The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill initiates the download of the 'jacobmarks/image-quality-issues' plugin using the download_plugin tool and suggests the installation of the 'umap-learn' package via 'pip'. These resources are associated with the FiftyOne ecosystem and its well-known community contributors.- [REMOTE_CODE_EXECUTION]: The skill enables and executes code from downloaded plugins through the execute_operator tool, allowing for the execution of logic defined in external FiftyOne plugins.- [COMMAND_EXECUTION]: Uses shell commands to start the FiftyOne delegated service in the background and to suggest installation of environment dependencies.- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection (Category 8) as it processes untrusted dataset content such as labels and metadata.
Ingestion points: Data from dataset fields (e.g., '<label_field>.label') and metadata is loaded into the agent's context during inspection and audit phases.
Boundary markers: There are no protective delimiters or instructions to ignore potential commands embedded in the dataset content.
Capability inventory: The skill possesses powerful capabilities including execution of arbitrary operators, file system tagging, and background process management.
Sanitization: There is no evidence of sanitization or validation of dataset values before they are used in curation workflows.- [COMMAND_EXECUTION]: The skill performs dynamic execution by discovering and running FiftyOne operators at runtime based on the environment's current state via the 'list_operators' and 'execute_operator' tool chain.

fiftyone-dataset-curation