# AI-Driven Development — Motion-Plate Prompt Guide

Use generated video as a quiet moving plate beneath the existing edit. Dialogue,
captions, reports, graphs, UI and music remain deterministic layers in HTML and
the final compositor. The model supplies physical life, not the storytelling.

## The shot contract

Write prompts in this order:

1. **Source contract** — say this is the exact source frame and name everything
   whose identity, wardrobe, geometry or placement must remain unchanged.
2. **Shot purpose** — one sentence describing what the audience should feel or
   notice.
3. **Subject motion** — allow one small, physically plausible action.
4. **Camera motion** — choose one restrained move, or explicitly lock the camera.
5. **Environment motion** — add only low-amplitude movement such as reflections,
   rain, display light or breathing.
6. **Look** — name lighting, contrast, lens character and mood.
7. **Exclusions** — prohibit speech, lip movement, identity drift, new objects,
   text generation, shake, rapid moves, cuts and transitions.
8. **Exit frame** — request a one-second steady hold for an editable cut.

## Default series prompt

> Animate this exact source frame as a restrained premium technology-thriller
> insert. Preserve the subject's identity, age, facial structure, hair, wardrobe,
> environment, framing and colour palette. The subject performs one subtle natural
> action: [ACTION]. The camera makes one extremely slow stable [CAMERA MOVE]. Only
> faint monitor reflections and natural breathing move in the environment. Deep
> blacks, controlled practical lighting, subtle film texture, realistic physics.
> Silent shot: no speaking and no lip movement. No cuts, camera shake, fast zoom,
> face morphing, wardrobe change, new objects, subtitles, logos or generated text.
> End on a steady one-second hold.

## Evidence and document plates

AI video models are not trusted to reproduce legible copy. Use the generated clip
for the desk, page, hand, light and camera movement; composite the real report and
highlighted words over it afterward.

> Animate this exact source frame as a forensic evidence insert. Preserve the A4
> page position, rectangular geometry, margins and surrounding dark bank
> environment. Do not invent or rewrite text. The page remains perfectly flat and
> still. An extremely slow, stable camera push-in creates subtle parallax while a
> faint monitor reflection passes across the desk. No hands enter frame. No cuts,
> shake, rotation, page warping, new graphics, subtitles or logos. End on a steady
> one-second hold with clean compositing space over the page.

## Sam continuity plate

> Animate this exact source frame as a quiet premium streaming-drama shot. Preserve
> Sam's identity, age, facial proportions, long dark hair, black blazer, charcoal
> top, desk, laptop, papers and glass-walled bank environment. Sam breathes
> naturally, makes one small eye movement towards the report and one subtle hand
> adjustment. The camera performs an extremely slow stable push-in. Faint monitor
> reflections move in the background; everything else remains still. Deep blacks,
> restrained practical lighting, natural skin texture, subtle film grain. Silent:
> no speaking or lip movement. No cuts, shake, face morphing, wardrobe change,
> additional people, text, logos or dramatic gestures. End on a steady one-second
> hold.

## Seedance input choice

- Use `image` when the still must be the exact first frame and composition.
- Use `reference_images` when identity, wardrobe or visual style matter more than
  reproducing the starting composition. Name them as `[Image1]`, `[Image2]`, etc.
- Do not combine first/last-frame images with reference images in Seedance 2.0.
- Keep generated audio disabled for this series. Existing narration, score and
  effects provide continuity and avoid fighting the edit.
- Prototype at five seconds and 720p. Increase duration or resolution only after
  the motion language and continuity pass review.

Reference-image example:

> Sam from [Image1], with exactly the same face, long dark hair, black blazer and
> charcoal top, sits in the secure bank investigation room shown in [Image2]. Quiet
> late-night concentration. She glances once from the report to the monitor. Locked
> medium camera with only a nearly imperceptible push-in. Preserve the realistic
> visual style, practical lighting and deep black palette of [Image2]. Silent, no
> speech, no lip movement, no new people, no cuts, no identity or wardrobe drift.
> End on a steady one-second hold.

## Time-coded shots

Use one generated shot wherever possible. If a longer Seedance shot genuinely needs
beats, state the camera, subject action and light state for every interval:

> [0–2s] Locked medium shot; Sam studies the report, breathing naturally.
> [2–4s] Extremely slow push-in; her eyes move once towards the monitor.
> [4–5s] All motion settles into a steady hold for the cut.

Do not ask the model to create an edit that the compositor can perform exactly.
Highlights, extracted words, charts, interface messages and cuts remain code-native.

## Dialogue coverage

For precise dialogue, generate coverage as separate silent plates and cut them in
the compositor:

1. establish both speakers with a two-shot;
2. cut to the current speaker just before their line begins;
3. hold their face through the sentence, allowing only restrained eye, breath and
   head movement;
4. cut to the listener for the reaction or pause;
5. return to a wider shot when the conversation changes direction.

This produces an exact `speaker A → speaker B` edit and lets the existing voice
track determine timing. Do not generate new dialogue audio. Lip movement is
optional: for most current scenes, restrained non-speaking reaction coverage is
more reliable and still feels cinematic under voiceover.

Seedance 2.0 accepts time-coded multi-shot prompts up to 15 seconds, and Kling 3.0
can accept up to six shot definitions. Use those features for exploratory montages
or action contained inside one generation. Do not depend on them for a critical
line-by-line edit when separate plates can be cut precisely and revised cheaply.

## Review test

A clip passes only if it answers yes to all four questions:

- Does it feel cinematic at full-screen size?
- Does the source identity and set remain recognisably unchanged?
- Is the motion stable enough to sit under narration without distraction?
- Is it actually useful in the episode after overlays are restored?

The comparison page records each answer from 1–5 and saves the review locally.