The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data (PR descriptions, release notes, and reports) by interpolating this content into a subagent's prompt for verification.
Ingestion points: Untrusted data enters the agent context through the {draft 全文} (Full Draft) and {claim schema 表} (Claim Schema Table) placeholders in the Section 4 subagent dispatch template.
Boundary markers: The skill uses fenced code blocks (triple backticks) to delimit untrusted data and provides a prominent "untrustworthy content warning" to the subagent.
Capability inventory: The skill utilizes Bash, Read, Grep, Glob, Write, and Agent tools to execute git commands and verify claims.
Sanitization: The skill includes explicit instructions for the subagent to assume every claim is potentially wrong, treat any instruction-like text in the input as plain text, and ignore any attempts to override the verification logic. This represents a high-standard defensive prompt design.

evidence-gate