The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of Mermaid text, BPMN XML, and images of diagrams.
Ingestion points: Mermaid code blocks, .mmd files, .bpmn files, and various image formats processed via Vision.
Boundary markers: The skill mandates a multi-stage pipeline (Detect, Extract, Validate, Generate). The Intermediate Representation (IR) validation phase (Step 4) acts as a strict structural boundary before any code generation occurs.
Capability inventory: Reads local files, performs image inference, generates source code in five languages, and manages workflows on the Diagrid Catalyst platform using tools like catalyst_start_workflow.
Sanitization: Instructions explicitly forbid using eval() or expression engines on diagram labels, mandating they be treated only as static text or field names.
[EXTERNAL_DOWNLOADS]: The skill references official Dapr SDK packages across multiple registries (e.g., dapr-ext-workflow for Python, Dapr.Workflow for NuGet, @dapr/dapr for NPM). These are well-known, trusted libraries for building distributed workflows.
[COMMAND_EXECUTION]: Instructions provide guidance for generating code that interacts with the Diagrid Catalyst platform. The skill emphasizes secure practices, such as using catalyst_get_connection to handle tokens via environment variables rather than persisting them in chat history or files.
[DYNAMIC_EXECUTION]: The skill generates source code from the diagram IR. This process uses static templates and language-specific SDK patterns to ensure the generated code is deterministic and suitable for replay-based workflow engines.

catalyst-workflow-from-diagram