The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the setup of an inference server (llama-server) that ingests untrusted user input via HTTP endpoints.
Ingestion points: Text prompts and chat messages received at /v1/chat/completions and /completion endpoints in references/server-api.md.
Boundary markers: The provided documentation and code snippets do not demonstrate the use of delimiters or "ignore" instructions to prevent the model from obeying instructions embedded in user data.
Capability inventory: As detailed in references/server-api.md, when enabled via the --tools all flag, the server possesses the ability to execute shell commands (exec_shell_command), read files (read_file), and write files (write_file).
Sanitization: No input validation or output sanitization logic is shown in the C++ or HTTP integration examples.
[COMMAND_EXECUTION]: The skill provides instructions for enabling and using tool execution capabilities that interact with the host system.
Evidence: references/server-api.md describes the --tools flag for llama-server, which specifically enables system-level operations like exec_shell_command and file system access (write_file).

llama-cpp