Refines codebase context retrieval for Claude subagents through a multi-phase evaluation and refinement loop.
The Iterative Retrieval skill addresses the 'context problem' in multi-agent workflows where subagents often lack sufficient information or exceed token limits. By implementing a sophisticated four-phase loop—Dispatch, Evaluate, Refine, and Loop—this skill enables Claude to progressively discover relevant files, learn codebase-specific terminology, and fill information gaps without overwhelming the context window. It is particularly effective for complex bug fixes and feature implementations in large, unfamiliar repositories where a single search query is often insufficient to capture all necessary dependencies.
主な機能
01Sophisticated relevance scoring (0.0 to 1.0) to filter out noise from search results
024-phase refinement loop (Dispatch, Evaluate, Refine, Loop) for deep codebase exploration
03Explicit gap identification to drive targeted subsequent search cycles
04Token usage optimization by stopping once 'good enough' context is achieved
050 GitHub stars
06Automatic terminology discovery to adapt search queries to project-specific naming conventions
ユースケース
01Implementing new features that require understanding existing architectural patterns and middleware
02Resolving complex bugs in large repositories where relevant code is spread across multiple layers
03Orchestrating subagents to prevent 'context too large' failures during complex task execution
