banana-scene-prompt

name	banana-scene-prompt
description	Use when converting scene analysis results into Banana JSON prompts for the current Banana/Veo pipeline. This stage should lead with one decisive narrative identity sentence, then give layered scene control rather than one giant brute-force room dump.

Banana Scene Prompt Generation

Overview

This skill is part of the current Banana + Veo mainline.

Its job is not to summarize a room loosely.

Its job is to turn the final reusable master shot scene analysis into one explicit, layered asset prompt.

This model understands semantics.

So do not assume that brute-force enumeration is always the best prompt form.

The best prompt should usually:

• open with one decisive sentence that tells the model what this space is and what dramatic job it serves
• explicitly state whether the scene asset must be empty or may contain only background crowd
• the exact master-shot camera view
• the upstream scene zone naming when available
• the key visible environment anchors
• the exact scene light / color / contrast
• a realistic, detail-rich, professional scene-building tone

Authority

Follow the active project spec first.

Current mainline reference:

• ${OPENCLAW_HOME:-~/.openclaw}/agents/flashxiaolagu制作/production-optimization-spec.md

Input

• analysis/scene-analysis.json
• analysis/film-type-analysis.json

This stage must inherit:

• selected_option.ratio_design
• selected_option.style_design

If scene-analysis.json already provides:

• control_table.style_inheritance
• control_table.master_shot_lighting_design
• control_table.scene_zone_map

use those fields as the scene-lighting and scene-layout authority instead of reconstructing them loosely.

Authority order:

1. scene-analysis.json -> control_table.style_inheritance
2. film-type-analysis.json -> selected_option

If these sources disagree on ratio_design or style_design, stop and flag the mismatch instead of choosing a convenient crop or rewriting the look.

Output

• prompts/banana-scene-prompts.md
• prompts/banana-scene-prompts.json

Required JSON Shape

{
  "ratio_design": "16:9",
  "style_design": "[窗光] + [青橙色调] + [高反差] + [电影数字质感]",
  "scenes": [
    {
      "scene_id": "scene_001",
      "scene_name": "场景名称",
      "prompt": {
        "aspect_ratio": "16:9",
        "scene_identity_sentence": "一句话说明这个空间是谁的空间、承担什么剧情作用",
        "human_presence_rule": "默认 no visible people；特殊场景可允许 background crowd only",
        "master_shot_view": "终极master shot视角描述",
        "foreground_midground_background": "前中后景的关键元素分层",
        "key_visual_elements": "关键环境锚点，不必把所有次级细节都写成长篇穷举",
        "spatial_type": "开放空间/封闭空间/过渡空间",
        "lighting": {
          "setup": "布光方法",
          "color_style": "色彩风格",
          "contrast": "反差影调"
        },
        "shot_language": "镜头类型、构图方式、机位关系",
        "realism_goal": "细节丰富、真实感强、层次明确等约定性目标",
        "technical_specs": "8K resolution, photorealistic PBR textures, environment concept art, single-point perspective, zero lens distortion",
        "style_reference": "match selected style design exactly",
        "negative_prompt": "文字、字幕、水印、logo、卡通、动漫、3D渲染、插画、油画、模糊、低画质、塑料质感、错误透视、空间错位、光影不一致、过曝、复杂畸变"
      }
    }
  ]
}

Writing Rules

1. scene_identity_sentence comes first

This is the highest-weight opening sentence.

It should tell the model:

• whose room or space this is
• what dramatic use the space serves in this scene
• what kind of human and narrative identity the environment should imply

Example direction:

• "A cramped bachelor bedroom where a half-awake young man is cornered by an unexpected mother-and-child intrusion."

1.5. human_presence_rule must be explicit

Default wording:

• empty environment plate, no visible people, no human figure, no character

Only switch to a crowd-allowed rule when crowd mass is truly part of the environmental identity.

Recommended crowd wording:

• background crowd only, no singled-out hero person, crowd acts as environmental mass

2. master_shot_view is the camera truth

Describe the final reusable master-shot camera view directly.

State:

• what this view covers
• where the bed is
• where the door is
• where the window / desk / wardrobe are
• what confrontation corridor or empty path remains visible
• what framing ratio this view is built for

3. foreground_midground_background should stay layered

Do not dump one flat paragraph if a layered field is clearer.

Split the visible set into front / middle / back layers when helpful.

4. key_visual_elements must stay concrete but selective

Mention the decisive visible anchors, not every low-value micro detail.

Prioritize:

• bed, mattress, bedding, pillows, blanket
• wall behind the bed
• bedside table and lamp
• door leaf, handle, frame
• window shape and curtains
• desk, chair, computer, books
• wardrobe, shelf, drying rack, cabinet
• floor, ceiling, ceiling light
• visible clutter and life traces

Let the model complete minor implied detail, but do not omit the anchors that protect continuity.

5. lighting must inherit the selected scene light exactly

Do not replace scene lighting with generic studio light.

It must match the selected StyleDesign.

If scene-analysis.json already contains master_shot_lighting_design, use that block as the first authority for:

• key light source
• fill behavior
• practical light policy
• color temperature balance
• contrast goal
• exposure behavior

Do not overwrite that block with generic language.

5.5. aspect_ratio is mandatory

Every scene prompt must explicitly write prompt.aspect_ratio.

Default rule:

• it must match control_table.style_inheritance.ratio_design when that field exists
• otherwise it must match selected_option.ratio_design
• do not silently switch to portrait because Banana is often vertical
• do not silently switch to landscape because a hotel scene looks cinematic

6. shot_language should feel like a director-level scene setup

Include only the most useful professional control:

• shot size
• camera height or side if important
• composition logic
• master-shot facing relation

If scene-analysis.json provides a scene_zone_map, keep those zone names stable in your internal reasoning so later stages do not drift.

7. realism_goal is encouraged

Use short phrases such as:

• rich lived-in detail
• strong realism
• layered visual depth
• believable apartment scale
• tactile materials

8. Do not overbuild dead weight

Do not stuff the prompt with low-value abstract explanation if the same control is already carried by:

• the opening sentence
• the master-shot description
• the layered visual anchors
• the lighting block

9. technical_specs stays fixed

Use:

• 8K resolution, photorealistic PBR textures, environment concept art, single-point perspective, zero lens distortion

Inheritance Checklist

Before writing the final JSON, verify:

• top-level ratio_design matches the selected type-analysis option
• top-level style_design matches the selected type-analysis option
• every scenes[*].prompt.aspect_ratio matches the inherited ratio
• style_reference matches the inherited style
• if scene-analysis provides style_inheritance, you did not drift from it

Constraints

1. Output must be valid JSON.
2. The prompt must begin with a decisive narrative identity sentence.
3. The prompt must explicitly control whether people are forbidden or only background crowd is allowed.
4. The prompt must describe the final master-shot scene view.
5. The prompt must preserve exact scene lighting from the selected style.
6. If scene analysis already provides master_shot_lighting_design, that block should be treated as lighting authority.
7. If scene analysis already provides scene_zone_map, later wording should not silently rename the zones.
8. Every scene prompt must explicitly carry inherited aspect_ratio.
9. The prompt must stay layered and concrete without collapsing into useless brute-force text.
10. Output both markdown and json.