The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses yt-dlp via a shell command to download videos. It interpolates the user-provided <URL> directly into the command string: yt-dlp -o "/tmp/react_%(id)s.%(ext)s" -f "mp4/best" "<URL>". Without strict validation or escaping of the URL, this represents a potential command injection vector if a malicious URL containing shell metacharacters is processed.
[EXTERNAL_DOWNLOADS]: The skill relies on several external tools and services to fetch content, including yt-dlp, the Apify actor shu8hvrXbJbY3Eb9W, and the clockworks/tiktok-video-scraper. While these are established tools for this purpose, they involve downloading executable-like binary data (video files) from untrusted third-party platforms.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted data from external video captions, transcripts, and metadata.
Ingestion points: Video metadata and content fetched from YouTube, Instagram, TikTok, and X (Twitter).
Boundary markers: The skill includes a high-level warning header explicitly stating that scraped content is untrusted data and should not be followed as instructions.
Capability inventory: The skill can execute subprocesses (yt-dlp), write to Airtable databases, and upload files to Google Drive.
Sanitization: While there is a conceptual instruction to treat content as untrusted, there are no technical sanitization steps (e.g., escaping input before command execution or validating schemas) to prevent data from influencing agent behavior during the 'Per URL' pass.

react