The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates a Python script (loop.py) that invokes the claude CLI with the --dangerously-skip-permissions flag. This flag is designed to bypass the agent's security model by suppressing all user approval prompts for tool execution, creating a high-risk environment for unvetted system changes.\n- [PROMPT_INJECTION]: The skill is vulnerable to Indirect Prompt Injection (Category 8) by creating an autonomous loop that processes untrusted data.\n
Ingestion points: The skill extracts tasks from .claude/plans/*.md files and detects 'quality gate' commands from various project configuration files like package.json, Makefile, and pyproject.toml.\n
Boundary markers: No protective delimiters or 'ignore embedded instructions' warnings are included in the generated prd.json or CLAUDE.md files.\n
Capability inventory: The generated loop has unrestricted access to all agent tools (Bash, File Write, etc.) via the bypassed permission flag.\n
Sanitization: There is no evidence of sanitization or validation of the extracted implementation steps or quality gate commands before they are integrated into the execution logic.\n- [COMMAND_EXECUTION]: During the initialization phase, the skill uses the Bash tool to inspect the local filesystem for testing and linting scripts. This process relies on the contents of files like package.json, which could be manipulated to inject malicious commands into the CLAUDE.md file's mandatory 'Quality Gates' section.

ralph-loop-init