The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of shell commands to compile source code via make, perform authentication using hf auth login, and run local model servers using binaries like llama-server and llama-cli.- [EXTERNAL_DOWNLOADS]: The skill fetches the llama.cpp repository from its official GitHub source (ggml-org/llama.cpp) and downloads model checkpoints from the Hugging Face Hub. These are well-known services for the AI developer community.- [PROMPT_INJECTION]: The skill defines a surface for indirect prompt injection as it processes data from the Hugging Face Hub (filenames, repository metadata, and API responses) to construct shell commands.
Ingestion points: API endpoints (api/models/<repo>/tree/main) and model pages (huggingface.co/<repo>?local-app=llama.cpp) are used to retrieve file lists and quantization parameters.
Boundary markers: No explicit delimiters or warnings to ignore embedded instructions in external data are provided in the workflow.
Capability inventory: The retrieved data is used as arguments for command-line tools including llama-server, llama-cli, and python conversion scripts.
Sanitization: The skill does not specify validation or escaping of the strings retrieved from external API responses before they are interpolated into shell commands.

huggingface-local-models