The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted audio content from TikTok and converts it to text, which is subsequently written to a Markdown file. Malicious instructions hidden in the audio could theoretically influence downstream agents processing the resulting file.
Ingestion points: The audio_path is generated from a user-provided TikTok URL in scripts/transcribe.py.
Boundary markers: The skill does not use specific delimiters or ignore-instructions warnings when writing the transcribed text into the Markdown body.
Capability inventory: The skill has the capability to write to the file system using os.makedirs and file write operations in scripts/transcribe.py.
Sanitization: Filenames are sanitized using markdown.sanitize_filename, but the transcribed text is interpolated directly into the file without escaping or validation.
[REMOTE_CODE_EXECUTION]: The documentation in SKILL.md suggests loading the iic/SenseVoiceSmall model with the trust_remote_code=True flag.
Evidence: The configuration snippet in SKILL.md (Step 2) includes trust_remote_code=True. This allows the execution of arbitrary code defined in the model repository on the ModelScope platform.
Context: The model is hosted by the Institute of Intelligent Computing on ModelScope, which is a well-known service.
[EXTERNAL_DOWNLOADS]: The skill downloads media from TikTok and machine learning models from ModelScope.
Evidence: Uses yt-dlp for media extraction and modelscope/funasr for model fetching.
[COMMAND_EXECUTION]: The skill executes external command-line tools to perform its primary function.
Evidence: Subprocess execution of yt-dlp for downloading and audio conversion as described in SKILL.md and implemented in the chubby_common dependency used by scripts/transcribe.py.

tiktok-transcribe