What is the subagent context problem?

It occurs when secondary agents are spawned with limited information, leading to errors because they don't know which files contain relevant code or project-specific terminology.

Does this replace standard search tools?

No, it acts as a management layer over search tools, using the results of one search to intelligently inform and narrow down the next.

Can this skill help with unfamiliar codebases?

Yes, it is specifically designed to learn codebase terminology (like 'throttle' vs 'rate-limit') during the first cycle to improve subsequent searches.

How does the evaluation stage work?

Files are scored on a scale of 0 to 1. Files with a score above 0.7 are kept, while those below 0.2 are excluded from future refinement cycles.

How many cycles should the Iterative Retrieval pattern run?

The pattern typically recommends a maximum of 3 cycles. This is usually enough to gather 'sufficient' context without exceeding time or token limits.

Iterative Context Retrieval

Name: Iterative Context Retrieval
Author: saar120

bysaar120

0•

生产力与工作流

Refines AI subagent context retrieval through a progressive four-stage loop to solve the limited context problem in complex workflows.

This skill implements a sophisticated 'Iterative Retrieval' pattern designed for complex agentic workflows where subagents often lack the necessary project context to begin work effectively. By utilizing a 4-stage cycle—Dispatch, Evaluate, Refine, and Loop—it progressively narrows down search criteria based on discovered code patterns and terminology. This prevents context window overflow from sending too much data while avoiding errors caused by missing information, ensuring Claude possesses the exact files and definitions needed for high-accuracy bug fixes and feature implementations within large codebases.

主要功能

01Four-stage refinement loop (Dispatch, Evaluate, Refine, Loop)

02Relevance scoring system (0.0 to 1.0) for context quality

03Configurable cycle limits to balance retrieval depth with performance

04Dynamic search criteria expansion and exclusion based on findings

050 GitHub stars

06Automated terminology discovery to adapt to codebase-specific naming

使用场景

01Providing precise context for subagents in multi-agent orchestration workflows

02Gathering middleware and architectural patterns for new feature implementation

03Locating specific logic and dependencies for bug fixes in large repositories

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add saar120/skillgate-registry iterative-retrieval

For use in Claude.ai and ChatGPT

主要功能

01Four-stage refinement loop (Dispatch, Evaluate, Refine, Loop)

02Relevance scoring system (0.0 to 1.0) for context quality

03Configurable cycle limits to balance retrieval depth with performance

04Dynamic search criteria expansion and exclusion based on findings

050 GitHub stars

06Automated terminology discovery to adapt to codebase-specific naming

使用场景

01Providing precise context for subagents in multi-agent orchestration workflows

02Gathering middleware and architectural patterns for new feature implementation

03Locating specific logic and dependencies for bug fixes in large repositories

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add saar120/skillgate-registry iterative-retrieval

For use in Claude.ai and ChatGPT