The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and 'follow each step exactly' from a plan file, which creates a surface where malicious instructions in processed data could override agent behavior. \n- Ingestion points: Implementation plan files read during the 'Load and Review Plan' step. \n- Boundary markers: The skill does not prescribe the use of delimiters (e.g., XML tags or JSON schemas) to separate external plan content from its own instructional framework. \n- Capability inventory: The workflow involves batch execution of implementation tasks and running verifications, which typically requires file system access and command execution permissions. \n- Sanitization: There is no defined process for sanitizing or filtering the content of the plan file, relying entirely on the agent's critical review to identify concerns.

executing-plans