The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from user queries and external web search results. It implements a robust mitigation strategy by requiring the agent to label all citations with source-specific tags (e.g., [联网检索——需复核] for web results, [模型知识——需验证] for internal memory) to ensure the user can evaluate the reliability of the advice.
Ingestion points: User-provided legal questions and text snippets from external web searches.
Boundary markers: Mandatory source tags and structured output sections for research versus conclusions.
Capability inventory: Web search, specialized legal research tool (yuan dian MCP) integration, and local file access within project-specific workspaces.
Sanitization: No programmatic sanitization is defined, but the workflow enforces strict human verification of all AI-generated citations.
[COMMAND_EXECUTION]: The skill utilizes integrated tools including a legal research MCP tool and web searching to fulfill its primary function. These are standard tool invocations within the agent's expected capabilities.
[DATA_EXFILTRATION]: The skill accesses a configuration file at ~/.claude/plugins/config/claude-for-legal-zh/employment-legal/CLAUDE.md and reads/writes to local matter workspaces. This access is scoped to the skill's own operational configuration and specific project files, with no evidence of exfiltrating sensitive system or user credentials to external endpoints.

wage-hour-qa