The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to capture the user's direction and lyrics verbatim into a brief.md file, noting that "Every later agent reads them." This design pattern creates a vulnerability where malicious instructions embedded in user-supplied text could influence the behavior of the AI in subsequent automated steps.
Ingestion points: User-provided lyric sheets, director's notes, and song files.
Boundary markers: The instructions do not specify any delimiters or safety warnings to distinguish between data and instructions when the brief is processed by downstream agents.
Capability inventory: The skill utilizes powerful tools including Bash (curl, ffmpeg, python3, bun, npx) and file system access (Read, Write, Edit).
Sanitization: No sanitization or validation of user-provided strings is mentioned; content is used verbatim.
[EXTERNAL_DOWNLOADS]: The skill workflow involves installing external software packages such as demucs via pip and using npx or bun for graphics rendering. While these are common developer tools, they represent a supply chain surface.
[COMMAND_EXECUTION]: Broad shell access is requested for ffmpeg, ffprobe, and python3. These tools are used to process external media files whose metadata or contents could be crafted to exploit vulnerabilities in the media handlers.
[DYNAMIC_EXECUTION]: The skill uses npx and bun to run Node.js based graphics engines like HyperFrames or Remotion. These engines generate code or visual assets dynamically based on the project timing data.

varg-music-video