The Agent Skills Directory

[SAFE]: The skill uses official and well-known repositories for its operations. It fetches the source code from the official llama.cpp repository on GitHub and suggests downloading models from Hugging Face. These are industry-standard locations for machine learning assets.
[INDIRECT_PROMPT_INJECTION]: The skill describes processes that involve reading and transforming external data, such as model weights and calibration text files.
Ingestion points: Untrusted data enters the context through local model directories and calibration text files (e.g., calibration.txt) used for creating importance matrices.
Boundary markers: The instructions treat these inputs as raw data for the llama-quantize and llama-imatrix tools.
Capability inventory: The skill demonstrates the use of local command-line tools for model processing and inference, including shell-based file creation and compilation via make.
Sanitization: Standard toolchain practices are followed; the agent acts as an operator for specific model transformation commands.
[PRIVILEGE_ESCALATION]: The documentation includes a command to install the NVIDIA CUDA toolkit using sudo apt install. This is a standard system administration task for installing hardware-specific dependencies and is clearly identified as a troubleshooting step for build failures.

gguf-quantization