The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/fetch-tiktok-transcript.py executes yt-dlp and mlx-whisper via subprocess.run. All commands are constructed as lists (not strings) and do not use shell=True, which prevents shell injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill downloads TikTok video metadata, captions, and audio using yt-dlp. It also facilitates the download of the mlx-whisper transcription model from the mlx-community repository (a recognized ML source for Apple Silicon). The use of uvx for mlx-whisper execution is a standard practice for environment isolation.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted TikTok captions and audio transcripts, which are potential ingestion surfaces for indirect prompt injection.
Ingestion points: Transcript files (transcript.json, transcript.md, transcript.txt) are generated from external TikTok content in scripts/fetch-tiktok-transcript.py.
Boundary markers: The workflow encourages reading metadata.compact.json first and using specific commands like preview, search, or excerpt to limit the volume of untrusted data loaded into the agent's context.
Capability inventory: The skill's capabilities include reading/writing files in the tiktok-transcripts directory and executing external tools for data retrieval.
Sanitization: Subtitle content is sanitized using html.unescape and strip_vtt_markup before being written to transcript artifacts, reducing the risk of control character injection.

tiktok-transcript