The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the user/agent to install an external Python package claude-real-video (including a [whisper] variant) from a public registry. This is an unverified third-party dependency.
[COMMAND_EXECUTION]: The skill requires the execution of multiple shell commands for installation and setup.
It instructs the agent to run bash install-skill.sh, an opaque script included in the skill package designed to symlink files into various agent platform directories.
It provides commands for manual installation using cp to sensitive hidden directories like ~/.claude/skills/, ~/.codex/skills/, etc.
It uses the crv CLI tool which executes logic for video processing and file writing.
[PRIVILEGE_ESCALATION]: The installation instructions for Linux systems include the use of sudo apt install ffmpeg, which requires administrative privileges.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection as it is designed to ingest and process untrusted external data.
Ingestion points: Processes remote video URLs (e.g., YouTube), local video files, subtitle tracks, and audio for transcription.
Boundary markers: The documentation includes a "Notes" section warning the agent that media content is untrusted and that it should describe instructions found in videos rather than obeying them.
Capability inventory: The skill can execute CLI commands (crv), write output to directories (-o), save to a knowledge base (--kb), and read local browser cookies (--cookies-from-browser).
Sanitization: The skill relies on natural language instructions for the agent to maintain safety, but there is no technical sanitization of the extracted transcript or frame metadata shown.

claude-real-video-for-agents