The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external customer data (family structure, income, insurance needs) which could potentially contain prompt injection attacks.
Ingestion points: Untrusted data enters the context through the input collection phase defined in SKILL.md.
Boundary markers: The instructions use markdown tables and structured output templates to separate user data from agent instructions.
Capability inventory: The skill utilizes MCP tools (customer-system.get_customer_profile, customer-system.get_interaction_history) and includes a Python validation script (scripts/validate_input.py).
Sanitization: A validation script is included to ensure required fields are present and formatted correctly before processing.
[METADATA_POISONING]: The TEST.md file contains several hardcoded absolute local paths (e.g., /Users/zhangjia/Documents/客户/保险/...). These appear to be artifacts from the author's local environment and do not pose a security risk, though they indicate the documentation was not fully sanitized for distribution.
[EXTERNAL_DOWNLOADS]: The requirements.txt file specifies the agentscope package. This is a known development framework and is appropriate for the skill's implementation.

insurance-agent-client-consulting