The Agent Skills Directory

[SAFE]: The skill implements professional best practices by requiring the agent to perform legal research using primary sources rather than relying on internal model knowledge, which mitigates the risk of hallucination or providing outdated legal advice.
[SAFE]: The skill uses localized configuration paths (~/.claude/plugins/config/claude-for-legal-zh/employment-legal/CLAUDE.md) to define jurisdictional rules and organizational policies. This is a standard practice for specialized AI plugins to maintain consistency and local compliance.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of employment offer letters or descriptions of hiring situations. While this presents a surface for indirect prompt injection, the skill includes strong internal logic that mandates research-based verification, source attribution (e.g., [yuandian检索], [联网检索]), and explicit warnings for non-lawyer users, which effectively manages the risk.
Ingestion points: Reads user-provided offer letters or job descriptions.
Boundary markers: The instructions implement a rigid multi-step workflow that forces the agent to cross-reference data against established jurisdictional tables and external research tools before providing conclusions.
Capability inventory: The skill possesses capabilities to read local configuration files, manage matter-specific workspaces, and perform legal searches. It does not possess arbitrary code execution or unauthorized network exfiltration capabilities.
Sanitization: The skill emphasizes source tagging ([模型知识——需验证]) and requires user confirmation before proceeding with lower-confidence information.

hiring-review