The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes external data from solicitation portals (SAM.gov, Grants.gov) and user-uploaded documents. This presents a potential surface for indirect prompt injection, where malicious instructions could be embedded in the data. The skill addresses this by referencing shared/untrusted-content.md and explicitly instructing the agent to treat text from messages, documents, and web pages as data only, never as commands.
Ingestion Points: Data enters the context via WebFetch from public portals and Read operations from Google Drive or M365.
Boundary Markers: The skill utilizes standard shared safety protocols and explicit instructions to disregard embedded commands.
Capability Inventory: The skill utilizes Read, WebFetch, and potentially Trello or DocuSign if connected by the user.
Sanitization: Malicious patterns like credential requests or payment changes found in external data are redirected to the owner for manual verification without execution.
External Data Retrieval: The skill fetches information from well-known government and agency portals to identify bidding opportunities. These operations are aligned with the skill's primary purpose and target established, reputable services.
Data Access Controls: The skill includes instructions to verify organizational scope using shared/tenant-scope.md before accessing document stores. This helps ensure the agent does not inadvertently access or process files from a different tenant context.
Human-in-the-Loop Approval: Sensitive operations, such as proposal submission or the creation of tracking boards in Trello, require explicit approval from the owner. This ensures that the agent does not take binding actions autonomously.

grant-rfp-writer