The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions direct the agent to fetch content from external Hugging Face repositories and treat the extracted snippets as a "source of truth." This creates a surface where a malicious repository could influence the agent's actions.
Ingestion points: Hugging Face model pages and API endpoints (e.g., huggingface.co/<repo>?local-app=llama.cpp, huggingface.co/api/models/<repo>/tree/main) are used to fetch model configuration and command snippets.
Boundary markers: Absent. The instructions explicitly tell the agent to "copy the exact llama-server or llama-cli command" from the external source.
Capability inventory: The skill involves shell command execution, file system access for model management, and network operations.
Sanitization: Absent. The instructions encourage the agent to reconstruct commands based on external quant labels and filenames without validation.
[EXTERNAL_DOWNLOADS]: Fetches model weights, metadata, and configuration from Hugging Face and the official llama.cpp GitHub repository. These sources are well-known and established services for the skill's intended purpose.
[COMMAND_EXECUTION]: The skill provides numerous instructions for executing shell commands to install software (via brew, winget, or cmake), manage local inference servers, and run CLI tools.
[PRIVILEGE_ESCALATION]: The troubleshooting documentation includes instructions for using sudo apt install to install system-level dependencies such as the NVIDIA CUDA toolkit.
[DYNAMIC_EXECUTION]: The skill references Python scripts that use subprocess.run to execute system tools like nvidia-smi for VRAM detection and provides a utility for inspecting GGUF file headers at runtime.

llama-cpp