The Agent Skills Directory

Local File System Interaction: The skill reads from and writes to specific configuration paths within the user directory (e.g., ~/.claude/plugins/config/). While this is used to maintain a 'practice profile' and 'matter workspaces' for consistent legal drafting, it involves accessing local files. This behavior is a security consideration as it creates a dependency on the integrity and privacy of the data stored in these application-specific directories.
Indirect Prompt Injection Surface: The skill is designed to ingest data from external web searches to find model policies and regulations. This creates a potential surface for indirect prompt injection if a retrieved source contains instructions designed to influence the agent's behavior.
Ingestion points: External content from web search results (ABA, state bars, EU AI Act guidance, etc.).
Boundary markers: The skill instructions do not explicitly define delimiters for external content, but they enforce a highly structured output format that may mitigate some impact.
Capability inventory: The skill has the capability to read local files, perform web searches, and write output files to specific matter folders.
Sanitization: The skill relies on human-in-the-loop review by incorporating [review] flags on all choices and open questions, ensuring an attorney must approve the draft content.
External Information Retrieval: The skill performs web searches based on data found in the practice profile (such as jurisdiction and company profile). While this is intended to provide relevant legal sources, it involves sending specific context related to the user's practice to a search engine.

policy-starter