The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from CSV, JSON, and Feishu tables to generate diagnosis reports. This ingestion of external content creates a surface for indirect prompt injection where malicious instructions embedded in the data could attempt to influence the agent's reasoning.
Ingestion points: The skill reads data via the scripts/diagnose_notes.py script and the lark-cli tool for Feishu integration.
Boundary markers: The skill includes robust mitigation strategies, such as mandatory evidence tags (【事实】, 【派生】, 【推断】, 【待验证】) and clear instructions in references/output-spec.md to prevent the agent from treating untrusted data as authoritative conclusions.
Capability inventory: The agent uses a local data processing script and the lark-cli tool. These tools are restricted to reading/writing data and do not have arbitrary execution capabilities.
Sanitization: The scripts/diagnose_notes.py script performs extensive data normalization and validation of numeric and boolean fields.
[COMMAND_EXECUTION]: The skill requires the execution of a local Python script (scripts/diagnose_notes.py) and an external CLI tool (lark-cli).
Evidence: The instructions in SKILL.md and references/feishu-cli-policy.md direct the agent to run these tools to process performance data. The Python script is provided with the skill and is focused on data validation, while lark-cli is used for interacting with the well-known Feishu/Lark service. The skill includes policies to ensure these tools are used safely, such as requiring user confirmation for write operations and avoiding the use of sensitive credentials.

xhs-note-performance-diagnosis