The Agent Skills Directory

Management of Untrusted Data: The skill is designed to handle external inputs such as vendor statements and invoices. The author has proactively addressed security concerns by including an explicit instruction to treat these sources as untrusted, which helps the agent distinguish between data to be processed and instructions to be followed.
Indirect Prompt Injection Assessment: The skill possesses an ingestion surface for untrusted data but manages it through specific architectural choices.
Ingestion points: The agent reads data from supporting invoices and vendor statements.
Boundary markers: The prompt includes a clear warning that these external sources are untrusted.
Capability inventory: The agent uses tools to query internal-gl data and produce text-based reports; it does not have authority to post to the ledger.
Sanitization: The workflow requires the agent to apply fixed firm policies to extracted data, reducing the likelihood of malicious input influencing the final output.
Human-in-the-Loop Safeguards: The output is restricted to producing a schedule and a draft journal entry. By explicitly stating that the output is for controller approval and not for posting, the skill ensures that no financial transactions occur without human verification.

accrual-schedule