The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill retrieves model data and repository structures from Hugging Face (huggingface.co) and clones the official repository from GitHub (github.com/ggml-org/llama.cpp). These are well-known and trusted platforms in the technology sector.
[COMMAND_EXECUTION]: The skill provides instructions for executing various command-line tools including llama-cli, llama-server, and llama-quantize to manage local inference and model optimization.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it is designed to ingest and parse untrusted metadata from Hugging Face repositories to construct local shell commands.
Ingestion points: Metadata extraction from Hugging Face tree API and local-app page text (found in SKILL.md and hub-discovery.md).
Boundary markers: Absent; the instructions do not specify delimiters or warnings for the agent when processing external repository strings.
Capability inventory: Subprocess execution for model inference and Python script execution for weights conversion (found in SKILL.md and advanced-usage.md).
Sanitization: Absent; there is no explicit instruction to validate or sanitize filenames or quantization labels before they are interpolated into executable commands.

llama-cpp