The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted span data and experiment runs which are interpolated into LLM judge prompts. Ingestion points are identified in ax spans export and ax experiments export command workflows. The skill utilizes structured templates and binary classification labels as boundary markers. Capability inventory includes tool-mediated shell execution via the ax CLI and the execution of evaluation tasks. Sanitization is implemented via recommendations for low-temperature model settings and explicit instructions for the judge model's output format.\n- [DYNAMIC_EXECUTION]: The skill enables the creation of custom Python code evaluators. The agent is instructed to generate Python classes that subclass the vendor-provided CodeEvaluator base class. This dynamic execution surface is controlled through the use of specific class contracts and import requirements defined by the Arize platform, reducing the risk associated with arbitrary code generation.

arize-evaluator