How does the skill handle implementation failures?

If the judge agent provides a score below 4.0, the orchestrator parses the specific issues identified, feeds them back to the implementation agent, and triggers a retry loop (up to 2 times).

Which LLM models are used for implementation?

The skill provides a selection guide: Opus for complex logic, Sonnet for high-volume pattern updates, and Haiku for trivial mechanical operations.

What is the do-and-judge pattern?

It is a workflow where the orchestrating agent delegates a task to an implementation sub-agent and then uses a separate, independent agent to evaluate the results, ensuring an objective quality gate.

Why does this skill use sub-agents instead of the main Claude session?

Sub-agents operate with a fresh context window, which reduces distractions from the main conversation and allows for more focused, specialized reasoning on specific tasks.

What scoring system does the judge use?

The judge uses a weighted 5-point scale evaluating Correctness (35%), Quality (25%), Completeness (25%), and Pattern Adherence (15%), requiring a score of 4 or higher to pass.

Task Orchestrator & Quality Judge

Name: Task Orchestrator & Quality Judge
Author: NeoLabHQ

byNeoLabHQ

•

542

•

開発者ツール

Executes coding tasks using a specialized implementation sub-agent and an independent LLM judge to ensure high-quality, verified results.

This skill implements a robust 'do-and-judge' pattern for Claude Code, transforming the primary agent into an orchestrator that manages a rigorous quality control workflow. Instead of performing tasks directly, it dispatches specialized implementation sub-agents with clean context windows and subsequently verifies their work using an independent judge sub-agent. By enforcing a weighted scoring system and an automated retry loop based on specific feedback, this skill catches blind spots, ensures adherence to codebase patterns, and prevents low-quality code from being committed, making it essential for complex refactoring and critical feature development.

主な機能

01LLM-as-a-judge verification using weighted scoring (Correctness, Quality, Completeness, Patterns)

02Multi-agent orchestration with specialized implementation and judge roles

03542 GitHub stars

04Automated feedback loop with up to two retries for failing implementations

05Structured 'Self-Critique' verification mandatory for all implementation tasks

06Zero-shot Chain-of-Thought (CoT) prompting for enhanced sub-agent reasoning

ユースケース

01Implementing new features that require strict adherence to existing architectural conventions

02Critical code refactoring where logic errors or pattern deviations are high-risk

03Automating complex, multi-file updates that need independent verification before completion

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add neolabhq/context-engineering-kit do-and-judge

For use in Claude.ai and ChatGPT

Download Skill