How does the scoring system work?

It ranks files from 0 to 1 based on relevance. Files scoring 0.7 or higher are typically kept, while low-scoring files are used to refine search exclusions and identify terminology gaps.

How many cycles should the retrieval loop run?

The pattern recommends a maximum of 3 cycles, as this usually provides a 'good enough' context of high-relevance files without wasting processing time.

What is the primary benefit of the Iterative Retrieval pattern?

It ensures AI agents have the exact context they need by refining searches based on what they learn in each step, preventing context overflow while avoiding missing information.

Can this skill handle large codebases?

Yes, it is specifically designed for large codebases where sending every file is impossible and guessing the right files is often inaccurate.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: x-cmd

byx-cmd

•

Developer Tools

Optimizes AI agent performance by progressively refining codebase searches through an intelligent multi-phase evaluation and feedback loop.

The Iterative Context Retrieval skill implements a robust 4-phase pattern—Dispatch, Evaluate, Refine, and Loop—to solve the context discovery problem in multi-agent AI workflows. By starting with broad queries and using automated relevance scoring to identify terminology gaps and specific file paths, it ensures agents receive high-quality, high-relevance code context without exceeding token limits. This approach is particularly effective for complex bug fixes and feature implementations in large codebases where the necessary project-specific terminology or architecture is not immediately known.

Key Features

01Four-phase iterative loop for progressive context discovery and refinement

02Explicit gap identification to drive precise search refinements in subsequent cycles

03Intelligent relevance scoring system to filter high-impact files (0.0 to 1.0)

04Automated terminology discovery to adapt to project-specific naming conventions

05Configurable cycle limits to balance context depth with operational efficiency

069 GitHub stars

Use Cases

01Onboarding AI agents to projects with non-standard directory structures

02Gathering implementation patterns and boilerplate for new feature development

03Locating relevant logic for bug fixes in unfamiliar or massive codebases

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add x-cmd/skill iterative-retrieval

For use in Claude.ai and ChatGPT

Download Skill