The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the ffmpeg system utility via subprocess.run to perform scene detection and keyframe extraction from video files. It also executes scripts from other local skills (scillm/run.sh and memory/run.sh) to perform analysis and persistence tasks.
[PROMPT_INJECTION]: User-provided strings from the --director command-line option are interpolated directly into the instructions sent to the Vision-Language Model (VLM). This represents a direct prompt injection surface where a user could manipulate the model's analysis output.
[PROMPT_INJECTION]: The skill processes external video content and parses the resulting VLM analysis into structured memory. This introduces an indirect prompt injection surface, as adversarial content within the processed video frames could potentially influence the model's output or the data subsequently stored in the agent's memory.

learn-movie