The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to ingest and process potentially untrusted data from regulatory summaries, user-pasted text, or web search results. This external content is interpolated into prompt context for analysis and report generation.
Ingestion points: Regulatory change text is provided via user arguments or retrieved through web/research tools in SKILL.md (Step 1).
Boundary markers: The skill does not define specific structural delimiters (e.g., XML tags) for the external text, though it uses citation tags for attribution.
Capability inventory: The skill performs file-write operations to the matter folder and invokes research/web search tools.
Sanitization: The skill mitigates risks by enforcing source attribution tags (e.g., [web search — verify]) and requiring manual verification of AI-generated citations.
Local File System Interaction: The skill reads from and writes to specific paths within the user's plugin configuration directory to manage policy indices and matter contexts.
Evidence: Accesses paths such as ~/.claude/plugins/config/claude-for-legal/regulatory-legal/CLAUDE.md and matter-specific subdirectories.
Context: These interactions are central to the skill's primary function of policy management and include explicit cross-matter isolation checks to prevent unauthorized data access between projects.

policy-diff