worker-judge

Coordinate an independent worker-and-judge review loop until task output is approved.

5|Updated Feb 8, 2026
One-click install
npx skills add https://github.com/tiancaiamao/ai --skill worker-judge
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: worker-judge
Source: https://github.com/tiancaiamao/ai/tree/main/skills/worker-judge
Command: npx skills add https://github.com/tiancaiamao/ai --skill worker-judge

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Use a separate judge agent to verify another agent's work, reducing self-review bias and catching issues that a single pass can miss.

Core Features & Use Cases

  • Independent review loop: Keep a worker agent alive while spawning a fresh judge each round for unbiased evaluation.
  • Iterative refinement: Feed judge feedback back to the worker until the output is approved or the maximum round limit is reached.
  • Quality-gated tasks: Best for planning, coding, and review work that has clear acceptance criteria and needs reliable verification.

Quick Start

Ask the skill to run a worker-judge loop on your task and keep iterating until the output is approved.

Frequently Asked Questions about worker-judge

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I use a multi-agent review loop to reduce self-review bias in code generation?

Iterative refinement for quality assurance works by feeding explicit judge feedback back to a persistent worker agent, repeating the evaluation across multiple rounds until the output is approved or the round limit is reached.

How do I set up an independent worker and judge agent workflow for task verification?

You need a task with clear acceptance criteria, a persistent worker agent state to apply changes, and the ability to spawn fresh judge agents that deliver explicit approve-or-changes verdicts across multiple review rounds.

What is the best way to automate quality assurance checks for planning and implementation tasks?

The worker-judge approach differs from single-pass verification by maintaining a persistent worker state while spawning a fresh judge agent each round, preventing reviewer fatigue and eliminating self-review bias through independent evaluation.

What happens if the judge agent never approves the worker's output during iterative refinement?

A limitation of the worker-judge loop is that it requires a controlled cleanup mechanism for spawned child agents, and tasks without clear acceptance criteria may cause the judge to iterate indefinitely until the maximum round limit is reached.