The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/generate_video.py uses subprocess.run to call the ffmpeg command-line utility. This is used to concatenate individual scene videos into a final episode file. While this involves system command execution, it is the standard and necessary implementation for the skill's stated purpose of video processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSON script files provided by the user to extract video prompts and scene metadata, creating a surface for indirect prompt injection.
Ingestion points: The ProjectManager.load_script method in scripts/generate_video.py reads local script files which contain the video_prompt fields passed to the generative API.
Boundary markers: The prompt construction logic does not use clear delimiters or systemic warnings to instruct the underlying model to ignore instructions that might be embedded within the script data.
Capability inventory: The skill is capable of calling the Google Veo API for video generation, executing shell commands via ffmpeg, and performing file writes to the local project directory.
Sanitization: There is no evidence of sanitization or character filtering for scene IDs or prompt text, which are used to build file paths and API payloads.

generate-video