The Agent Skills Directory

External Downloads and Repository Access: The skill facilitates downloading AI models from the Hugging Face Hub and cloning the llama.cpp repository from GitHub. These sources are well-known and established platforms within the AI development ecosystem.
Local Command Execution: Instructions are provided for compiling software using make and executing local inference servers (llama-server, llama-cli). These operations are required to perform the skill's primary task of running models locally.
Untrusted Data Ingestion (Indirect Prompt Injection Surface): The skill processes data from external Hugging Face repository pages and API endpoints to identify model quants and filenames.
Ingestion points: Repository metadata is retrieved from the Hugging Face Hub API (/tree/main) and the local-app view of model pages as described in references/hub-discovery.md.
Boundary markers: The instructions do not currently include explicit delimiters or instructions to ignore potential commands embedded within repository descriptions or metadata.
Capability inventory: The agent has the capability to execute commands for model conversion, quantization, and server initialization across the provided reference files (e.g., llama-server, llama-quantize).
Sanitization: There are no explicit sanitization steps mentioned for the content retrieved from the Hub before it is used to build execution commands.

huggingface-local-models