The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to clone the official llama.cpp repository from GitHub and install various Python packages (e.g., llama-cpp-python, transformers, numpy) from the standard PyPI registry. These operations target trusted domains and well-known open-source projects necessary for the tool's functionality.
[COMMAND_EXECUTION]: The documentation includes standard build commands using make, quantization binaries (llama-quantize, llama-imatrix), and CLI inference tools (llama-cli, llama-server). These commands are essential for the conversion and deployment workflow described in the skill.
[PRIVILEGE_ESCALATION]: The troubleshooting guide suggests using sudo to install the nvidia-cuda-toolkit via the system package manager on Linux. This is a standard and necessary procedure for installing the system-level drivers and libraries required for GPU acceleration.
[DYNAMIC_EXECUTION]: The skill includes Python scripts that use the subprocess module to query GPU memory status via nvidia-smi and dynamically modify the system path (sys.path) to import conversion logic from the llama.cpp directory. These are routine tasks for hardware-aware AI tooling.
[INDIRECT_PROMPT_INJECTION]: The skill handles external model files and processes user prompts during inference, which represents an inherent attack surface for indirect prompt injection.
Ingestion points: GGUF model files loaded via Llama(model_path=...) and user prompts passed to the generation methods in SKILL.md and references/advanced-usage.md.
Boundary markers: The code snippets mention stop tokens but do not demonstrate explicit boundary markers for untrusted model data.
Capability inventory: The skill has the capability to write files (--outfile), execute subprocesses for hardware checks, and perform network operations during setup.
Sanitization: Standard sanitization is not explicitly detailed in the instructional code snippets.

gguf-quantization