The Iterative Retrieval skill implements a structured four-stage pattern—Dispatch, Evaluate, Refine, and Loop—to solve the common 'context problem' in multi-agent workflows. When sub-agents are tasked with complex coding challenges, they often lack the initial knowledge of which files are relevant or what specific terminology the codebase uses. This skill allows Claude to progressively narrow down its search, scoring file relevance on a 0-1 scale and learning project-specific patterns dynamically. By iterating up to three times, it ensures the agent receives precise, high-value context while avoiding the pitfalls of exceeding token limits or operating with insufficient data.
