What is the sub-agent context problem?

Sub-agents often start tasks without knowing which files are relevant or what specific terminology is used in a codebase, leading to errors or hallucinations.

Can I limit the number of retrieval cycles?

Yes, the skill is designed to run for a maximum of 3 cycles by default, ensuring a balance between thoroughness and execution speed.

Does this skill help save on token costs?

Yes, by identifying and excluding low-relevance files early, it prevents the context window from being filled with unnecessary data.

How does the 4-stage loop improve AI results?

It starts broad, evaluates file relevance, learns project-specific terms from the results, and narrows down the search to find the most critical implementation files.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: xu-xiang

byxu-xiang

•

Productivity & Workflow

Optimizes AI agent accuracy by progressively refining file retrieval through a specialized 4-stage discovery loop.

The Iterative Retrieval skill solves the 'sub-agent context problem' where AI agents lack initial knowledge of a codebase's structure and terminology. By implementing a cyclical process of dispatching queries, evaluating relevance, refining search criteria, and looping, this skill ensures that Claude identifies the most pertinent files while avoiding context window overflow. It is an essential pattern for working on large-scale repositories where finding the right starting point is half the battle, allowing sub-agents to operate with high-precision context.

Key Features

014-stage refinement cycle: Dispatch, Evaluate, Refine, and Loop

02Context gap identification to drive targeted secondary searches

03Intelligent file exclusion to minimize context noise and token usage

04Relevance scoring system on a 0.0 to 1.0 scale

05Automated discovery of project-specific terminology and patterns

0661 GitHub stars

Use Cases

01Identifying architectural patterns for consistent new feature implementation

02Reducing LLM hallucination by ensuring agents have exactly the right code context

03Locating bug sources in large, unfamiliar codebases where keywords are unknown

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add xu-xiang/everything-claude-code-zh iterative-retrieval

For use in Claude.ai and ChatGPT

Download Skill