The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to clone the llama.cpp repository from the official GGML organization on GitHub for local building and execution.
[COMMAND_EXECUTION]: Instructions include standard software development workflows such as compiling C++ source code using make, installing Python packages via pip, and managing models using command-line tools like huggingface-cli and ollama.
[DYNAMIC_EXECUTION]: One of the advanced usage examples includes a Python snippet that utilizes subprocess.run to call nvidia-smi. This is used legitimately to query system VRAM and calculate the optimal number of model layers to offload to the GPU.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest external model files from HuggingFace and process user-provided calibration text for creating importance matrices. While these are external data ingestion points, they represent standard model optimization surfaces and the skill follows established safety patterns for these workflows.

gguf-quantization