The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation provides recipes for executing shell commands and scripts to manage the agent's environment, such as bunx skills add, bunx skills remove, and scripts/audit-global-skills. These commands perform file operations and environment modifications in user-level directories including ~/.agents, ~/.claude, and ~/.pi.
[REMOTE_CODE_EXECUTION]: The skill automates the download and installation of external scripts and packages. The scripts/audit-global-skills script materializes remote content from GitHub and installs it into local agent directories. While the recommended source is the vendor repository https://github.com/sjunepark/agent-scripts, the bunx skills CLI allows installation from any GitHub repository or git remote.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a gateway for installing 'executable instructions' from external sources.
Ingestion points: External data enters the system through bunx skills add <source> and scripts/audit-global-skills --apply (documented in SKILL.md and references/cli.md).
Boundary markers: The system uses skill-registry.json to define authoritative scope and source, and applies 'when' conditions for project-scoped recommendations to restrict ingestion.
Capability inventory: The skill utilizes subprocess calls (via bunx), file-system writes to agent-specific global paths (~/.agents, ~/.claude), and network operations for fetching remote packages.
Sanitization: The skill relies on SHA256 digest verification and quarantine mechanisms for integrity, but does not implement instruction-level sanitization for the downloaded skill content.

skills-cli