The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: During project initialization in code mode, the scripts/mv.py script clones a code template from an external GitHub repository (https://github.com/JohnHeibel/ClaudeAnimationBase.git).
[REMOTE_CODE_EXECUTION]: The agent is instructed to execute node render.mjs within the cloned template to generate animation frames. This results in the execution of unverified external code on the local system.
[CREDENTIALS_UNSAFE]: The workflow in SKILL.md instructs the agent to check for the presence of an API key at ~/.config/noiz/api_key. Accessing hardcoded sensitive file paths for credentials increases the risk of exposure.
[PROMPT_INJECTION]: The file references/opus-5.5-prompt.md contains meta-instructions that attempt to override the model's standard operating behavior with directives such as "Run the host at xhigh effort" and instructions to ignore certain reasoning or budget constraints.
[COMMAND_EXECUTION]: The script scripts/mv.py frequently uses subprocess.run to interact with the operating system, calling tools like ffmpeg, ffprobe, git, and node. While it avoids shell injection by using list-based arguments, it processes various user-controlled file paths and generated configuration files.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (audio, lyrics, and video assets) to generate creative briefs and storyboards. There are no explicit sanitization steps for these inputs, which could allow malicious instructions embedded in file metadata to influence the agent's creative planning process.

music-video