The Agent Skills Directory

Processing Untrusted External Data: The skill ingests trademark and legal data from external integrations and tools (such as Solve Intelligence, CourtListener, or Descrybe). This introduces a surface for indirect prompt injection, where malicious instructions embedded in a third-party database record could attempt to influence the agent's behavior.
Ingestion points: Mark descriptions, goods/services metadata, and search results from external registries.
Boundary markers: The skill uses structured markdown templates for output but lacks explicit instructions to ignore embedded instructions in external data.
Capability inventory: The skill can read/write local files and invoke external search tools.
Sanitization: No explicit sanitization or filtering of search results is defined.
Local File System Interaction: The skill reads configuration files and writes triage memos to specific local paths (~/.claude/plugins/config/claude-for-legal/ip-legal/). This is standard for its role as a legal assistant plugin but involves managing potentially sensitive matter data and configuration.
Safety Guardrails and Refusals: The instructions include rigorous, non-softenable disclaimers and refusal logic to prevent the AI from making definitive legal conclusions or clear/not-clear determinations. These guardrails help mitigate risks associated with the AI providing unauthorized or incorrect legal advice.

clearance