The Agent Skills Directory

[PROMPT_INJECTION]: The skill employs extremely directive and repetitive instructions (e.g., 'MANDATORY', 'STRICT', 'No exceptions', 'No shortcuts') to enforce a specific execution path. While intended for technical accuracy, such patterns are analyzed for their potential to override core model constraints.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests untrusted data from the userPrompt and local project files (inflightMetadata). There are no defined boundary markers or sanitization steps to prevent embedded instructions in these sources from hijacking the flow generation process.
Ingestion points: userPrompt input and inflightMetadata (scanned from local sfdx project files).
Boundary markers: Absent; data is passed directly into tool parameters without delimiters.
Capability inventory: Calls execute_metadata_action to fetch schema, select elements, and generate XML metadata.
Sanitization: Not observed in the instruction set.
[COMMAND_EXECUTION]: The skill orchestrates multiple calls to the execute_metadata_action tool, including a repetitive loop in the final step. The instructions mandate that the agent must not pause or seek user confirmation during this loop, which reduces human oversight over the generated output.

generating-flow