The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes workflows for ingesting model metadata and configuration snippets directly from Hugging Face model pages and their tree API. This creates a surface for indirect prompt injection via external data, although the sources are limited to a well-known service.
Ingestion points: Hugging Face repository pages and API endpoints (e.g., https://huggingface.co/<repo>?local-app=llama.cpp).
Capability inventory: Shell command execution for model inference (llama-server, llama-cli) and file system access for loading models.
Boundary markers: None specified for the data parsed from web pages.
Sanitization: The skill focuses on extracting and using command snippets as provided by the external source.
[EXTERNAL_DOWNLOADS]: Provides instructions for downloading the llama.cpp project from its official GitHub repository and fetching GGUF models from the Hugging Face Hub. These operations target well-known and reputable services relevant to the skill's primary function.
[COMMAND_EXECUTION]: Includes Python code snippets that execute shell commands via subprocess.run to query hardware information (e.g., nvidia-smi) for calculating GPU memory availability and optimizing model layer offloading.
[PRIVILEGE_ESCALATION]: Mentions the use of sudo in the troubleshooting documentation for the installation of system-level dependencies such as the CUDA toolkit. This is standard behavior for configuring specialized hardware environments.
[DYNAMIC_EXECUTION]: Documents the use of LlamaGrammar for defining constrained output formats at runtime and utilizes subprocess calls for dynamic hardware introspection.

llama-cpp