The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external sources and processes it through an LLM.
Ingestion points: YouTube video subtitles and audio transcripts are fetched in scripts/transcribe.py using ytdlp and funasr wrappers.
Boundary markers: None. The transcript text is interpolated directly into the translation prompt in translate_text without delimiters or instructions to ignore embedded commands.
Capability inventory: The skill has the ability to write files to the local system (.md output) and execute subprocesses (yt-dlp, ffmpeg, funasr).
Sanitization: There is no evidence of sanitization, filtering, or escaping of the transcript content before it is sent to the DeepSeek API for translation.
[COMMAND_EXECUTION]: The skill relies on external CLI tools and subprocess management.
Evidence: scripts/batch_transcribe.py uses subprocess.run() to execute scripts/transcribe.py. Additionally, scripts/transcribe.py uses the ytdlp wrapper to invoke the yt-dlp command-line tool for metadata and media retrieval. While these calls use list-based arguments (reducing shell injection risk), they represent a significant local execution capability.
[EXTERNAL_DOWNLOADS]: The skill performs several network operations to external services.
Evidence: It fetches video data from YouTube, interacts with the DeepSeek API (api.deepseek.com) for text translation, and downloads AI models from ModelScope (iic/SenseVoiceSmall). These are well-known services and organizations, but the skill's functionality is fundamentally dependent on external data retrieval.

youtube-transcribe