The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill relies on the agent generating and executing sophisticated Python scripts to handle data extraction, transformation, and asset management. Detailed code templates are provided in references/migrate-data/extraction-recipes.md, references/migrate-maps/arcade-translation.md, and references/migrate-maps/marker-upload.md, which utilize libraries like geopandas, pandas, PIL, and arcgis for runtime processing.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted metadata and content from external ArcGIS Portals, such as item titles, descriptions, and Arcade expressions, as described in references/discover.md and references/migrate-maps/app-absorption.md. This data is subsequently used to construct SQL queries and CARTO Builder configurations, creating a surface where malicious instructions within ArcGIS content could potentially manipulate the agent's logic.
Ingestion points: Fetches data via the ArcGIS REST API (/sharing/rest/content/items/.../data) during the discovery and map migration phases.
Boundary markers: No specific delimiters or warnings are used to instruct the LLM to ignore potentially malicious commands within the ingested ArcGIS metadata.
Capability inventory: The skill can execute SQL queries in the user's data warehouse, create and modify maps via the CARTO platform, and manage files on the local disk.
Sanitization: Although the skill uses sqlglot to validate translated SQL and applies slugification to identifiers, these measures do not protect the LLM from instruction-following attacks based on the ingested content.
[COMMAND_EXECUTION]: The skill frequently invokes shell commands, including the carto CLI, curl, and jq, to facilitate API communication and manage the migration manifest, as documented throughout the reference materials.

carto-arcgis-migration