The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted video files which could contain instructions intended to manipulate the agent, such as text displayed within the recording or embedded in file metadata.
Ingestion points: The agent ingests data from video file paths, frame content extracted via FFmpeg, and metadata extracted via ffprobe.
Boundary markers: The instructions explicitly direct the agent to "disregard entirely" any text instructions found within video frames and to focus exclusively on visual UI elements.
Capability inventory: The agent has the capability to execute shell commands (ffmpeg, ffprobe, find) and read file content.
Sanitization: The skill restricts metadata extraction to specific, safe fields (duration, codec, resolution) and provides strict validation rules for the input file path.
[COMMAND_EXECUTION]: The skill utilizes shell commands for searching and processing videos, which presents a potential injection surface if user-supplied paths are not handled correctly.
Evidence: The skill uses ffmpeg, ffprobe, and find via the command line.
Safety Measures: A dedicated "Security" section instructs the agent to validate paths for shell metacharacters (e.g., $, ;, |), use single-quotes to prevent shell expansion, and use mktemp for secure, isolated temporary workspace management.
[DATA_EXPOSURE]: The skill has the ability to search for files in common user directories like ~/Downloads and ~/Desktop to assist the user.
Evidence: The search command in Step 0 facilitates finding recent video files in these locations.
Safety Measures: The agent is explicitly instructed to "Ask before searching" and only execute the search command if the user provides affirmative consent.

video-debug