The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to process external, untrusted data from invoices and vendor statements, which could potentially contain instructions aimed at influencing the agent's calculations or draft outputs.
Ingestion points: Untrusted supporting invoices and vendor statements are processed via a "reader worker" and ingested into the skill's logic.
Boundary markers: The skill explicitly labels external sources as "untrusted," which serves as a conceptual warning, though technical delimiters for the extracted data are not specified in the instructions.
Capability inventory: The skill utilizes an internal general ledger tool (internal-gl MCP) to read prior accruals and produces text-based draft journal entries and schedules.
Sanitization: The skill employs a multi-step process where a separate worker extracts amounts before this skill applies policy, which provides a layer of data validation and reduces the risk of direct instruction execution from the raw documents.

accrual-schedule