The Agent Skills Directory

Handling of Untrusted Data: The skill provides instructions for building agents that ingest user messages and Adaptive Card data. This represents a potential surface for indirect prompt injection if not properly handled.
Ingestion points: Handlers in files like experts/teams/runtime.routing-handlers-ts.md receive activity text and values directly from users.
Boundary markers: The prompt engineering expert (experts/prompt-engineer.md) recommends the use of structured section tags to delineate data from instructions.
Capability inventory: The resulting agents are capable of performing Microsoft Graph API operations and sending proactive notifications.
Sanitization: A dedicated expert (experts/security/input-validation-ts.md) provides guidance on using zod for schema validation and custom sanitization functions to mitigate these risks.
Dynamic Execution in Examples: A code pattern in experts/teams/a2a.client-basics-ts.md demonstrates the use of eval() for math expression parsing.
Context: This is presented as an illustrative example with an explicit comment advising developers to use a safe parser in production environments.
Local Command Execution: The skill contains patterns for programmatically interacting with official command-line tools using execSync (e.g., in experts/models/foundry-local-ts.md).
Context: These commands are used to manage local model services and are necessary for the intended integration functionality.
External Resource Management: The instructions cover the installation of official developer tools and the download of AI models from well-known repositories.
Context: These resources originate from established organizations and are required for the development workflow the skill teaches.

teams-app-developer