What problem does it solve? Writing answer templates for LLM-as-judge evaluation is error-prone: wrong base classes, mismatched ground truth types, vague field descriptions, and misused verification primitives all cause silent failures. This Skill guides the authoring of karenina answer templates (BaseAnswer subclasses with VerifiedField) so extracted values are verified correctly against ground truth. ## Core Features & Use Cases - Interactive six-step authoring workflow: from domain understanding through field identification, structure proposal, code generation, validation, and user confirmation. - Verification primitive guidance: covers BooleanMatch, ExactMatch, SetContainment, NumericGraded, LiteralMatch, trace primitives, and composition nodes (AllOf, AnyOf, AtLeastN) with a full primitive catalog reference. - Template validation script: runs structural checks on generated templates and reports errors for iterative fixing. - Use Case: You need to evaluate whether LLM responses correctly identify a drug's modality. The Skill walks you through defining a boolean VerifiedField with a judge-facing description, correct ground truth, and BooleanMatch, then validates the template before pipeline use. ## Quick Start Ask the assistant to create a karenina answer template for your evaluation question, describing the domain and what a correct answer looks like.
