The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a feedback mechanism that returns rule violation messages from project-level configuration files back to the agent context during file operations.
Ingestion points: The engine/check.py script reads and parses the .claude/guardrails.yaml file from the target project to extract rule definitions and their associated message strings.
Boundary markers: Violation messages are concatenated into the reason field of a JSON response returned to the agent. No specific delimiters or "ignore embedded instructions" warnings are applied to these user-defined strings, allowing potential instructions to be interpreted by the LLM as valid commands.
Capability inventory: The skill is capable of reading project files, blocking Write and Edit operations, and providing textual feedback that directly influences the agent's next steps in the tool-use loop.
Sanitization: The message content from the YAML configuration is not sanitized, escaped, or validated against a schema before being interpolated into the feedback string provided to the AI agent.

sdlc-guardrails