The Agent Skills Directory

[SAFE]: The skill instructions proactively address security risks by mandating the redaction of API keys and secrets from all logs and evidence records. It also provides explicit warnings against exposing unauthenticated servers to the network and recommends disabling experimental built-in filesystem or shell tools when operating in untrusted environments.
[EXTERNAL_DOWNLOADS]: The skill facilitates the acquisition of source code and models from highly trusted platforms. It provides procedures for cloning the official llama.cpp repository from GitHub and downloading GGUF models from Hugging Face, including instructions for using exact revisions and dry-run verification to ensure artifact integrity.
[COMMAND_EXECUTION]: The core of the skill involves executing local binaries (llama-cli, llama-server, llama-bench) and standard build tools (git, CMake). The procedures are designed to be discovery-first, requiring the agent to verify device support and build configurations before proceeding with resource-intensive operations.

llama-cpp