The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by reading and writing to files that serve as source-of-truth instructions for AI assistants. This process creates a potential surface for indirect prompt injection if external or user-provided data contains malicious instructions that are subsequently obeyed by the agent.
Ingestion points: Reads existing project configuration files such as CLAUDE.md, AGENTS.md, and modular rule files to identify tool conventions.
Boundary markers: The generated and edited files do not use explicit delimiters or "ignore" warnings to encapsulate user-provided rule content.
Capability inventory: The skill is authorized to use Write, Edit, and restricted Bash commands (git, mkdir) to modify the project environment.
Sanitization: No explicit sanitization or input validation is performed on the user-provided rule name or description before they are committed to the filesystem.

create-rule