The Agent Skills Directory

[COMMAND_EXECUTION] (MEDIUM): The skill provides instructions for enabling shell command hooks that execute automatically during agent events. This provides a mechanism for arbitrary local code execution and persistence if the configuration file is modified. Evidence: SKILL.md documentation of 'command' type hooks.
[PROMPT_INJECTION] (MEDIUM): The 'prompt' hook type creates an indirect injection surface by passing tool inputs directly into LLM prompts via the $ARGUMENTS variable. Malicious data in files or commands being processed could influence the agent's decision-making logic. Evidence: SKILL.md 'hook_type name="prompt"' section.
[COMMAND_EXECUTION] (LOW): The logging example demonstrates the ability to capture tool usage data (including command arguments) and write it to local files, which could be extended to exfiltrate or expose sensitive session information. Evidence: SKILL.md 'Log all bash commands' example.

create-hooks