SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsbilibili-analyzer

bilibili-analyzer

Writing & research

View repository

Updated 10-15 16:43

About this skill

Bilibili Video Analyzer

Description

Bilibili video content analysis tool. After providing a video URL, it automatically downloads the video, breaks it down into frame images, then uses AI to analyze the content, ultimately generatinghigh-quality feature documents or hands-on tutorials。

Core Features:

  • It is not a simple timeline record, but ratherreorganizes and arrangesinto a complete document
  • Hands-on videos → generatedirectly usable operation tutorials
  • Knowledge videos → generatestructured feature documents
  • Insert key screenshots into the report, using the ![描述](./images/frame_xxxx.jpg) format

Source & Documentation

Tool Purpose Documentation
FFmpeg Video frame extraction Official Website / Documentation
Bilibili API Video download API Documentation

Installation

1. Install .NET 10 SDK

The script uses the .NET 10 single-file execution feature, so the .NET 10 SDK needs to be installed.

Download address: https://dotnet.microsoft.com/download/dotnet/10.0

Verify installation:

dotnet --version

2. Install FFmpeg

Windows:

# Chocolatey
choco install ffmpeg

# 或 Scoop
scoop install ffmpeg

# 或手动下载: https://ffmpeg.org/download.html

macOS:

brew install ffmpeg

Linux:

# Ubuntu/Debian
sudo apt install ffmpeg

# CentOS/RHEL
sudo yum install ffmpeg

Verify installation:

ffmpeg -version

Trigger

  • /bilibili-analyzer command
  • The user requests analysis of a Bilibili video
  • The user provides a Bilibili video link and requests analysis

Provided Script

This skill provides the scripts/prepare.cs script for downloading videos and extracting frame images.

Script location: skills/tools/bilibili-analyzer/scripts/prepare.cs

How to run: Use the .NET 10 single-file execution feature

Usage

# 基本用法
dotnet run scripts/prepare.cs "<视频URL>" -o <输出目录>

# 示例
dotnet run scripts/prepare.cs "https://www.bilibili.com/video/BV1xx411c7mD" -o ./output

# 长视频（降低帧率）
dotnet run scripts/prepare.cs "https://www.bilibili.com/video/BV1xx411c7mD" -o ./output --fps 0.5

Parameter description

Parameter Description Default value
url Bilibili video URL (required) -
-o, --output Output directory Current directory
--fps Number of frames extracted per second 1.0
--similarity Similarity threshold (0-1); adjacent frames exceeding this value will be deduplicated 0.80
--no-dedup Disable similar frame deduplication false
--video-only Only download the video, do not extract frames false
--frames-only Only extract frames (video.mp4 must already exist) false

Similar frame deduplication

The script automatically performsadjacent framessimilarity detection and removes duplicate frames whose similarity exceeds the threshold (80% by default):

  • Use ffmpeg's SSIM/PSNR algorithm to calculate similarity
  • Only compare adjacent frames, no cross-frame comparison
  • After deduplication, automatically renumber (frame_0001.jpg, frame_0002.jpg, ...)
  • Can be adjusted via --similarity 0.85 to adjust the threshold
  • Can be disabled via --no-dedup to disable deduplication

Output structure

<输出目录>/
├── video.mp4           # 下载的视频文件
└── images/             # 帧图片目录
    ├── frame_0001.jpg
    ├── frame_0002.jpg
    ├── frame_0003.jpg
    └── ...

Workflow (Prompt)

You are a video content analysis assistant. When the user provides a Bilibili video link, execute the following steps:

Step 1: Download the video and extract frames

Use the provided script to download the video and break it down into frame images:

dotnet run skills/tools/bilibili-analyzer/scripts/prepare.cs "<视频URL>" -o <输出目录>

Notes:

  • Short videos (<10 minutes): use the default --fps 1
  • Medium videos (10-30 minutes): use --fps 0.5
  • Long videos (>30 minutes): use --fps 0.2

Step 2: Analyze the frame images

Use the Task toolto analyze the images in the images/ directory in batches in parallel.

Batching strategy(calculated dynamically based on the total number of images):

Total number of images Number of batches Images per batch
1-30 1 batch All
31-60 2 batches About 15-30 images/batch
61-120 3 batches About 20-40 images/batch
121-200 4 batches About 30-50 images/batch
200+ 5 batches Evenly distributed

Calculation formula:

总图片数 <= 30: 1 批
总图片数 <= 60: 2 批
总图片数 <= 120: 3 批
总图片数 <= 200: 4 批
总图片数 > 200: 5 批

每批图片数 = 总图片数 / 分批数量（向上取整）

Example: Suppose there are 85 images → split into 3 batches

Task 1: 分析 frame_0001.jpg ~ frame_0029.jpg（29张）
Task 2: 分析 frame_0030.jpg ~ frame_0058.jpg（29张）
Task 3: 分析 frame_0059.jpg ~ frame_0085.jpg（27张）

Task Prompt template:

读取并分析 <输出目录>/images/ 目录下的 frame_0001.jpg 到 frame_0020.jpg（共20张图片）。

【重要要求】
你的响应必须是这些图片内容的【完整详细报告】，不要省略任何信息。

对每张图片，详细记录：

1. **帧号**: frame_xxxx.jpg
2. **场景类型**: 代码编辑器/终端/浏览器/PPT/对话/其他
3. **界面内容**:
   - 窗口标题、菜单、按钮等UI元素
   - 当前打开的文件/页面
4. **文字内容**:
   - 完整转录屏幕上的所有文字
   - 代码内容（完整复制，保留格式）
   - 终端命令和输出
   - 注释和说明文字
5. **操作动作**:
   - 鼠标位置、点击目标
   - 正在进行的操作
6. **关键信息**:
   - 重要的配置项
   - 关键步骤说明
   - 错误信息或警告

【输出格式】

## frame_0001.jpg
- 场景: [场景类型]
- 内容: [详细描述]
- 文字/代码:

[complete text or code content]

- 操作: [正在进行的操作]
- 要点: [关键信息]

## frame_0002.jpg
...

【注意】
- 不要省略任何图片
- 代码和文字必须完整转录
- 信息越详细越好

Analysis points:

  1. Fully transcribe all text and code content
  2. Describe UI elements and operation steps in detail
  3. Record the key information of each image
  4. Mark important screenshot frame numbers (such as frame_0042.jpg)

Step 3: Generate document

According to the video type, reorganize and arrange the analysis resultsreorganize and arrangeinto 视频分析.md：

Determine the video type:

  • Practical type: programming tutorials, software operations, configuration demonstrations, etc.
  • Knowledge type: concept explanations, principle analysis, experience sharing, etc.

[Key] Images and content must strictly correspond:

错误示例 ❌:
### 安装 Node.js
首先下载 Node.js...
![截图](./images/frame_0001.jpg)  ← 图片可能是其他内容

正确示例 ✅:
### 安装 Node.js
首先下载 Node.js...
![frame_0015: Node.js官网下载页面](./images/frame_0015.jpg)  ← 图片确实是下载页面

Correct process for generating documents:

  1. First organize the analysis results returned by all Tasks

    • Summarize the analysis content of all frames
    • Establish the correspondence between "frame number → content"

  2. Reorganize the content by theme(not in chronological order)

    • Classify related content into the same chapter
    • Determine which frames' information each chapter needs

  3. Must verify when inserting images

    • Only insert images that are directly related to the current contentdirectly relatedimages
    • Image descriptions must accurately reflect the actual content of the image
    • Use format: ![frame_xxxx: 图片实际内容描述](./images/frame_xxxx.jpg)

  4. Code must come from the actual code in the image

    • Do not make up code yourself
    • Code block source annotation: <!-- 来自 frame_0025 -->

Important principles:

  1. Image-text correspondence - Each image must match its contextual content
  2. Do not write a chronological running account - Reorganize the content, like writing an article
  3. Clear structure - Has chapter divisions and logical order
  4. Authentic code - Only use code that appears in the images, do not fabricate
  5. Independently readable - Can be fully understood without watching the video

Output Format

Practical tutorial type

# {教程主题}

## 简介

{教程目标}
{前置条件和要求}

## 环境准备

{需要安装的软件}
{配置要求}

## 操作步骤

### 1. {步骤标题}

{详细说明，内容必须与下方图片对应}

![frame_xxxx: 图片实际内容的准确描述](./images/frame_xxxx.jpg)

<!-- 代码来自 frame_xxxx -->
```代码块```

### 2. {步骤标题}

{详细说明}

![frame_xxxx: 准确描述](./images/frame_xxxx.jpg)

...

## 完整代码

<!-- 汇总自 frame_xxxx, frame_xxxx, frame_xxxx -->
{汇总所有代码片段，标注来源帧号}

## 常见问题

{可能遇到的问题和解决方案}

## 总结

{核心要点回顾}
{延伸学习建议}

Knowledge document type

# {主题}

## 概述

{主题背景介绍}
{为什么重要}

## {章节1标题}

{内容，必须与配图对应}

![frame_xxxx: 图片实际内容描述](./images/frame_xxxx.jpg)

## {章节2标题}

{内容}

![frame_xxxx: 图片实际内容描述](./images/frame_xxxx.jpg)

## 核心要点

- 要点1
- 要点2
- 要点3

## 延伸阅读

{相关资源和建议}

Image insertion specifications

Rule Description
Frame number must be marked ![frame_0015: 描述](./images/frame_0015.jpg)
Description must be accurate Describe the actual content of the image, not the expected content
Content must match The text above/below the image must be related to the image content
Code source labeling <!-- 代码来自 frame_0025 -->
Do not insert images randomly If there is no suitable image, do not insert one; do not force an image into the text

API Reference

Bilibili API

The script uses Bilibili's official API to download videos:

# 获取视频信息
GET https://api.bilibili.com/x/web-interface/view?bvid=BV1xx411c7mD

# 获取播放地址
GET https://api.bilibili.com/x/player/playurl?bvid=BV1xx411c7mD&cid={cid}&qn=80&fnval=1

API documentation: https://github.com/SocialSisterYi/bilibili-API-collect

FFmpeg frame extraction command

# 每秒1帧
ffmpeg -i video.mp4 -vf "fps=1" -q:v 2 images/frame_%04d.jpg

# 每秒0.5帧（每2秒1帧）
ffmpeg -i video.mp4 -vf "fps=0.5" -q:v 2 images/frame_%04d.jpg

# 指定时间范围
ffmpeg -i video.mp4 -ss 00:01:00 -to 00:05:00 -vf "fps=1" -q:v 2 images/frame_%04d.jpg

# 提取关键帧（场景变化）
ffmpeg -i video.mp4 -vf "select='gt(scene,0.3)'" -vsync vfr -q:v 2 images/frame_%04d.jpg

More options: https://ffmpeg.org/ffmpeg.html

Examples

Example 1: Analyze a programming tutorial

# 1. 下载并拆帧
dotnet run scripts/prepare.cs "https://www.bilibili.com/video/BV1xx411c7mD" -o ./react-tutorial

# 2. 分析图片（使用 Task 工具）
# 3. 生成 react-tutorial/视频分析.md

Example 2: Analyze a long video

# 降低帧率，减少图片数量
dotnet run scripts/prepare.cs "https://www.bilibili.com/video/BV1xx411c7mD" -o ./long-video --fps 0.2

Example 3: Download only the video

dotnet run scripts/prepare.cs "https://www.bilibili.com/video/BV1xx411c7mD" -o ./output --video-only

Quality Checklist

Before generating the document,check item by itemthe following requirements:

Content quality

  • Content is reorganized, not a chronological running account
  • Chapter structure is clear and has a logical order
  • All content can be understood without watching the video
  • Includes environment preparation/prerequisite explanations
  • Includes a summary and key points

Image-text correspondence (important!)

  • Each image is labeled with its frame number: ![frame_xxxx: 描述](...)
  • Image descriptions accurately reflect the imageactual content
  • The text in the context of an image isdirectly related
  • No unrelated images are inserted arbitrarily
  • Code blocks are labeled with the source frame number: <!-- 来自 frame_xxxx -->

Code quality

  • Code comes from the actual code in the images, not fabricated
  • Code snippets are complete and can be copied and used directly
  • The complete code chapter summarizes all code and labels its source

Tags

bilibili, video-analysis, ai, frame-extraction, markdown, tutorial, ffmpeg, dotnet

Compatibility

  • Codex: Yes
  • Claude Code: Yes