skills.rest
NeoLabHQ
NeoLabHQOFFICIAL
Explore all skills by this author →
WebsiteGitHub

do-and-judge

Executes tasks via sub-agent implementation with LLM-as-a-judge verification and retry loops.

1.5k|154|Updated Nov 13, 2025
Search Tags:#quality gate#model selection#code verification#retry loop#sub-agent orchestration#LLM-as-a-Judge#Task Automation
One-click install
npx skills add https://github.com/NeoLabHQ/context-engineering-kit --skill do-and-judge
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: do-and-judge
Source: https://github.com/NeoLabHQ/context-engineering-kit/tree/main/plugins/sadd/skills/do-and-judge
Command: npx skills add https://github.com/NeoLabHQ/context-engineering-kit --skill do-and-judge

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Delegating coding or documentation tasks to AI agents often produces unverified output with hidden defects, forcing developers to manually review every change. This Skill orchestrates an implementation sub-agent, a meta-judge that builds evaluation criteria, and an independent judge that scores the result, iterating with feedback until quality thresholds are met.

Core Features & Use Cases

  • Parallel Meta-Judge and Implementation Dispatch: Generates tailored evaluation rubrics while the implementation agent works, then verifies output against those criteria.
  • Automatic Retry with Feedback: Retries up to 3 times with specific judge-identified issues, escalating model tiers (haiku → sonnet → opus) when quality gaps indicate capability limits.
  • Model Selection Policy: Selects cost-appropriate model tiers based on task scope, complexity, and risk, with cross-provider tier mapping for non-Anthropic models.
  • Use Case: Run a refactoring task like "Refactor the UserService class to use dependency injection" and receive verified, judge-scored output without reading or writing any code yourself.

Quick Start

Ask the agent to run /do-and-judge followed by your task description, optionally adding --strict to require a score of 4.0 or higher before passing.

Frequently Asked Questions about do-and-judge

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

FAQPage Schema
How do I run a task with automatic AI verification and retry?▼

Invoke /do-and-judge followed by your task description. The skill dispatches an implementation sub-agent and a meta-judge in parallel, then a judge scores the result against generated rubrics, retrying with feedback up to 3 times until the score reaches 4.0.

How does the skill choose between haiku, sonnet, and opus models?▼

It evaluates task scope, complexity, and risk against a selection rules table. Mechanical single-file edits use haiku, standard code writing uses sonnet, and critical or multi-file contract changes use opus. You can override with the --model flag.

What does the --strict flag do in do-and-judge?▼

The --strict flag disables the Iteration Discretion Rule, so the task passes only when the judge score reaches 4.0 or higher. Without it, scores between 3.0 and 4.0 with only low-priority nitpicks may be accepted to avoid wasting retries.

Can this skill work with non-Anthropic models like GPT or Gemini?▼

Yes, the Cross-Provider Equivalence table maps tiers by capability rather than name. Haiku maps to gemini-flash-lite class models, sonnet to gemini-pro or GPT-5-mini class, and opus to frontier reasoning tiers like GPT-5.5 or deep-think modes.

What happens when the task fails after all retries?▼

After 3 failed retries, the skill stops and escalates to the user with a failure report showing all verdicts, persistent issues, and options: provide more guidance, escalate the model tier, modify requirements, or abort. It never proceeds without a user decision.

Related Skills

Explore More Software Engineering Skills→
fabiomilennials1234-a11yfabiomilennials1234-a11y

agent-prompt-engineer

Translate Conductor briefs into parallel sub-task prompts with acceptance criteria.

Community
Advanced
redsub-captainredsub-captain
2

redsub-ship

Enforce a sequential release pipeline with semantic versioning and user approval.

Community
Advanced
RamonsDkaRamonsDka
1

judgment-day

Run two blind judge sub-agents in parallel to review code or architecture.

Community
Advanced
EntityProcessEntityProcess
15

agentv-eval-orchestrator

Orchestrate AgentV evaluations by simulating LLM responses without API keys.

Official
Intermediate