The Agent Skills Directory

[COMMAND_EXECUTION]: The workflow in SKILL.md instructs the agent to execute shell commands (ffprobe and ffmpeg) using the $VIDEO variable, which is derived directly from user-provided local paths or URLs. Without explicit instructions to sanitize shell metacharacters in the filename or path, this presents a command injection risk.
Evidence: ffprobe -v error -show_entries format=duration,size:stream=width,height,r_frame_rate -of json "$VIDEO" in SKILL.md.
Evidence: ffmpeg -y -i "$VIDEO" -vf fps=1 /tmp/video-frames/frame-%03d.jpg in SKILL.md.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted external data (video files and metadata) to drive its core analysis and output generation.
Ingestion points: The skill explicitly accepts local paths, uploaded files, and URLs provided by the user (as described in SKILL.md workflow step 1).
Boundary markers: The skill suggests wrapping the final output in a single fenced text block, which helps separate the prompt but does not protect the agent during the analysis phase.
Capability inventory: The agent is instructed to use file system tools (mkdir) and media analysis tools (ffprobe, ffmpeg).
Sanitization: The instructions do not define any validation or escaping mechanisms for handling external strings before they are used in shell commands or prompt interpolation.

video-to-superprompt