The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from Power BI models (such as table names, column names, descriptions, and DAX code) which could potentially contain malicious instructions intended to influence the agent's behavior during refactoring.
Ingestion points: The skill reads model metadata and data via get_model_overview, get_object_details, and execute_dax_query.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within the ingested metadata.
Capability inventory: The skill has the capability to write changes to the model via execute_csharp_script and interact with reports/services using fabric-cli and pbir-cli.
Sanitization: There is no mention of sanitizing or validating ingested strings before they are incorporated into generated scripts or descriptions.
[DYNAMIC_EXECUTION]: The skill's primary method of action is generating and executing C# scripts at runtime to modify the semantic model.
Evidence: The workflow in SKILL.md explicitly instructs the agent to use execute_csharp_script for applying changes across multiple passes, including renaming objects, setting properties, and formatting DAX.
[COMMAND_EXECUTION]: The skill invokes several external command-line tools and other agent skills to perform administrative and deployment tasks.
Evidence: References to the te CLI, fabric-cli, pbir-cli, and lineage-analysis skills indicate the orchestration of external processes to manage Power BI Desktop Project files (PBIP) and Fabric service resources.

organize-semantic-model