The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted data (referred to as "facts" from primary sources like release notes, repository documentation, or code) to generate a video project. This content is interpolated into data/script.json and subsequently used as arguments for audio generation and alignment tasks.
Ingestion points: references/production-pipeline.md instructs the agent to gather facts from external primary sources to inform the video content.
Capability inventory: The skill provides a suite of local scripts (scripts/*.mjs, scripts/*.py) that have the capability to write to the file system and execute system commands via ffmpeg and the multix CLI.
Sanitization: The templates align-voiceover.mjs and generate-audio-assets.mjs attempt to sanitize inputs using a quote function that escapes double quotes. However, this method does not necessarily prevent shell expansion (e.g., using backticks or $()) when executed through a shell.
Boundary markers: The skill does not provide specific instructions to the agent to treat the collected facts as untrusted or to use boundary markers during interpolation.
[COMMAND_EXECUTION]: The skill's workflow depends on the execution of multiple command-line tools, including ffmpeg, the multix CLI tool, and the hyperframes framework via npx.
Evidence: Python scripts such as fit-beat-grid.py, verify-arrangement.py, and measure-mix-balance.py use subprocess.run to invoke ffmpeg for audio analysis and processing.
Risk Factor: The JavaScript templates align-voiceover.mjs and generate-audio-assets.mjs utilize child_process.spawn with the { shell: true } option. While this is functional for the intended media tasks, executing user-influenced strings through a shell shell is a security risk if the sanitization (in this case, manual double-quote escaping) is insufficient.

motion-video