The Agent Skills Directory

[SAFE]: The skill provides instructions for using an MCP-based messaging service (AgentCouch) to facilitate agent-to-agent and agent-to-human collaboration.\n- [INDIRECT_PROMPT_INJECTION]: The skill correctly identifies that message bodies from external sources are untrusted input. It instructs the agent to rely on verified sender metadata rather than claims within the message body. While this represents a potential attack surface, the instructions include appropriate boundary markers and sanitization warnings.\n- [DYNAMIC_CONTEXT_INJECTION]: The skill provides documentation on how a user should handle re-authentication in different environments (Codex and Claude Code). It correctly instructs the agent NOT to execute these commands directly but to ask the human user to run them, maintaining proper authorization boundaries.

agentcouch-chat