video-creation-collaborator

Yingpin Zhichuang - Multi-Agent Collaborative Video Creation Management

Task Objectives

This Skill is used to: standardize and guide the collaborative creation workflow of 10 agents, and solve issues such as distorted generated images (multiple limbs/deformities) and video compositing flaws.
Capabilities include:
- Structured division of labor among 11 agents (copywriting creation/story planning/script creation/storyboard director/storyboard artist/subtitle artist/voiceover artist/sound effects artist/video engineer/quality inspection/data feedback)
- 5-stage collaborative process management (requirement intake → content creation → image generation creation → audio subtitles → video compositing → full-process quality inspection → data iteration)
- Quality control and closed-loop feedback (pre-quality inspection interception, data accumulation and iteration)
- Material generation support (images/audio/subtitles/music)
- Technical detection tool support (image quality, video quality, audio-video synchronization)
Trigger conditions: The user needs to create short videos, or needs to standardize a multi-agent collaborative workflow.

Prerequisites

Dependency description:

opencv-python>=4.8.0
pillow>=10.0.0
moviepy>=1.0.3
numpy>=1.24.0

Non-standard file/folder preparation:

# 创建工作目录
mkdir -p ./input/{materials,scripts}
mkdir -p ./output/{drafts,storyboards,final}
mkdir -p ./cache/{images,audio}

Operation Steps

Standard process (5 stages, 10 agents)

Stage 1: Requirement intake and content framework building

Trigger node: User inputs core requirements (product information + video requirements)

Step 1: Copywriting Creator Agent (first node)

Responsibility: Based on product information, distill core selling points and generate copy suitable for short videos.
Input: Product category, core selling points, target audience, video style, duration
Output:
- Core slogan (8-12 characters, concise and powerful)
- Storyboard-adapted copy (3-5 sentences, each 4-8 characters)
Quality standards: No ambiguity, no exaggeration, fits the video tone
Next-level trigger: Immediately trigger [Story Planner] and [Subtitle Artist] (temporarily stored)

Step 2: Story Planner Agent

Responsibility: Based on selling points and copy, design a scenario-based storyline suitable for the short video duration.
Input: Slogan, storyboard copy, product information, video duration, style
Output:
- Scenario-based storyline (shot breakdown, scene/action annotations)
- Number of shots (8-10 for 30 seconds, 12-15 for 45 seconds, 15-18 for 60 seconds)
Quality standards: Logically coherent, duration strictly matched, no irrelevant scenes
Next-level trigger: Immediately trigger [Script Creator]

Step 3: Script Creator Agent

Responsibility: Transform the storyline into a standardized, executable storyboard script.
Input: Storyline, copy, video duration, resolution, aspect ratio
Output:
- Standardized storyboard script (shot number, duration, scene description, visual actions, copy, sound effect notes)
- Visual action description (clearly specify human body posture, avoid vague expressions)
Quality standards: Precise duration allocation, specific visual actions, unified format
Next-level trigger: Immediately trigger [Storyboard Director], [Storyboard Artist], [Subtitle Artist], and [Sound Effects Artist]

Phase 2: Visual Content Creation and Quality Inspection

Trigger Node: Storyboard script confirmation

Step 4: Storyboard Director Agent

Responsibility: Plan the visual details, lighting, and composition of each shot, providing precise guidance for the storyboard artist
Input: Storyboard script, product detail images (if any), video style
Output:
- Shot visual specifications (composition/lighting/color tone/body specifications/product details)
- Negative avoidance checklist (no extra limbs, deformities, blurring, etc.)
Quality standards: Clear composition, even lighting, proper body specifications, clear product details
Next-level trigger: Immediately trigger [Storyboard Artist]

Step 5: Storyboard Artist Agent

Responsibility: Generate high-quality, flawless storyboard images based on visual specifications
Input: Visual description, storyboard script, product information
Output: 1080P qualified storyboard images (one image per shot, naming convention compliant)
Quality standards:
- Normal human body structure (no extra limbs/deformities/ghosting)
- Accurate product details (no deformation/blurring)
- Fine image quality (1080P, no graininess)
Negative avoidance: No extra limbs, hand deformities, product deformation, image noise, etc.
Next-level trigger:
- After image generation, trigger [Video Engineer] material preprocessing
- Simultaneously trigger [Quality Inspection Agent] for dedicated storyboard image quality inspection

Phase 3: Audio and Subtitle Creation (Executed in Parallel)

Trigger Node: Storyboard script confirmation (can be started in parallel, no need to wait for visual content)

Step 6: Voice Actor Agent

Responsibility: Generate high-quality narration/voiceover based on storyboard copy
Input: Storyboard copy, video style, emotional tone
Output:
- Narration text (corresponding to storyboard copy, appropriately colloquial)
- Voice suggestion (energetic male voice/professional female voice/steady male voice/friendly female voice, etc.)
- Speech rate requirement (medium/slightly fast/slightly slow)
- Emotion annotation (positive/professional/steady/friendly, etc.)
Quality standards: Natural and fluent speech, emotion fits the scene, no obvious mechanical feel
Next-level trigger: Output synchronized to [Sound Effects Artist], callaudio_generator.pyGenerate audio file

Step 7: Subtitle Artist Agent

Responsibility: Create subtitles adapted to the visuals, ensuring display effect and readability
Input: Storyboard copy, storyboard script, video style, aspect ratio
Output:
- Subtitle information package (text + corresponding shot duration + overlay position)
- Format specifications (font/font size/color/position)
Quality standards: No typos, word count adapted (1-2 characters per second), avoid the product subject
Next-level trigger: Output synchronized to [Video Engineer]

Step 8: Sound Effects Artist Agent

Responsibility: Recommend suitable background music and scene sound effects, ensuring audio-visual coordination
Input: Storyboard script, video style, duration, core scenes
Output:
- Sound effect plan package (background music + scene sound effects + time nodes + volume parameters)
- Sound effect name, style description, duration, volume suggestion
Quality standards: Style fits, rhythm matches, moderate volume, no copyright issues
Next-level trigger: Output synchronized to [Video Engineer]

Phase 4: Video Synthesis and Finished Product Quality Inspection

Trigger Node: Storyboard images pass quality inspection

Step 9: Video Engineer Agent

Responsibility: Complete high-quality video synthesis based on materials output by each agent
Input: Qualified storyboard images, subtitle information, sound effect plan
Output: Synthesized MP4 finished video
Quality standards:
- Material preprocessing (call quality inspection script to identify defects)
- Synthesis parameters (1080P, 25fps, 8Mbps, no black edges/stuttering)
- Transition effects (fade in/fade out, 0.3-0.5 seconds, natural transition)
- Audio-video synchronization (audio offset ≤0.1 seconds)
Next-level trigger: Immediately trigger [Quality Inspection Agent] after finished product output

Step 10: Quality Inspection Agent (interspersed throughout the entire process)

Responsibility: Conduct quality inspection of the outputs of each agent across all stages, intercept defects, and drive rectification
Input: Outputs of each stage, quality standards, negative avoidance checklist
Output:
- Stage-by-stage quality inspection report (qualified/unqualified, defect type, rectification suggestions)
- Quality inspection ledger (defect data, rectification results, pass rate)
Quality inspection dimensions:
- Content layer (copy/storyline/script/subtitles)
- Visual layer (storyboard images/finished video)
- Audio layer (background music/sound effects/audio-video synchronization)
- Format layer (aspect ratio/script format/video parameters)
Processing rules:
- Minor defects (minor copy word order adjustments) → direct feedback for optimization
- Major defects (multiple limbs/screen tearing) → intercept output and require remaking
- Pre-inspection → mandatory quality inspection for storyboard images/finished video, rejected if unqualified
Next-level trigger:
- Unqualified → trigger the corresponding agent to redo
- Qualified → trigger [Data Feedback Agent]

Phase 5: Data Accumulation and Iterative Optimization

Trigger Node: Full-process quality inspection passed

Step 11: Data Feedback Agent

Responsibility: Collect and analyze full-process data, output iteration suggestions, and optimize agent prompts and collaboration logic
Input: Quality inspection ledger, creation time data, user feedback
Output:
- Data statistics report (pass rate, high-frequency defects, creation time, rectification rate)
- Iteration optimization suggestions (optimize prompts and adjust collaboration process for high-frequency defects)
- Data accumulation (closed loop of defect type-optimization plan-effect verification)
Analysis dimensions:
- Defect analysis (TOP3 high-frequency defects, locate root causes)
- Efficiency analysis (identify process bottlenecks)
- Optimization effect (verify iteration effectiveness)
Next-level trigger: Output synchronized to the skill development side, update agent configuration, complete the iteration closed loop

Parallel Collaboration and Closed-Loop Control

Parallel Execution

Subtitle Artist, Sound Effect ArtistCan be started in parallel after the Script Creator outputs the storyboard script, without waiting for visual content completion
Efficiency Improvement: Reduce waiting time, shorten the overall creation cycle

Closed-Loop Control

All outputs must be verified by the Quality Inspection AgentOnly after passing can it proceed to the next stage
Major defects are directly intercepted, avoiding problem flow
Track rectification results, ensuring 100% resolution

Iterative linkage

Data feedback agent optimization suggestionsdirectly act on the underlying configuration of each agent
**Form a "creation-quality inspection-optimization"** continuous iteration mechanism

Quality inspection tool usage

Image quality inspection

# 调用图片质量检测脚本
python scripts/image_quality_checker.py --image ./cache/images/shot_001.jpg --resolution 1920x1080

Inspection dimensions:

Limb abnormalities (extra limbs/deformities/ghosting)
Image quality (blur/noise/distortion)
Product details (texture/outline/buttons)
Resolution/aspect ratio

Video quality inspection

# 调用视频质量检测脚本
python scripts/video_quality_checker.py --video ./output/final/video.mp4 --resolution 1920x1080 --duration 30 --fps 25

Inspection dimensions:

Resolution/duration/frame rate/bitrate
Image quality (stuttering/tearing/transition defects)
Audio quality (clarity/noise)
Audio-video sync (offset ≤0.1 seconds)
Subtitle occlusion detection

Material generation and video composition process

Material generation process

1. Image material generation

Trigger timing: After the storyboard artist agent completes the visual design

Execution steps:

# 准备分镜脚本JSON文件
cat > ./scripts/storyboard.json << 'EOF'
{
  "shots": [
    {
      "shot_id": "L01",
      "description": "智能手表特写,手指滑动屏幕",
      "duration": 3.0,
      "resolution": "1920x1080"
    }
  ],
  "style": "科技风,冷色调,清晰锐利"
}
EOF

# 调用图片生成脚本
python scripts/image_generator.py \
  --storyboard ./scripts/storyboard.json \
  --output ./output/images \
  --style "科技风,冷色调"

Output specifications:

Path: ./output/images/shot_{shot_id}.jpg
Format: JPG, 1080P (1920x1080)
Quality requirements: no limb abnormalities, clear product details, delicate image quality

Reference: references/asset-specifications.md - Image material specifications

2. Audio material generation

Trigger timing: After the voice actor agent completes the narration text and the sound effects artist agent completes the sound effects plan

Execution steps:

# 生成旁白音频
python scripts/audio_generator.py \
  --mode narration \
  --text "腕上未来,触手可及" \
  --voice "professional_male" \
  --output ./output/audio/narration_S01.wav

# 生成背景音乐
python scripts/music_generator.py \
  --style "科技风,轻快" \
  --duration 30 \
  --output ./output/audio/background_music.wav

# 混合音频(旁白+背景+音效)
python scripts/audio_generator.py \
  --mode mix \
  --narration ./output/audio/narration_S01.wav \
  --background ./output/audio/background_music.wav \
  --effects ./output/audio/sound_effects/ \
  --output ./output/audio/merged_audio.wav

Output specifications:

Path: ./output/audio/
Format: WAV (44.1kHz, 16bit/24bit)
Quality requirements: no noise, balanced volume, clear sound quality

Reference: references/asset-specifications.md - Audio material specifications

3. Subtitle material generation

Trigger timing: After the subtitle artist agent completes the subtitle design

Execution steps:

# 准备分镜脚本
cat > ./scripts/storyboard.json << 'EOF'
{
  "shots": [
    {
      "shot_id": "L01",
      "duration": 3.0,
      "text": "腕上未来"
    },
    {
      "shot_id": "L02",
      "duration": 3.0,
      "text": "触手可及"
    }
  ]
}
EOF

# 生成SRT字幕
python scripts/subtitle_generator.py \
  --storyboard ./scripts/storyboard.json \
  --format srt \
  --output ./output/subtitles/subtitle.srt

Output specifications:

Path: ./output/subtitles/subtitle.srt or .ass
Format: SRT/ASS (UTF-8 encoding)
Quality requirements: no typos, adapted word count, no occlusion in position

Reference: references/asset-specifications.md - Subtitle material specifications

Video composition process

Trigger timing: After all materials (images/audio/subtitles) are ready and pass quality inspection

Execution steps:

# 准备项目配置
cat > ./scripts/project_config.json << 'EOF'
{
  "images_dir": "./output/images",
  "audio_file": "./output/audio/merged_audio.wav",
  "subtitle_file": "./output/subtitles/subtitle.srt",
  "shots": [
    {
      "shot_id": "L01",
      "duration": 3.0,
      "transition": "fade"
    },
    {
      "shot_id": "L02",
      "duration": 3.0,
      "transition": "cut"
    }
  ],
  "width": 1920,
  "height": 1080,
  "fps": 25,
  "duration": 30,
  "bitrate": "8000k"
}
EOF

# 验证项目配置
python scripts/video_compositor.py \
  --config ./scripts/project_config.json \
  --output ./output/temp/validate.mp4 \
  --validate_only

# 合成最终视频
python scripts/video_compositor.py \
  --config ./scripts/project_config.json \
  --output ./output/final/final_video_$(date +%Y%m%d).mp4

Output specifications:

Path: ./output/final/final_video_{日期}.mp4
Format: MP4 (H.264 encoding)
Parameters: 1080P, 25fps, 8Mbps
Quality requirements: no stuttering, audio-video sync ≤0.1 seconds, subtitle sync without delay

Reference: references/asset-specifications.md - Video output specifications

Quality inspection process

1. Image quality inspection

python scripts/image_quality_checker.py \
  --image ./output/images/shot_L01.jpg \
  --resolution 1920x1080 \
  --check_limb_anomaly \
  --check_blur \
  --check_deformation

2. Audio quality inspection

python scripts/audio_quality_checker.py \
  --audio ./output/audio/narration_S01.wav \
  --sample_rate 44100 \
  --check_noise

3. Video quality inspection

python scripts/video_quality_checker.py \
  --video ./output/final/final_video_20240122.mp4 \
  --resolution 1920x1080 \
  --duration 30 \
  --fps 25 \
  --check_sync

Reference: references/quality-standards.md - Quality inspection standards

Resource index

Required scripts:
- scripts/image_generator.py - Image generation script (generates 1080P images based on the storyboard script)
- scripts/audio_generator.py - Audio generation script (TTS narration + background music mixing)
- scripts/subtitle_generator.py - Subtitle generation script (SRT/ASS format)
- scripts/music_generator.py - Background music generation script
- scripts/video_compositor.py - Main video composition script (integrating images/audio/subtitles)
- scripts/image_quality_checker.py - Image quality detection tool (limb abnormalities/blurriness/deformation)
- scripts/video_quality_checker.py - Video quality detection tool (including audio-video synchronization)
Domain Reference:
- references/agent-prompts.md - Detailed prompt templates for 11 agents
- references/quality-standards.md - Quality inspection standards and negative avoidance checklist
- references/asset-specifications.md - Material generation and output specifications (images/audio/subtitles/video)
- references/data-analysis.md - Data statistics and iterative optimization templates
Output Assets:
- assets/templates/storyboard-template.md - Storyboard script standard template
- assets/reports/quality-report-template.md - Quality inspection report template

Notes

Agent Collaboration: Strictly follow the trigger order and collaboration rules of the 11 agents
Material Specifications: All material generation must followasset-specifications.mdspecifications in
Video Composition: Before composition, ensure all materials (images/audio/subtitles) are in place and pass quality inspection
Quality First: All outputs must be verified by the quality inspection agent; major defects are directly intercepted
Closed-Loop Iteration: Optimization suggestions from the data feedback agent must be applied to the next round of creation
Technical Detection: The quality detection script must be called before video composition; materials that fail inspection must not be composed
Data Retention: Quality inspection ledgers and creation data must be completely recorded for iterative optimization

Usage Examples

Example 1: Complete Creation Workflow

# 用户请求: "为智能手表创作30秒短视频,科技风,目标年轻人"

# 智能体执行流程:
1. 文案创作师: 输出Slogan "腕上未来,触手可及" + 分镜文案
2. 故事策划师: 输出9个镜头的故事线
3. 脚本创作师: 输出标准化分镜脚本(时长、场景、动作、文案)
4. 分镜导演: 输出9个镜头的画面规范 + 负面规避清单
5. 分镜画师: 生成9张1080P分镜图片(调用质量检测脚本)
6. 字幕师: 输出字幕信息包(并行执行)
7. 音效师: 输出音效方案包(并行执行)
8. 质检智能体: 检测分镜图片,发现第3张图片肢体异常,反馈分镜画师重绘
9. 分镜画师: 重新生成第3张图片,质检合格
10. 视频工程师: 合成视频(调用视频质量检测脚本)
11. 质检智能体: 检测成品视频,音画同步0.08秒,合格
12. 数据反馈: 输出数据统计报告 + 迭代建议

Example 2: Image Quality Detection

from scripts.image_quality_checker import check_image_quality

result = check_image_quality(
    image_path="./cache/images/shot_003.jpg",
    expected_resolution=(1920, 1080),
    check_limb_anomaly=True,
    check_blur=True,
    check_deformation=True
)

if not result['overall']:
    print(f"发现瑕疵: {result['issues']}")
    # 触发分镜画师重绘
else:
    print("图片质量合格,可用于合成")

Example 3: Quality Inspection Closed-Loop Example

质检智能体检测:
- 分镜图片: 合格率92%(10张图片,8张合格,1张轻微瑕疵,1张重大瑕疵)
  - 重大瑕疵: shot_007.jpg 发现多余肢体,驳回分镜画师重绘
  - 轻微瑕疵: shot_003.jpg 轻微模糊,反馈优化

成品视频: 合格
- 分辨率: 1920x1080 ✓
- 帧率: 25fps ✓
- 音画同步: 0.08秒 ✓
- 字幕遮挡: 无 ✓

数据反馈智能体:
- 高频瑕疵TOP3: 
  1. 肢体异常(占比8%) → 优化分镜导演/画师提示词
  2. 轻微模糊(占比6%) → 调整画质标准
  3. 字幕位置偏差(占比4%) → 优化字幕师模板
- 效率瓶颈: 分镜画师重绘耗时过长 → 调整前置质检规则

About this skill

Task Objectives

Prerequisites

Operation Steps

Standard process (5 stages, 10 agents)

Stage 1: Requirement intake and content framework building

Phase 2: Visual Content Creation and Quality Inspection

Phase 3: Audio and Subtitle Creation (Executed in Parallel)

Phase 4: Video Synthesis and Finished Product Quality Inspection

Phase 5: Data Accumulation and Iterative Optimization

Parallel Collaboration and Closed-Loop Control

Parallel Execution

Closed-Loop Control

Iterative linkage

Quality inspection tool usage

Image quality inspection

Video quality inspection

Material generation and video composition process

Material generation process

1. Image material generation

2. Audio material generation

3. Subtitle material generation

Video composition process

Quality inspection process

1. Image quality inspection

2. Audio quality inspection

3. Video quality inspection

Resource index

Notes

Usage Examples

Example 1: Complete Creation Workflow

Example 2: Image Quality Detection

Example 3: Quality Inspection Closed-Loop Example