The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for downloading models and tools from well-known and trusted sources. It references official repositories and services including the Hugging Face Hub (huggingface.co), the llama.cpp GitHub repository (github.com/ggml-org/llama.cpp), and standard package managers like Homebrew (brew), Winget, and PyPI (pip). These are documented as standard deployment workflows for the intended purpose of the skill.
[COMMAND_EXECUTION]: The skill includes numerous shell command examples for installing dependencies, compiling software, and running local inference servers. These commands (e.g., cmake, make, llama-server, curl) are standard practices for the local AI deployment use-case and are presented as user-facing instructions rather than automated or hidden scripts.
[DYNAMIC_EXECUTION]: The skill provides Python code snippets for using llama-cpp-python bindings. These snippets demonstrate how to load models, generate text, and manage contexts. While these involve loading binary model files (.gguf), this is the primary and legitimate function of the skill, and the examples use static or user-defined paths.
[SAFE]: Analysis of the 11 threat categories, including prompt injection, data exfiltration, and obfuscation, reveals no malicious patterns. The skill follows best practices for secret management (referencing .env files for configuration) and utilizes established industry domains and organizations for its external dependencies.

llama-cpp