The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill manages the installation of orchestration components and domain-specific tools (e.g., VectorCraft, PhotoCraft) using package management commands. It also facilitates the retrieval of AI models, such as Whisper for speech recognition, from Hugging Face using specific commit revisions to ensure content integrity.
[COMMAND_EXECUTION]: The skill executes local Python scripts for bootstrapping, project workflows, and management. It also invokes domain-specific CLI tools to perform media editing operations. These executions are performed with safety flags (e.g., -I -B for Python) and are limited to the user's data directory.
[INDIRECT_PROMPT_INJECTION]: As an orchestration tool that processes complex JSON-based workflow plans and external media assets (WAV audio, PNG images), the skill has an inherent attack surface for indirect injection. However, it implements robust mitigations including structural validation of inputs and checksum verification of all assets before processing.
Ingestion points: Loads JSON workflow definitions and external media files provided by the user or upstream tasks.
Boundary markers: Uses explicit schema versions and verifies that input assets match expected metadata.
Capability inventory: Uses subprocess calls to specialized media processing tools and local file system operations within defined project directories.
Sanitization: Employs file hash checking (SHA-256) and formal schema validation for project plans.

artcraft-use