The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess module in analyze.py to execute ffmpeg for extracting audio and performing media transcoding. This is a primary function of the skill but involves direct interaction with the system shell.
[PROMPT_INJECTION]: The skill contains an indirect prompt injection surface where data extracted from external video files could potentially influence agent behavior.
Ingestion points: Audio from video files is transcribed into text via mlx-whisper and included in the final analysis report.
Boundary markers: Transcribed content is formatted within Markdown blockquotes in the output report, but the skill lacks explicit instructions to the agent to ignore any commands found within the transcribed text.
Capability inventory: The agent has access to file system operations and shell command execution through the skill's primary script and environment.
Sanitization: There is no evidence of filtering or sanitization performed on the transcribed text to prevent embedded instructions from being processed by the LLM.
[EXTERNAL_DOWNLOADS]: The mlx-whisper library used in analyze.py downloads pre-trained model weights from Hugging Face repositories. While Hugging Face is a well-known service, this involves fetching external assets at runtime.
[DATA_EXFILTRATION]: Documentation in SKILL.md contains hardcoded absolute file paths (e.g., /Users/baimengke/Documents/...) that expose local directory structures and the author's username.

video-reverse-engineer