The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads video/audio content from Zhihu domains as part of its core functionality. It also utilizes the modelscope library to download pre-trained speech recognition models from official repositories.
[COMMAND_EXECUTION]: The script relies on external command-line tools ffmpeg and yt-dlp for media downloading and processing. These tools are invoked through vendor-provided wrapper modules in chubby_common.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio data from Zhihu videos and transcribes it into text. This presents an attack surface where audio could contain malicious instructions for a downstream LLM, though the skill's local capabilities are restricted.
Ingestion points: Video URLs provided as arguments and the resulting audio content from Zhihu.
Boundary markers: None present in the generated Markdown output to separate transcribed content from potential system instructions.
Capability inventory: File system write access (os.makedirs, open) and directory cleanup (shutil.rmtree).
Sanitization: The script uses markdown.sanitize_filename to ensure that video titles cannot be used for path traversal attacks during file creation.

zhihu-transcribe