hook-analyzer

Video First-Three-Seconds Hook Analysis (Hook Analyzer)

Overview

The video first-three-seconds hook analysis skill extracts the first-three-seconds shot data from the storyboard breakdown result and constructs structured analysis context, including visual content descriptions, keyframe image URLs, camera movement methods, and other information, providing input for subsequent LLM multi-dimensional scoring analysis.

Use Cases

Short Video Optimization: Evaluate the attractiveness of the video opening and improve the 3-second retention rate
Creative Review: Quantitatively analyze hook quality from 5 dimensions
Competitor Benchmarking: Compare hook strategies of different video openings

Analysis Dimensions

The data extracted by this skill is used for scoring across the following 5 dimensions (scoring is completed by the LLM):

Dimension	Weight	Evaluation Points
Visual Impact	30%	Image composition, color, lighting, camera movement dynamics
Verbal Hook	25%	Copy and image coordination, suspense creation
Emotional Arousal	15%	Image emotion, character expressions, atmosphere creation
Information Density	15%	Effective information volume, core value delivery
Pacing Control	15%	Storyboard switching rhythm, platform characteristic adaptation

Usage Steps

Method 1: Read storyboard data from a file

# 1. 先使用 video-breakdown skill 处理视频并获取分镜结果
python ../video-breakdown-skill/scripts/process_video.py "https://example.com/video.mp4" > breakdown.json

# 2. 提取前三秒分镜数据
python scripts/analyze_hook_segments.py breakdown.json

Method 2: Pass in via stdin pipe

cat breakdown.json | python scripts/analyze_hook_segments.py -

Output Format

{
  "segment_count": 3,
  "total_duration": 2.8,
  "total_video_segments": 15,
  "analysis_mode": "multimodal",
  "segments": [
    {
      "index": 0,
      "start_time": 0.0,
      "end_time": 1.0,
      "duration": 1.0,
      "visual_content": "画面描述",
      "speech_text": "语音文字",
      "shot_type": "特写",
      "camera_movement": "推镜头",
      "function_tag": "开场",
      "frame_images": [
        {"type": "image_url", "image_url": {"url": "https://..."}}
      ],
      "frame_count": 3
    }
  ]
}

Hook Type Classification

The analysis results can be used to identify the following hook types:

Pain Point Type: Directly addresses user pain points and evokes resonance
Curiosity Type: Sets up suspense and arouses curiosity
Conflict Type: Creates contrast or conflict to attract attention
Value Type: Directly demonstrates value promise
Emotional Type: Moves users through emotional resonance
Visual Impact Type: Attracts through strong visual effects
Suspense Type: Leaves suspense to guide continued viewing

Scoring Standards

Score Range	Grade	Description
9-10	Top-tier	Extremely strong attractiveness and creativity
7-8	Excellent	Has good attractiveness
5-6	Average	Room for improvement
3-4	Weak	Needs significant improvement
1-2	Very Poor	Basically unattractive

Notes

This script only performsdata extraction, and does not perform LLM scoring (scoring is completed by the Agent's LLM)
The extracted frame_images field contains keyframe URLs, which can be directly analyzed by the Vision model
Each storyboard takes at most the first 3 keyframes to avoid exceeding the token limit
The input must be process_video the complete JSON data returned

Troubleshooting

Problem	Solution
Output is empty segments	Confirm that the input JSON contains the `segments` field and that it is not empty
No storyboard in the first three seconds	The video may start with a still image; check the original data
Keyframe URL is invalid	The TOS signed URL may have expired; re-decompose to obtain it again