The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The agent analyzes video transcripts generated from external sources to identify clip segments. This creates a surface where malicious instructions spoken within a video could theoretically influence the agent's behavior during the analysis phase.\n
Ingestion points: Video metadata from yt-dlp and transcriptions from Whisper or other providers.\n
Boundary markers: The system prompt does not specify delimiters or 'ignore' instructions for the transcribed text.\n
Capability inventory: The agent has access to shell_exec (system commands), file_write, and file_read.\n
Sanitization: There are no explicit steps to sanitize the transcript content before the agent processes it for viral segment analysis.\n- [COMMAND_EXECUTION]: The skill relies extensively on the shell_exec tool to perform its core tasks by running binaries like ffmpeg, ffprobe, yt-dlp, and whisper. This is the primary function of the skill, providing it with significant control over the local command line.\n- [EXTERNAL_DOWNLOADS]: The skill is designed to download video content from user-provided URLs using yt-dlp. It also communicates with external APIs from well-known providers such as OpenAI, Groq, Deepgram, ElevenLabs, Meta (WhatsApp), and Telegram to perform transcription, speech-to-text, and publishing.\n- [DATA_EXFILTRATION]: The skill documentation references the use of the --cookies-from-browser flag in yt-dlp to access local browser cookies for downloading age-restricted content. While this is a standard and documented feature of the legitimate tool used for its intended purpose, it involves the retrieval and transmission of sensitive session data to the video platform.\n- [PRIVILEGE_ESCALATION]: The HAND.toml file includes installation instructions for Linux that utilize sudo to install dependencies via system package managers (e.g., apt, dnf). This is a standard requirement for installing software at the system level on these operating systems.

clip-hand-skill