Does this skill increase my token usage?

It is designed to optimize token usage by ensuring that only the most relevant files are included in the final task context, preventing the 'send everything' approach that wastes tokens.

Can this skill be used for languages other than C++?

Yes, while it was developed as part of a C++ HPC collection, the pattern of iterative search, evaluation, and refinement is language-agnostic and works for any codebase.

How many cycles does the retrieval loop perform?

The skill is configured to run for a maximum of three cycles, stopping earlier if sufficient high-relevance context (relevance score >= 0.7) is found.

How does the skill determine if a file is relevant?

It uses a scoring system from 0 to 1.0; files implementing target functionality score 0.8+, while those with related patterns score 0.5-0.7. Only files with 0.7 or higher are typically retained.

What is the 'context problem' this skill solves?

It solves the issue where AI subagents either have too much irrelevant information, which exceeds token limits, or too little information to understand the codebase accurately.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: ysyecust

byysyecust

•

生产力与工作流

Optimizes AI subagent performance by progressively refining codebase context through a multi-phase search and evaluation loop.

The Iterative Retrieval skill solves the common 'context problem' in multi-agent workflows where subagents struggle with either too much irrelevant data or insufficient information to complete tasks. It implements a structured four-phase loop—Dispatch, Evaluate, Refine, and Loop—that allows Claude to dynamically discover relevant files, learn project-specific terminology, and identify implementation patterns. By scoring file relevance on a 0-1 scale and stopping after three cycles, this skill ensures that Claude maintains a high-signal, low-noise context window, making it particularly effective for large-scale C++ HPC projects and complex debugging scenarios.

主要功能

01Cycle-limited execution to prevent infinite loops and token waste

02Intelligent exclusion of noise like test files and benchmarks

03Automatic terminology discovery to adapt to unique project naming

042 GitHub stars

05Dynamic relevance scoring to filter high-impact source files

06Four-phase refinement loop for precise context gathering

使用场景

01Gathering necessary patterns for adding new features to existing architectures

02Pinpointing memory leaks or bugs across large, unfamiliar codebases

03Synchronizing multi-agent workflows with consistent domain-specific context

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add ysyecust/everything-claude-code iterative-retrieval

For use in Claude.ai and ChatGPT

Download Skill