The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted legal documents (demand letters) to extract information and perform analysis, creating an attack surface where a document could contain hidden instructions intended to manipulate the agent. However, the skill implements strong mitigations including a rigid output schema for triage results and explicit directives that forbid the agent from asserting legal conclusions without verified sources.
Ingestion points: Processes external files through the '读取提供的来函文件' step.
Boundary markers: Results are structured into a specific triage.md format with mandatory source attribution labels like [需审查] and [模型知识].
Capability inventory: Writes analysis files to pre-defined directory paths and suggests transitions to other management skills.
Sanitization: Explicitly warns the agent not to fill legal research gaps with model knowledge without user consent and requires manual review of all citations.
[COMMAND_EXECUTION]: The skill acts as a coordinator, preparing pre-filled arguments for subsequent skill calls such as /matter-intake and /demand-intake. These transitions are triggered based on the analysis of the incoming document but require user interaction to proceed, preventing automated tool abuse.

demand-received