What problem does it solve? Authoring evaluation cases for prompts and skills often lacks grounded, trustworthy source material, leading to guessed or low-quality eval rows. This Skill researches public datasets, benchmarks, and documentation with a primary-source-first approach and delivers a standalone research brief before any eval synthesis begins. ## Core Features & Use Cases - Primary-source-first research plan: Builds a query plan tied to the prompt's task, placeholders, and scoring rule before searching, and asks for missing constraints instead of guessing. - Ranked source shortlist with approval bar: Scores candidates on authority, provenance, annotation quality, licensing, version stability, and contamination risk, and records explicit rejection reasons for weak sources. - Eval-authoring mapping notes: Maps approved source fields into prompt rows, expected outputs, optional files, and objective assertions, with a stop recommendation when no source clears the bar. - Use Case: Given a support-ticket classification prompt, the Skill derives the target evals/evals.json layout, shortlists official intent-classification datasets with license notes, rejects derivative mirrors, and hands off mapping guidance for downstream eval authoring. ## Quick Start Research grounded public datasets and benchmarks for the prompt at skills/researcher-research/evals/files/support_intent_prompt.md and return a ranked primary-source shortlist with licensing notes and eval mapping guidance.
