The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The toolkit processes user-supplied text briefs and markdown scripts to generate audio, images, and video clips. These inputs are interpolated into command-line arguments for underlying AI models, creating a surface where malicious instructions in the input could influence agent behavior.
Ingestion points: VOICEOVER-SCRIPT.md, scenes.json, and direct command-line prompts provided by the user.
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions for untrusted data.
Capability inventory: The skill has capabilities for file system writes, network communication with cloud APIs, and shell command execution.
Sanitization: No explicit sanitization or escaping of external content is defined before it is passed to tool scripts.
[COMMAND_EXECUTION]: The skill uses shell commands for project management, media processing (ffmpeg, ffprobe), and infrastructure management (npm, modal). It uses the yieldMs pattern to monitor long-running processes, which involves a continuous polling loop with shell access.
[DYNAMIC_EXECUTION]: The toolkit uses modal deploy to upload and execute local Python application code on remote cloud GPU infrastructure. This runtime deployment of code to external environments is a core part of its architecture.
[EXTERNAL_DOWNLOADS]: The skill fetches software packages via pip3 and npm from standard registries. It also interacts with external services including Modal for GPU compute, AceMusic for AI music generation, and Hugging Face for model access.
[CREDENTIALS_UNSAFE]: The agent is instructed to read sensitive environment variables from a .env file, such as ACEMUSIC_API_KEY, R2_SECRET_ACCESS_KEY, and HF_TOKEN. While necessary for the tool's cloud operations, the agent has direct access to these credentials during the setup verification process.

video_toolkit