The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external video data which can include untrusted content such as subtitles, transcripts, and OCR text extracted from frames. This creates a risk where a video could contain embedded instructions designed to manipulate the agent's behavior.
Ingestion points: The skill processes content from any yt-dlp supported URL or local file path via the watch-skill watch command.
Boundary markers: The instructions do not define clear boundaries or 'ignore' directives for the agent when interpreting extracted video data.
Capability inventory: The skill utilizes the Bash tool for command execution and the Read tool for accessing local files.
Sanitization: No filtering or validation of the extracted OCR or transcript text is specified before it is provided to the agent context.
[COMMAND_EXECUTION]: The skill relies on the Bash tool to perform all operations, including installation and video processing.
Evidence: The skill instructs the agent to run commands such as watch-skill doctor, pip install watch-skill, and watch-skill watch.
[EXTERNAL_DOWNLOADS]: The skill downloads external software and dependencies during its setup and operation.
Evidence: The agent is instructed to install the watch-skill package via pip. Additionally, the watch-skill doctor command is described as auto-bootstrapping binary dependencies like ffmpeg and yt-dlp, which are well-known tools for media processing.

watch