The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted video files, filenames, and OCR text which could contain malicious instructions designed to influence the agent.
Ingestion points: Video bytes, filenames, metadata, and OCR text extracted from frames in SKILL.md.
Boundary markers: The skill includes an "Untrusted content boundary" section that explicitly instructs the agent to ignore any instructions, links, or tool-use prompts found within the media.
Capability inventory: Executes ffmpeg via shell, creates directories, and writes image files and JSON analysis records to the local filesystem.
Sanitization: Instructs the agent to validate output against a JSON schema, restrict path components to conservative basenames, and reject symlink or path traversal components.
[COMMAND_EXECUTION]: The skill executes shell commands for frame extraction and environment verification.
Evidence: Bash scripts in SKILL.md for running ffmpeg and verifying the Pillow installation.
Security Controls: Explicit instructions are provided to sanitize variables used in shell commands (like {platform} and {video_id}) to prevent command injection. It also uses --require-hashes for package installations.

video-frames