The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands if a shell is available. This includes running local Python scripts (swipe.py, hookscore.py) and utilizing the yt-dlp utility for data collection.
[INDIRECT_PROMPT_INJECTION]: The skill ingests YouTube video titles and channel metadata from external sources. Because this content is attacker-controlled, it represents an indirect prompt injection risk. Malicious titles could be crafted to influence the agent's analysis or subsequent responses.
Ingestion points: Data from the search_youtube_videos tool, yt-dlp output, or user-pasted video lists.
Boundary markers: Absent. The instructions do not specify delimiters or warnings to ignore embedded instructions in the video titles.
Capability inventory: The agent has the capability to execute shell commands and read/write files via the provided scripts.
Sanitization: The swipe.py script uses regex to classify titles but does not perform sanitization to prevent the agent from interpreting instructions contained within the titles during the 'formula naming' or 'full plan' phases.

yt-viral