The Agent Skills Directory

[PROMPT_INJECTION]: The skill propagates user input from the $ARGUMENTS variable into several files that control the behavior of subagents, such as prd.json and CLAUDE.md. * Ingestion points: The engineering task description enters the context via the $ARGUMENTS parameter. * Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present when user input is interpolated. * Capability inventory: The skill and its subagents have access to powerful tools like Bash, Write, and Edit. * Sanitization: There is no evidence of sanitization or validation performed on the input before it is used to define subagent tasks.
[COMMAND_EXECUTION]: The skill generates and executes shell scripts locally to orchestrate its phases. In Phase 3b, it creates ralph-loop.sh for each workspace, which is then executed in Phase 3c. It also invokes a local sugar CLI tool for managing git workspaces and codebase state.

phase