The Agent Skills Directory

[PROMPT_INJECTION]: The instructions utilize strong imperative language (e.g., "MANDATORY", "No exceptions", "NEVER pause") to enforce a specific autonomous workflow. While this directs the agent to execute multiple tool calls without user intervention, the behavior is scoped to the generation of Salesforce metadata using the execute_metadata_action tool and does not attempt to override the underlying system's safety filters.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted user input and local project metadata to generate complex business logic (Salesforce Flows).
Ingestion points: User-provided natural language in userPrompt and custom object/field metadata scanned from the local project for inflightMetadata.
Boundary markers: There are no explicit delimiters or instructions to treat the interpolated userPrompt as data rather than instructions when passing it to the metadata actions.
Capability inventory: The skill uses the execute_metadata_action tool to fetch org schema, select flow elements, and generate XML metadata. It also performs repeated automated loops for element generation.
Sanitization: No validation or sanitization of the input content is performed before it is used to influence the generated XML logic.

generating-flow