القائمة
视觉素材获取与生成。根据 resources.yaml 定义，通过库存图片/视频搜索（Pexels/Pixabay）或 AI 生图（Caro LLM）获取视觉素材。包含搜索策略、视觉隐喻搜索（counterpoint 声画关系）、构图感知搜索（framing 字段）、prompt 工程指南、素材质量评估标准。
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
استنادا إلى تصنيف SOC المهني
| name | carocut-media-visual |
| description | 视觉素材获取与生成。根据 resources.yaml 定义，通过库存图片/视频搜索（Pexels/Pixabay）或 AI 生图（Caro LLM）获取视觉素材。包含搜索策略、视觉隐喻搜索（counterpoint 声画关系）、构图感知搜索（framing 字段）、prompt 工程指南、素材质量评估标准。 |
Retrieves stock images and generates custom images for video production. This step fulfills visual resource requirements defined in manifests/resources.yaml.
digraph visual_decision {
"Resource from resources.yaml" [shape=doublecircle];
"source = existing?" [shape=diamond];
"Already in raws/" [shape=box];
"source = retrieve?" [shape=diamond];
"Search stock images" [shape=box];
"source = generate?" [shape=diamond];
"Generate with AI" [shape=box];
"Needs transparency?" [shape=diamond];
"Remove background" [shape=box];
"Save to raws/" [shape=box];
"Update resources.yaml" [shape=doublecircle];
"Resource from resources.yaml" -> "source = existing?";
"source = existing?" -> "Already in raws/" [label="yes"];
"source = existing?" -> "source = retrieve?" [label="no"];
"source = retrieve?" -> "Search stock images" [label="yes"];
"source = retrieve?" -> "source = generate?" [label="no"];
"source = generate?" -> "Generate with AI" [label="yes"];
"Search stock images" -> "Needs transparency?";
"Generate with AI" -> "Needs transparency?";
"Needs transparency?" -> "Remove background" [label="yes"];
"Needs transparency?" -> "Save to raws/" [label="no"];
"Remove background" -> "Save to raws/";
"Already in raws/" -> "Update resources.yaml";
"Save to raws/" -> "Update resources.yaml";
}
资源类型完整定义见 carocut-shared-schema skill。Media 只处理 visual 类型（image/video/sprite），不处理 data 和 components 类型。
| Source | When to Use |
|---|---|
existing | User-provided materials already in raws/ |
retrieve | Objective backgrounds, stock photos, stock videos (Pexels/Pixabay) |
generate | Custom characters, foreground elements with transparency |
When the storyboard includes animated characters, mascots, or any element with multiple poses/actions, prefer generating sprite sheets over individual images. Sprite sheets provide:
When the storyboard requires motion footage (product demos, b-roll, ambient shots), use type: video in resources.yaml:
- id: demo_recording
type: video
description: "产品演示录屏"
source: retrieve
search_keywords: ["product demo", "screen recording"]
duration_range: "5-15s"
videos_searchraws/videos/raws/
videos/
retrieved/
product_demo_001.mp4
aerial_city_002.mp4
- id: vid_001
type: video
description: "Aerial city timelapse"
source: retrieve
search_keywords: ["aerial city", "timelapse", "urban"]
duration_range: "5-10s"
path: videos/retrieved/aerial_city_001.mp4
Search royalty-free images from Pexels and Pixabay.
| Need | Keywords |
|---|---|
| Video backgrounds | abstract, texture, gradient, bokeh, blur |
| Business | office, meeting, professional, workspace |
| Nature | mountain, forest, ocean, sunset, aerial |
| Technology | technology, computer, coding, digital |
| Minimal | minimal, white background, flat lay, clean |
When the storyboard specifies audio_visual_relation: counterpoint, the visual should not literally illustrate the narration. Instead, switch to metaphorical/associative search:
| Narrative Concept | Metaphor Search Keywords |
|---|---|
| 信息过载 | overflowing water, paper avalanche, cluttered desk |
| 快速增长 | sprouting plant timelapse, rising tide, rocket launch |
| 技术落后 | vintage technology, old computer, floppy disk |
| 复杂度 | tangled wires, maze, puzzle pieces |
| 突破/创新 | breaking through wall, sunrise, butterfly emerging |
| 竞争 | race track, chess game, climbing mountain |
| 连接/协作 | bridge, handshake, puzzle fitting together |
| 衰退/消失 | wilting flower, melting ice, fading light |
Rules:
audio_visual_relation: counterpointsync or lead-visual, use literal/direct search as usuallead-audio, visuals can be more abstract/atmospheric but not necessarily metaphoricalAdjust search keywords based on the storyboard's framing field to find images matching the intended shot composition:
| Framing | Append to Search Keywords |
|---|---|
ECU (Extreme Close-Up) | "close up", "detail", "macro", "texture" |
CU (Close-Up) | "close up", "detail", "macro", "texture" |
MCU (Medium Close-Up) | (no modification, default search) |
MS (Medium Shot) | (no modification, default search) |
LS (Long Shot) | "wide shot", "landscape", "panoramic", "aerial" |
ELS (Extreme Long Shot) | "wide shot", "landscape", "panoramic", "aerial" |
Rules:
MCU and MS, no framing keywords are needed| Platform | License | Attribution | Commercial Use |
|---|---|---|---|
| Pexels | Pexels License | Appreciated but not required | Yes |
| Pixabay | CC0 / Pixabay License | Not required | Yes |
Generate custom images via OpenAI-compatible API.
sign that reads "Chinese text here"Background image:
Abstract gradient background, soft blue to purple transition, subtle geometric patterns, clean modern aesthetic, suitable for tech presentation
Character with transparency:
A friendly 3D cartoon scientist character, female, wearing white lab coat and glasses, standing pose facing camera, on a plain white background, Pixar style rendering
Sprite sheet:
Grid: 8 columns x 4 rows.
Row 1: Idle - character breathing subtly
Row 2: Walking - 8-frame walk cycle from left
Row 3: Running - 8-frame run cycle
Row 4: Jumping - jump up and land sequence
Character: cute robot with round blue body, small antenna, expressive digital eyes.
Style: 3D cartoon, soft lighting, clean design.
Remove background from generated images to create transparent subjects. Works best when the original image has a clean, solid-color background.
After generating a sprite sheet, validation is performed automatically by the images_generate tool when using mode: sprite. You can also run validation independently with the images_validate_sprite tool.
When using images_generate with mode: sprite:
cols/rows parametersmax_retries times, default 3):
Recommended usage:
images_generate(
prompt="Grid: 8 columns x 4 rows. ...",
output="/path/to/sprite.png",
mode="sprite",
cols=8, # explicit is more reliable than auto-inference
rows=4,
)
Use images_validate_sprite for post-hoc validation or re-checking:
images_validate_sprite(
image="/path/to/sprite.png",
cols=8,
rows=4,
fix_chroma=true, # optional: replace magenta with transparency
output="/path/to/sprite_transparent.png",
extract_frames="/path/to/frames/", # optional: extract individual frames
)
Returns JSON with check results:
{
"success": true,
"checks": [
{"check": "dimensions", "passed": true, "frame_size": "512x512"},
{"check": "aspect_ratio", "passed": true, "frame_aspect_ratio": 1.0},
{"check": "chroma_key", "passed": true, "magenta_coverage": 42.3}
]
}
If sprite sheet generation fails after all retries:
partial_failureIf the tools are unavailable, use Python/Pillow to verify:
from PIL import Image
img = Image.open("raws/images/generated/robot_sprite.png")
cols, rows = 8, 4
w, h = img.size
assert w % cols == 0, f"Width {w} not divisible by {cols} columns"
assert h % rows == 0, f"Height {h} not divisible by {rows} rows"
frame_w, frame_h = w // cols, h // rows
print(f"Sprite: {w}x{h}, Frame: {frame_w}x{frame_h}, Grid: {cols}x{rows}")
| Check | Expected | Auto-handled | Manual Action if Fails |
|---|---|---|---|
| Width divisible by columns | width % cols == 0 | Yes (auto-retry with enhanced prompt) | Regenerate with explicit resolution |
| Height divisible by rows | height % rows == 0 | Yes (auto-retry with enhanced prompt) | Regenerate with explicit resolution |
| Frame aspect ratio ~1:1 | ratio ≈ 1.0 ± 0.3 | Warning only | Adjust cols/rows ratio |
| Magenta coverage > 15% | Background is clean | Yes (auto-retry) | Regenerate with stronger magenta instruction |
| No magenta on character | Character pixels distinct | Not auto-checked | Add "avoid magenta/pink colors on character" to prompt |
| Frame-to-frame continuity | Smooth motion | Warning only | Regenerate with "smooth transition" emphasis |
After validation, record sprite metadata for downstream consumption:
- id: vis_sprite_001
type: sprite
description: "Robot character animation"
source: generate
path: images/generated/robot_sprite.png
sprite:
cols: 8
rows: 4
frameWidth: 512
frameHeight: 512
chromaKey: "#ff00ff"
animations:
idle: { row: 0, frameCount: 8 }
walk: { row: 1, frameCount: 8 }
run: { row: 2, frameCount: 8 }
jump: { row: 3, frameCount: 8 }
raws/
images/
existing/ # User-provided (unchanged)
logo.png
retrieved/ # From Pexels/Pixabay
server_room_001.jpg
abstract_bg_002.jpg
generated/ # From Caro LLM API
robot.png # With transparency
scientist.png # With transparency
background_tech.png # No transparency needed
robot_sprite.png # Sprite sheet (magenta bg)
videos/
retrieved/ # From Pexels Video API
product_demo_001.mp4
aerial_city_002.mp4
manifests/resources.yaml to identify visual resourcessource: retrieve:
raws/images/retrieved/source: generate:
images_generate with mode: sprite, cols, rows params. Built-in auto-validation and retry handles dimension/chroma issues.images_validate_sprite with fix_chroma=true to convert magenta to transparencyraws/images/generated/resources.yaml with actual file paths (include sprite metadata for sprite sheets)当 dispatch context 中 mode: "incremental" 时：
After retrieving/generating images, update the resource entry with the actual path:
Before:
- id: vis_001
type: image
description: "Server room with blue lighting"
source: retrieve
After:
- id: vis_001
type: image
description: "Server room with blue lighting"
source: retrieve
path: images/retrieved/server_room_001.jpg
视觉素材增强进行中...
处理中: vis_003 (server room background)
来源: Pexels
状态: 下载完成
视觉素材增强完成。
检索素材:
- 3 张图片 (Pexels)
- 2 张图片 (Pixabay)
生成素材:
- 2 张图片 (带透明背景)
- 1 张背景图
已更新: manifests/resources.yaml
输出目录: raws/images/
- retrieved/ (5 个文件)
- generated/ (3 个文件)
请确认素材后继续下一步。
| Issue | Cause | Solution |
|---|---|---|
| API key not found | Environment variable not set | Set PEXELS_API_KEY, PIXABAY_API_KEY, or CARO_LLM_API_KEY |
| Background removal messy | Original image has complex background | Use "on plain white background" in generation prompt |
| Search returns irrelevant results | Query too vague | Use specific English keywords, add style descriptors |
| Generated image style inconsistent | Prompt lacks detail | Specify art style, lighting, and composition clearly |
Remotion Shot 组件实现。将 storyboard 转化为 React 组件。包含关键的帧计算规则（Math.round 强制）、interpolate 安全规则（最小 1 帧）、1080p 视觉标准、动画编排模式、音视频同步架构、组件复用模式。这是技术密度最高的 skill。
Remotion 资产管道。调用 project_migrate tool 将 raws/ 素材迁移到 template-project/public/ 并生成 resourceMap.ts、constants.ts、timing.ts。Agent 只需 edit constants.ts 中的 COLORS 和 FONTS。
音频素材生成与获取。批量 Edge TTS 旁白生成（支持 storyboard pacing 字段驱动语速）、BGM/SFX 检索（BGM 节奏匹配 BPM 规则）、音频时长提取。包含 Edge voice 配置、速度调整规则、durations.json 格式规范（含 audio_visual_relation 说明）和关键的音频时序规则。
旁白脚本口语化处理与情绪节奏标注。
视频制作策划。四阶段交互式策划流程：Creative Memo → Resources → Script → Storyboard。包含完整的影视标准（剪辑七律、景别、机位、灯光）、盖曼讲故事法则（叙事引擎、冰山理论、叙事声音、灵感堆肥、人物推进、幽默感、对白法则）、视听语言系统（情绪曲线、景别/运镜/节奏/张力/声画关系/转场）、文档模板和审批流程。
所有 agent 共享的协议定义。包含资源类型注册表、storyboard schema、dispatch context schema、amendment action 类型。Planner 写入时遵守、Media 处理时校验、Builder 实现时参考、Orchestrator 调度时构造。