The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions and code for building LLM-powered agents that process external data.
Ingestion points: The startThread and continueThread handlers in references/agent.md accept user input directly via the prompt argument.
Boundary markers: The provided code samples and instructions do not include the use of delimiters or specific framing to prevent the underlying LLM from executing instructions potentially embedded within the user-supplied prompt.
Capability inventory: Agents are configured with tool-calling capabilities (e.g., accountLookup, fileTicket) and have access to database mutations and queries within the Convex environment.
Sanitization: There is no evidence of input sanitization or validation to filter malicious instructions from the data before it is passed to the generation model.
[DYNAMIC_EXECUTION]: The skill uses a mechanism for resolving and executing functions based on string handles.
Evidence: In SKILL.md, the documentation promotes createFunctionHandle to pass callback references as strings across component boundaries, which are then cast back to executable handles.
Context: This is a design pattern of the Convex framework for cross-component communication but constitutes a dynamic execution surface where string identifiers determine runtime logic.
[EXTERNAL_DOWNLOADS]: The skill references several external dependencies from a well-known service ecosystem.
Evidence: Instructions include the installation of @convex-dev/agent, @convex-dev/rag, @convex-dev/workflow, and @convex-dev/workpool from NPM.
Context: These modules are official components of the Convex platform and are used as intended for backend functionality.

convex-components