The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements robust defenses against indirect prompt injection. It defines an explicit 'untrusted content boundary' using specific delimiters and explicitly instructs the agent to ignore any commands or policy change requests found within scraped metadata, titles, or descriptions.
[DATA_EXFILTRATION]: Network security is tightly controlled via an explicit allowlist of supported HTTPS hosts (Twitter/X, TikTok, YouTube, Instagram, Facebook). It includes specific instructions to block egress to loopback, private networks, and cloud metadata services, and prohibits the extraction or display of session credentials.
[COMMAND_EXECUTION]: The skill mandates safe execution patterns by requiring subprocess.run([...], shell=False) for invoking external tools like yt-dlp and ffmpeg. It further mitigates risks by sanitizing external video IDs into conservative alphanumeric basenames and rejecting symlinks or directory traversal components.
[EXTERNAL_DOWNLOADS]: The skill utilizes well-known, standard media tools. It promotes secure supply chain practices by recommending the use of requirements.lock with hashes and relying on the user's trusted OS package manager for binaries like ffmpeg rather than automatic unverified downloads.

video-download