the
   video; strongest for character-led scenes, but restricts the model choice
   and may vary voice/performance between independent clips.
2. **Off-camera TTS voiceover** — one consistent narrator over non-speaking
   performances; easiest to time and revise, but the characters do not talk.
3. **Hybrid** — native dialogue for key character moments and voiceover for
   B-roll or transitions; expressive and flexible, but more complex to plan,
   cost, and mix.

Explain the viable model family, performance and lip-sync result, voice
continuity, cost, and edit complexity for each. Recommend one and obtain the
user's explicit choice. An endpoint without native audio is a constraint to
disclose, not permission to silently convert dialogue into narration.

After that decision, classify every spoken beat. When a character or
presenter visibly speaks, prefer a video endpoint that creates the picture,
performance, dialogue, and synchronized audio together. Generating a separate
voice and trying to manufacture lip sync afterward is less reliable. Keep the
exact dialogue in the script and verify both wording and lip sync.

Use separate TTS only for off-camera voiceover. For instructional, tutorial,
and explainer voiceover, exact wording, a consistent voice, replaceable
corrections, and reliable timing matter more than native speech. Generate that
voiceover before visuals and use its measured duration as the edit clock.

Native model audio also remains useful for ambience, Foley, and location sound.
It does not make required off-camera voiceover optional. Never turn a spoken
instructional brief into a silent video because a generated visual clip
arrived without an audio stream.

## Make the direction visible in documents

The direction document must use explicit sections for:

- `Locations and backgrounds`;
- `Wardrobe, props, and continuity`;
- `Lighting and visual palette`;
- `Camera grammar`;
- `Speech and voices`;
- `Music`;
- `Ambience and sound effects`;
- `Captions`.

Carry those decisions forward instead of hiding them in prompts. The shot list
records the relevant background, look anchors, speech mode, and audio intent per
shot. The generation report records whether each clip contains native dialogue,
ambience/effects, or no usable audio. The render report lists the actual voice,
music, ambience, effects, and caption sources used in the final export.

Do not ask again about a choice already settled in the conversation.

## Build around real narration timing

When separate narration is selected:

1. Lock the exact script.
2. Generate narration in beat-level segments, not one monolithic file.
3. Measure every segment with `ffprobe`; never substitute word-count estimates.
4. Derive the shot list and visual duration from those measured values.
5. Generate and trim visual clips to that audio timeline.
6. Assemble narration, music, ambience, effects, and captions with the visuals.

If speech is explicitly not applicable, record the approved timing source—such
as a music edit, action beats, or native dialogue—so the shot list still has a
concrete clock.

## Keep previews honest

Show every generated visual clip before assembly. If a separate narration and
mix will be added later, label it clearly as a silent visual preview. Never
present that preview as the finished video or imply that missing audio will fix
itself.

## Enforce the final promise

At assembly and QA, compare the actual export with this direction document.
An instructional video that promised narration fails when narration is missing,
silent, unintelligible, materially different from the approved script, or
misaligned with the visuals. It may not mark audio or captions
`not_applicable`. Also verify voice consistency, pronunciation, music ducking,
ambience continuity, motivated effects, caption accuracy and timing, and the
visual invariants above.