The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to process various forms of external data, including call transcripts, voice memos, and RFP documents. This ingestion of untrusted content presents a potential surface for indirect prompt injection, where instructions hidden within the discovery materials could attempt to alter the agent's behavior. The skill proactively addresses this risk with explicit instructions to treat all such data as information rather than commands and to flag suspicious content—such as requests for credential changes or urgent payments—to the user without executing them.
Ingestion points: Processes external files such as transcripts, RFPs, and photos (referenced in SKILL.md and reference/discovery_extraction.md).
Boundary markers: The skill includes a 'What not to do' section that serves as a behavioral boundary, though explicit technical delimiters for inputs are not specified.
Capability inventory: Accesses file systems via the Read tool, performs network operations via WebFetch, generates DOCX/PDF files, and interacts with third-party connectors for invoicing and document signing.
Sanitization: Relies on clear instructional directives to ignore embedded commands and escalate suspicious patterns to the user for manual review.
Sensitive Data Handling: The instructions specifically caution against exposing sensitive proposal details via public URLs when using external signature services. This demonstrates a security-conscious approach to managing the lifecycle of documents that contain proprietary pricing and client information.

proposal-builder