The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill is designed to fetch content from external registries via git clone or web downloads. While this is its primary function, it introduces a surface for ingesting untrusted code.
[COMMAND_EXECUTION]: The installation process involves writing files to the local system (~/.claude/skills/) and updating configuration files (CLAUDE.md). It also analyzes hooks/hooks.json and .mcp.json, which are mechanisms for executing shell commands and connecting to external servers.
[PROMPT_INJECTION]: The skill implements logic to detect and warn the user about common prompt injection patterns in the skills it installs (e.g., instructions to ignore previous rules or reveal system prompts). However, it remains susceptible to sophisticated indirect prompt injections hidden within the third-party SKILL.md files it processes.
[DYNAMIC_EXECUTION]: During the installation phase, the skill generates a "Freshness Gate" preamble using a string template and injects it into the beginning of the installed skill. While it uses regex validation for the input data, this dynamic generation of instructions is a form of runtime content modification.

skill-installer