What is Iterative Retrieval in Claude Code?

It is a pattern used to progressively refine the context provided to an AI agent through multiple search and evaluation cycles, ensuring the most relevant code is used for the task.

Why is this better than sending all files at once?

Sending all files often exceeds token limits or leads to model hallucinations; iterative retrieval focuses the AI on high-signal data, much like a developer manually exploring a project.

How does this skill prevent context window limits?

By scoring files for relevance and excluding low-scoring content, it ensures only the most critical code is included in the agent's active memory, preventing token overflow.

What are the four stages of this pattern?

The cycle consists of Dispatch (initial broad query), Evaluate (scoring file relevance), Refine (updating search criteria based on findings), and Loop (repeating the process until enough context is found).

Can I limit the number of search cycles?

Yes, the pattern is designed to typically run for a maximum of 3 cycles to balance thoroughness with execution speed and cost-efficiency.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: unju-ai

byunju-ai

0•

生産性とワークフロー

Refines agent context retrieval through a progressive four-stage loop to ensure high-relevance code discovery without exceeding token limits.

Iterative Retrieval is a specialized pattern designed to solve the 'context problem' in multi-agent workflows where subagents often lack sufficient codebase knowledge to perform tasks accurately. This skill implements a systematic Dispatch-Evaluate-Refine-Loop cycle that allows Claude to start with broad queries, evaluate the relevance of discovered files, and iteratively narrow down its search based on discovered terminology and patterns. By simulating a human-like discovery process, it ensures the agent has exactly the right information—no more, no less—to solve complex bugs or implement new features without hitting context window constraints or suffering from information overload.

主な機能

01Automatic exclusion of irrelevant paths to optimize token usage

02Dynamic search criteria refinement based on discovered codebase terminology

030 GitHub stars

04Automated Relevance Scoring (0.0 to 1.0) to prioritize critical files

054-Stage Retrieval Loop (Dispatch, Evaluate, Refine, Loop) for deep discovery

06Context gap identification to drive targeted subsequent search cycles

ユースケース

01Tracing complex bugs across unfamiliar directories in large legacy codebases

02Locating distributed middleware and utility functions for API development

03Gathering existing architectural patterns before implementing new features

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add unju-ai/ecc iterative-retrieval

For use in Claude.ai and ChatGPT