The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to ingest and explain arbitrary workspace content, including source code, UI wiring, and various project artifacts. This creates a surface where malicious instructions embedded in the analyzed data could potentially influence agent behavior.
Ingestion Points: The skill reads files specified by the target argument, as well as open or attached files in the editor context.
Capability Inventory: While primarily read-only, the skill has the capability to modify source files through its 'reconciliation' feature if the user explicitly requests a change. It also dispatches subagents (Explore, rpi-research) which may perform broader codebase analysis or external searches.
Boundary Markers and Sanitization: The protocol uses segment headers and reference tables to structure output, but does not explicitly implement specific escaping or 'ignore instruction' delimiters for the content of the files it is walking through.
User Oversight: As a mitigation, the skill's instructions mandate that no source code changes occur unless the user explicitly chooses immediate reconciliation, providing a human-in-the-loop checkpoint.

rpi-walkthrough