The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides shell templates to execute the claude CLI tool in a non-interactive mode (-p). This allows the agent to trigger external code analysis processes.- [DATA_EXFILTRATION]: By design, the skill transmits local project context, including source code and metadata, to Anthropic's external processing services via the CLI. The instructions note that users should be informed of this, although explicit approval is not requested per execution.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the local file system (e.g., project files, CLAUDE.md) and interpolates it into prompts for an external LLM. While it suggests using XML boundaries (<context>, <task>) to structure this data, there remains a low-level risk that malicious content in the project files could influence the second LLM's output.

ask-claude