The Agent Skills Directory

[COMMAND_EXECUTION]: The templates/loop.sh script executes the claude CLI tool with the --dangerously-skip-permissions flag. This explicitly bypasses security guardrails designed to prevent unauthorized shell command execution or file system modifications without user approval.
[DATA_EXFILTRATION]: The templates/loop.sh script automatically executes git push origin "$CURRENT_BRANCH" after every iteration of the AI loop. This behavior can lead to the silent exfiltration of sensitive data, configuration files, or credentials if the AI agent commits them to the repository during its build process.
[PROMPT_INJECTION]: The templates/PROMPT_build.md and templates/PROMPT_plan.md files contain instructions that encourage the use of massive amounts of parallel subagents (up to 500) and directives like 'Ultrathink' or specific model overrides. While primarily operational, these instructions attempt to dictate high-level agent architecture and resource consumption in a way that could overwhelm local resources or bypass typical agent constraints.
[EXTERNAL_DOWNLOADS]: The REFERENCE.md file points to an external GitHub repository (https://github.com/ClaytonFarr/ralph-playbook). While documented as a reference, the skill is designed to bootstrap a project using templates derived from this external methodology, creating a dependency on third-party logic for core execution flow.

ralph-loop