yyl-video-prompt: Director-Style AI Video Prompts

Positioning

Step 0: Select a Tier

1. If the user specifies a platform or style, follow the user. Douyin, Xiaohongshu, fast pacing, and short videos use the short-video tier; premium series, cinematic narrative, slow pacing, and performance use the cinematic premium tier.
2. If the user gives no guidance, inspect the script: multiple dialogue lines, emotional turns, and two-hander scenes within the same setting indicate the cinematic premium tier; everything else uses the short-video tier.
3. Vertical-format short dramas → use the short-video tier while incorporating cinematic-tier dialogue embedding and reaction chains, as described at the end of the short-video guide.
4. If the tier cannot be determined, use the short-video tier and state in the Step 3 confirmation message that it can be switched to the cinematic premium tier.
5. Composite or editing tasks (placing A into B, changing only the background) do not use either tier. Follow the minimal format in the final section of tool-adapters.md.

modes/

Workflow

Step 1: Read the Input and Determine Three Things

1. Target tool and model: This determines the maximum duration of each segment and the syntax for referencing assets, as described in tool-adapters.md. If the user does not specify one, use the current project default. If the project has no default either, use

   Seedance 2.5 | 720p | 16:9

   .
2. Visual tone: Cinematic photography / smartphone live action / animation. Do not write "cinematic," "IMAX," "blockbuster," or "4K" for smartphone live action.
3. Reference assets: If assets are available, assign each image a single task. If there are none, clearly state what images should be prepared or switch to text-only generation.

Step 2: Divide the Content into Segments

modes/

Step 3: Director's Judgment — The Only Confirmation Gate

Tier: Short video (one continuous shot)——[one-sentence reason]; it can be switched to the cinematic premium tier for slow-paced acting.
I will divide it into N segments:
Segment 1 (0–15s): [one-sentence beat] | [generation mode] | [required reference images]
Segment 2 (…): …

My judgment:
- Segment 1 [specific moment]: I plan to [treatment], so viewers will feel [effect]; the alternative is [alternative].

To confirm, say "write it directly"; to merge, split, or change direction, just say so.

Step 4: Write Each Segment

Step 5: Self-Check and Deliver

python3 scripts/prompt_builder.py --validate <file>

Shell for Each Segment — Shared by Both Tiers

## Segment 1: [one-sentence beat] (0–15 seconds)

Tier: Short video (one continuous shot / multi-shot) or Cinematic premium
Generation mode: [text-to-video / first-frame image-to-video / first-and-last-frame / multi-image reference / universal reference / video editing]
Model parameters: [Model] | [Resolution] | [Aspect Ratio] | [Duration]
Reference assets:
Image 1: [single task]——locks [face / hairstyle / clothing / proportions…]; excludes [background / people within it…]
(For text-to-video, write "None"; if an image is missing, specify what image should be prepared)

Prompt:
```text
[Write using the selected tier's structure]
```

End-of-segment state: [pose, position, composition——the next segment continues from here]

@图片1

参考图1

{{Node "node name"}}

@图像1

Core Rules — Shared by Both Tiers

1. Every segment must be self-contained. Generation tools do not have cross-segment memory. Character locks, setting, lighting, props in hand, and the previous segment's ending state must be fully restated in every segment. Do not write "same as above" or "continue from the previous segment." Lock appearance and clothing through reference images rather than repeating them in the text.
2. Each reference image must serve only one task. Writing only "reference image" is equivalent to writing nothing. Always specify what is locked and what is excluded. Follow the target tool's reference syntax: LibTV requires

   {{Node "node name"}}

   ; plain-text references such as "Image 1" are invalid.
3. Timecodes are action budgets. Time windows must be continuous and non-overlapping, with only one core action per window. Write key events down to the exact second. Leave a 0.5-second buffer after the final beat because models often drift late at the end.
4. The first frame shows the state before the first action begins. If the character raises a leg at second 2, the first frame must show both legs together.
5. Characters must remain physically supported. Specify the support for every weight-bearing body part: what they lean against, where they sit, or where their feet are planted. Give every action a cause and an ending state.
6. Express emotion through visible actions, not isolated emotion labels such as "she is very sad." Use one expression beat for short video and a complete progression for cinematic premium content.
7. Embed dialogue in the timecodes. Do not append the lines at the end.
8. Write only the most important information. Excessive description may cause the model to promote elements that should remain within the scene to the entire frame.
9. Match style terms to the visual tone. Do not write "cinematic," "high definition," "blockbuster," or "4K" for smartphone live action. For cinematic photography, specify the camera, lens, and film stock. Use no more than one or two style references.
10. No background music by default. Retain only production sound. Add music only when the user requests it.

Shared Self-Check

• The correct tier was selected, and its

  modes/

  file and examples were read
• No segment exceeds the target tool's duration limit
• Every segment is self-contained, with no "same as above" or "continue from the previous segment"
• Every reference image serves only one task and includes both locked and excluded elements; every numbered reference in the prompt appears in the reference list; the reference syntax matches the target tool
• Actions have physical support and clear cause and effect; the first frame shows the state before the action begins
• Sound includes an ambient bed and event sounds, with no BGM unless requested
• Hard constraints use positive statements to lock subject occupancy, spatial logic, and action order; negative constraints target only known failure modes
• The ending state is written for every segment
• The tier-specific self-check has been completed

File Map

scripts/prompt_builder.py

--validate

--analyze-scene

--build