What problem does it solve? Turning a structured research plan into a trustworthy list of sources is error-prone: agents fabricate URLs, return duplicate or single-domain results, and silently leave sub-questions uncovered. This Skill enforces a fixed output schema, deduplication, diversity caps, and a validation gate so downstream stages receive only real, backend-returned sources. ## Core Features & Use Cases - Backend-agnostic search orchestration: Wraps whatever search tool the host configures (Tavily, Exa, Brave, Serper, Bing, Google CSE, or an MCP search server) behind one fixed source-record schema. - Depth-controlled query budgets: Maps quick, standard, and deep depths to 1-4 query variants per sub-question (keyword, semantic, recency-filtered, contrarian) with target source counts. - Quality gates: Applies URL and snippet-based deduplication, a 30% per-domain diversity cap, relevance scoring with a 0.3 cutoff, and a coverage side channel that flags thin or paywalled sub-questions instead of padding results. - Use Case: As stage 2 of the researcher workflow, feed it the research_plan from plan-research at standard depth and receive 10-20 normalized source records tagged with sub-question IDs, ready for extract-findings to consume. ## Quick Start Gather sources for this research plan at standard depth and return the normalized sources array with coverage notes.
