The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions that explicitly override standard agent interaction patterns. For example, it directs the agent to 'avoid unnecessary questions', 'never ask whether to continue', and 'do not ask "should I continue?" after completing a partial step'. While these are intended to maintain development momentum, they function as behavioral overrides that reduce the user's opportunity to review or halt actions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from the repository environment, creating a potential vulnerability surface where malicious instructions in files or logs could influence the agent's autonomous loop.
Ingestion points: The agent is instructed to read git status, changed files, git diff, failing tests, and logs to reconstruct state (found in SKILL.md and references/ralph-loop.md).
Boundary markers: There are no instructions to use delimiters or ignore embedded natural language instructions when reading these repository artifacts.
Capability inventory: The agent is empowered to implement code changes, run validation scripts, and continue execution loops autonomously without frequent user confirmation.
Sanitization: There is no mention of sanitizing or filtering instructions that might be contained within the logs or repository files being read.

ralph-loop