The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for installing official packages from a trusted organization's registry, such as @anthropic-ai/claude-code. It also references official Model Context Protocol (MCP) servers for GitHub and PostgreSQL.
[COMMAND_EXECUTION]: The documentation provides extensive examples for executing shell commands to drive the Claude Code CLI. This includes non-interactive print mode for one-shot tasks and interactive tmux sessions for multi-turn orchestration.
[INDIRECT_PROMPT_INJECTION]: The skill describes an orchestration surface that processes external data from codebases, piped inputs, and web searches, which could be leveraged for indirect prompt injection.
Ingestion points: Processes external content through tools like WebFetch and WebSearch as documented in SKILL.md.
Boundary markers: No specific boundary markers are mandated in the guide, though the tool uses internal safety guardrails.
Capability inventory: The documented environment includes capabilities for file modification, bash command execution, and network access.
Sanitization: The skill explicitly documents security hooks (PreToolUse) designed to sanitize tool inputs and prevent the execution of dangerous command patterns.
[SAFE]: A fork bomb pattern was flagged by static analysis; however, evaluation confirms this pattern is contained within a defensive regex matcher used in a security hook example. The code is intended to detect and block malicious shell inputs, not to execute them.

claude-code