1. Core narrative rules (narrative DNA)

3. Shot tech-selection matrix

4. Diffusion prompt protocol

This is what keeps visuals high quality and coherent. Every diffusion shot description should combine five parts:

$$ \text{Prompt} = \text{[Style anchor]} + \text{[Micro-timeline]} + \text{[Concrete entities]} + \text{[Camera physics]} + \text{[Physical bridge]} $$

A. Style anchors

• Force consistency: Start every shot with the same style phrase, e.g. 【Impressionist oil painting】 or 【Cyberpunk photoreal】.
• Push intensity: Use extreme wording—reject “fine.”
  • Weak: “sunflowers”
  • Strong: “Van Gogh sunflowers as extremely thick, rough impasto in blazing yellow”

B. Micro-timing

• Avoid even mush: State what happens each second.
  • Pattern: 【0–2s】action A, 【2–10s】action B.

C. Concrete entities

• Make everything physical: Turn abstractions into objects. Models don’t understand metaphor alone.
  • Weak: “falling into despair”
  • Strong: “the floor collapses underfoot into a bottomless pit of black tar”

D. Camera physics

• Lock direction: Say push in, pull back, pan.
• Keep inertia: If the last shot pushed in, this shot must continue pushing in—random moves cause visual whiplash.

E. Physical transitions

• Input dependency: Say explicitly: “this shot is generated from the last frame of the previous shot.”
• No pop in/out: Nothing vanishes without a process.
  • Weak: “the house disappears”
  • Strong: “the house crumbles from the roof into golden sand blown away by wind”

5. Execution workflow

1. Storyboard: Lock the story, split into $N$ shots.
2. Duration math:
   • Write lines $\rightarrow$ count characters $\rightarrow$ divide by speech rate ($4.5$) $\rightarrow$ round up to duration $T$.
   • Check: $T \le 10\text{s}$ for diffusion segments.
3. Continuity:
   • For each shot, define start frame and end frame sources.
   • Strategy A (Diff $\rightarrow$ Diff): previous end frame = next start frame (I2V).
   • Strategy B (Code $\rightarrow$ Diff): last code frame export = first diffusion frame.
4. Asset build:
   • Render all silent video segments.
   • Generate matching TTS and SRT.
   • Verify: $\sum(\text{segment durations}) = \text{total audio duration}$.
5. Final mux: Remotion combines video, audio, and subtitle layers into MP4.

6. Standard script output template

When writing a script, use this structure.

Video basics

• Theme: [e.g. a developer sorting a mountain of messy code]
• Estimated total length: $[xx]\ \text{s}$
• Resolution: $1920 \times 1080$ ($1080\text{p}$)
• Style keywords: [e.g. minimal, low-poly, cool palette]

Shot execution table

Document metadata