The Agent Skills Directory

COMMAND_EXECUTION (HIGH): The skill generates a bash script (ralph.sh) via init_ralph.py that executes the claude CLI using the --dangerously-skip-permissions flag.
This configuration explicitly bypasses the agent's built-in security sandbox and permission prompts.
Combined with the loop structure, it enables an agent to perform autonomous system modifications (e.g., file writes, shell commands) without human oversight for each step.
PROMPT_INJECTION (LOW): The skill facilitates Indirect Prompt Injection (Category 8) by design, as it iteratively processes tasks that may contain untrusted data from a codebase or external backlog.
Ingestion points: backlog.json and the native Task system (via TaskGet).
Boundary markers: Absent; the generated prompt.md does not use delimiters or instructions to ignore embedded commands within task descriptions.
Capability inventory: Full filesystem and shell access via the claude toolset.
Sanitization: None; task data is consumed directly to guide the agent's next action.

ralph-loop