The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill integrates capabilities for processing external data sources such as images, audio recordings, and PDF documents. This creates a surface area for indirect prompt injection, where malicious instructions embedded in media files could attempt to influence agent behavior.
Ingestion Points: The skill instructions involve reading external PDF files (via the Read tool) and processing user-provided images and audio for model analysis.
Boundary Markers: The implementation patterns lack explicit prompt boundaries or 'ignore-instruction' markers for the untrusted content being processed.
Capability Inventory: The skill is configured to use environment tools including Read, Glob, Grep, WebFetch, and WebSearch to facilitate document processing and information retrieval.
Sanitization: No explicit sanitization or validation logic is provided for the content extracted from vision or audio models before it is returned to the agent context.
[EXTERNAL_DOWNLOADS]: The skill recommends the use of official SDKs from trusted providers and well-known services. This includes anthropic, google-genai, openai, and the Vercel ai SDK. These references target official package registries and represent standard development practices.
[COMMAND_EXECUTION]: The documentation provides patterns for utilizing system-level tools such as Read, Glob, and Grep for file discovery and content extraction. These commands are used for their intended purpose of processing project documents.

multimodal-llm