The Agent Skills Directory

[SAFE]: The skill manages student progress by reading from and writing to local files within a specific plugin configuration directory (~/.claude/plugins/config/claude-for-legal-zh/law-student/). This is a standard and safe implementation for persistent state management in AI skills.
[SAFE]: A 'Real Case Check' mechanism is integrated to detect and intercept prompts involving real-world legal situations, correctly directing users to professional legal services instead of providing unauthorized legal advice.
[SAFE]: The skill incorporates a 'Confidence Discipline' protocol that requires the model to explicitly flag uncertainty or refuse to generate content if legal rules are not verified, which helps prevent factual hallucinations in a high-stakes domain.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from study-plan.yaml and CLAUDE.md. While this creates a theoretical injection surface if those files were compromised, the risk is minimal as they are local files managed by the skill itself, and the primary output consists of hypothetical exam questions.
Ingestion points: Reads local exam profile and study history from CLAUDE.md and study-plan.yaml (SKILL.md).
Boundary markers: None explicitly defined in instructions.
Capability inventory: File reading and writing to restricted local paths.
Sanitization: Not present.

bar-prep-questions