GameDev Study — Video Tutorial to Learning Notes

Generate structured learning materials from game development video tutorials across multiple domains. Supports Godot, Unity, Unreal Engine, Blender, and Pixel Art.

Trigger

/gamedev-study <video-url-or-file-path> [--output <output-path>] [--domain <domain>]

Usage Flow

1. Provide video URL → Skill will ask for output directory
2. Specify output path → Transcript and notes will be saved there
3. Domain auto-detection → Or use --domain to specify manually

Options

Examples

# Basic usage - will ask for output path
/gamedev-study https://www.bilibili.com/video/BV1xxx

# With output path specified
/gamedev-study https://www.bilibili.com/video/BV1xxx --output D:\MyNotes

# With domain specified
/gamedev-study https://www.bilibili.com/video/BV1xxx --output D:\MyNotes --domain godot

Output Directory Selection

When --output is not provided, the skill will ask:

请输入笔记输出路径（例如: D:\Godot学习笔记 或 ~/Documents/Notes）:

Supported paths:

• Absolute Windows path: D:\Notes\Godot
• Absolute Unix path: /home/user/Notes
• Relative path: ./notes (relative to current working directory)

URL Mode Flow

1. If --output not provided → ask user for output directory path
2. Run: python <skill-dir>/scripts/main.py "<url>" --output "<output-dir>" (timeout: 10 minutes)
3. Script outputs transcript JSON path to stdout
4. Read the transcript JSON file
5. Generate learning document in the specified output directory

File Mode Flow

1. If .json: read as transcript (expects {full_text, segments} format)
2. If .md: read as existing notes text

Domain Detection

After obtaining the transcript content, determine which domain the video belongs to:

1. If --domain argument is provided → use it directly, skip detection
2. Otherwise → read <skill-dir>/resources/DOMAIN_DETECTION.md and follow its rules:
   • First check URL / title for explicit domain keywords
   • Then scan the first 2000 characters of full_text for domain-specific indicators
   • If 3+ indicators match a single domain → use that domain (high confidence)
   • If indicators are ambiguous or no clear match → ask the user to choose: "无法自动判断该视频的领域，请选择: godot / unity / unreal / blender / pixel-art"

Generate Learning Document

1. Based on the detected domain, read the corresponding prompt file from <skill-dir>/resources/:
   • godot → GODOT_PROMPT.md
   • unity → UNITY_PROMPT.md
   • unreal → UNREAL_PROMPT.md
   • blender → BLENDER_PROMPT.md
   • pixel-art → PIXEL_ART_PROMPT.md
2. Combine transcript content with the domain prompt. When passing transcript data, include:
   • full_text — the complete transcript text
   • segments — the array of {start, end, text} objects (needed for timestamp matching)
   • source — the original video URL or file path (needed for building clickable timestamp links)
   • platform — "bilibili", "youtube", or "local" (determines timestamp link format)
3. Generate the learning document following the prompt structure
4. Timestamp validation: After generation, verify every content-level ### heading has a timestamp. If any is missing, scan segments for the best-matching text and add the timestamp with the correct platform-specific link format.
5. Create output directory if not exists
6. Write result to <output-path>/<sanitized-title>.md

Dependencies

Before first use, install Python dependencies:

pip install -r <skill-dir>/scripts/requirements.txt

Notes

• Transcription uses local faster-whisper (no API key needed)
• Supports Bilibili, YouTube, and local video files
• Audio files are automatically deleted after transcription to save space
• Use --output to specify where notes should be saved