Skip to main content
affpapa.org
RU EN
skill CX Codex

agent-evaluation · neolabhq — Claude Code skill: what it does, how to install

Framework for testing agents, prompts, and skills using multi-dimensional scoring rubrics and LLM-as-judge prompts.

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓ 
npx skills add NeoLabHQ/context-engineering-kit --skill agent-evaluation -a claude-code
Open on GitHub

What agent-evaluation does

This skill outlines an evaluation workflow for autonomous coding agents and prompt sets. Agents follow non-deterministic execution paths toward a goal, so step-by-step assertions usually fail. The skill shifts evaluation toward end outcomes using a weighted rubric: instruction following (0.30), output completeness (0.25), tool efficiency (0.20), reasoning quality (0.15), and response coherence (0.10).

It provides an LLM-as-judge template and references research showing that placing justification before the score improves evaluation reliability by 15-25%. Relying strictly on automated judges leaves silent hallucinations undetected on edge cases. Manual spot-checking remains necessary for subtle context drift.

What it helps with

  • evaluate agent outputs against a weighted multi-factor rubric
  • measure prompt and context engineering revisions
  • set up structured LLM-as-judge evaluation prompts
  • catch subtle failures across prolonged conversational contexts

Example prompts for the agent

  • Evaluate the agent output using the weighted rubric with evidence before scores
  • Compare two agent prompts based on instruction following and tool efficiency
  • Generate an LLM-as-judge prompt to assess agent reasoning and output completeness

How to install agent-evaluation

1Skills CLI (Vercel)

npx skills add NeoLabHQ/context-engineering-kit --skill agent-evaluation -a claude-code

2Manually: put the files in place

~/.claude/skills/agent-evaluation/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/agent-evaluation/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about agent-evaluation

How do I install agent-evaluation in Claude Code?

Run npx skills add NeoLabHQ/context-engineering-kit --skill agent-evaluation -a claude-code. Alternatively, copy the skill directly into ~/.claude/skills/agent-evaluation/SKILL.md.

What does agent-evaluation do?

It guides testing of agent workflows and prompt modifications where fixed assertions fall short. The skill defines outcome-focused scoring rubrics and instructions for model-based judges.

What drives agent performance variance according to the skill?

Drawing from BrowseComp research, it highlights three drivers: token usage (80%), tool call count (~10%), and model selection (~5%).

Similar skills

Everything for Testing and QA →

agent-evaluation by other authors comes first.

agent-evaluation sickn33 A skill for comparing agent, prompt or tool changes against versioned cases with explicit verifiers. Useful for reproducing a failure or running a regression pass. CX Codex +1 971installs agent-evaluation mlflow A skill that runs systematic LLM agent evaluation through MLflow: datasets, scorers, tracing. Useful when an agent picks the wrong tool or returns incomplete answers. CX Codex +1 1,100installs improve-codebase-architecture Matt Pocock This skill scans a codebase for architectural friction, then builds a visual HTML report with refactoring candidates. Use it before a deeper architecture review. 1,163,035installs design-taste-frontend Leonxlnx This skill directs the design of public pages and work screens, then checks the result in a browser through screenshots. Use it when an interface feels generic or unfinished. CX Codex +1 678,700installs tdd Matt Pocock Guides the red-green TDD loop: ensures tests are written first, keep them at public interfaces, and avoid coupling to implementation details. 1,141,062installs skill-creator official anthropics Creates and refines skills, then tests them with prompts, evaluations, and performance checks. 414,180installs agent-browser vercel-labs A CLI skill for browser control: it opens pages, fills forms, clicks buttons, and tests web apps. Use it when an agent needs to act inside a website. CX Codex +1 1,200,000installs prototype Matt Pocock Creates throwaway prototypes for checking state logic or exploring UI before committing the decision to the main codebase. 986,201installs

From the same set55

reflect A skill that makes the agent reread its own answer and list the flaws before you accept the work. CX Codex +1 1,400installs prompt-engineering Guides prompt design for agents, subagents, and LLM integrations. Helpful when building production templates, chain-of-thought prompts, or agent hooks. CX Codex +1 1,400installs memorize Transfers insights from agent reflections and critique sessions into CLAUDE.md. Keeps your context file sharp by structuring lessons instead of bloating it with raw logs. CX Codex +1 1,300installs subagent-driven-development Runs an implementation plan through fresh subagents and puts code review between tasks, so context stays clean and issues surface early. CX Codex +1 1,300installs judge-with-debate The skill runs three independent judges who argue over a solution's score until they agree. Useful when a single pass feels too thin for a real verdict. CX Codex +1 1,300installs update-docs Updates project documentation based on uncommitted git changes or the latest commit. Keeps README files, docs/, and JSDoc comments synchronized with your codebase. CX Codex +1 1,300installs root-cause-tracing A skill that walks the agent backward through the call stack until it finds the original trigger, then fixes there instead of at the crash site. CX Codex +1 1,300installs critique The critique skill runs finished work past several AI judges: they review it independently, debate their findings, and agree on a shared list of issues. CX Codex +1 1,300installs

Project sponsors

Companies that keep this analytics open
Gold sponsor
Nutra.Club

Impressions, clicks and code copies — open statistics. We keep the numbers public for the same reason NeBlask itself is public. Become a sponsor