The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external legal documents (NDAs) and has the capability to update its own configuration file based on patterns found in those documents.
Ingestion points: The skill ingests full-text non-disclosure agreements provided by the user in SKILL.md (via the commercial-legal:review mechanism), which are authored by external third parties.
Capability inventory: The instructions explicitly direct the agent to write new stances and classification criteria to the core configuration and instruction file at ~/.claude/plugins/config/claude-for-legal-zh/commercial-legal/CLAUDE.md.
Boundary markers: There are no explicit delimiters or "ignore embedded instructions" markers defined for the ingested document text; however, the skill mandates a human-in-the-loop step where the agent must ask the user before updating the configuration file.
Sanitization: The skill lacks automated sanitization or filtering for the external content before it is processed or used to suggest new logic for the CLAUDE.md file.

nda-review