The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests large volumes of documentation from the references/ directory to guide agent responses. While this content is derived from official sources, the lack of explicit boundary markers for external data is a structural characteristic of documentation skills.
Ingestion points: Reference files in references/ (e.g., advanced.md, getting_started.md).
Boundary markers: Absent.
Capability inventory: Command execution (pip, torchrun, llamafactory-cli), file system access (configuring JSON/YAML files), and network operations (downloading models from HuggingFace/ModelScope).
Sanitization: Absent.
[COMMAND_EXECUTION]: The documentation contains numerous examples of command-line operations for environment setup, model training (using torchrun and accelerate), and API deployment. These are standard procedures for machine learning development.
[PRIVILEGE_ESCALATION]: The installation guide for Linux mentions using sudo to execute the CUDA uninstaller. This is documented as a necessary administrative step for managing system-level GPU drivers and does not indicate malicious intent.

llama-factory