The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to read and execute instructions from an external file located at .agent/execplan-pending.md or a path specified by the user. This exposes the agent to indirect prompt injection risks, as instructions within these files are followed without validation.
Ingestion points: Reads instructions from .agent/execplan-pending.md or user-provided file paths.
Boundary markers: The skill does not define any boundary markers or instructions to ignore malicious commands embedded within the plan.
Capability inventory: The skill allows the agent to "implement" steps, which typically involves file system modifications, web searches, and potentially code execution.
Sanitization: There is no evidence of content sanitization or safety checks before the agent executes the steps in the plan.
[COMMAND_EXECUTION]: The workflow encourages the agent to proceed autonomously without human intervention. The instruction to "make a reasonable decision and proceed" when uncertain and to use "best judgment rather than stopping to ask" when a step is ambiguous reduces the likelihood of the agent flagging malicious or dangerous commands during the execution of the plan.

implement-execplan