Token导航 LogoToken导航TokenDH.com
开发需要联网github未标认证来源可访问许可证需确认审计通过

create-judgecreate judge 命令行

Agent Skill

create-judge 用于处理 GitHub 仓库、Issue、Pull Request 和代码协作信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要围绕仓库状态、代码变更或协作事项进行整理时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

总安装

212

周安装

9

GitHub Stars

公开资料未说明

下载量

74
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:create-judge(create judge 命令行)
来源仓库:https://github.com/zeroeval/zeroeval-skills
仓库路径:skills/create-judge
安装命令:
npx skills add https://github.com/zeroeval/zeroeval-skills --skill create-judge
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/zeroeval/zeroeval-skills --skill create-judge

简介

用于处理 GitHub 仓库、Issue、Pull Request 和代码协作信息,适合在需要整理仓库状态或协作事项时使用。

  • 适用于 Codex、Claude、Cursor、Gemini CLI 等宿主环境,支持评估模型输出的自动化裁判设计。
  • 通过命令行安装,需结合来源仓库文档核验具体用法,确保裁判模板与评估目标对齐。
  • 安装前建议确认权限范围,避免触发未授权的联网或文件操作。
  • create-judge 属于开发类 Skill,可作为该场景下的辅助能力补充。

SKILL.md

Create Judge

Guide users through designing and creating an automated judge that evaluates LLM outputs in ZeroEval.

When To Use

  • Creating a new judge from scratch for any evaluation goal.
  • Deciding between binary (pass/fail) and scored (numeric rubric) evaluation.
  • Writing judge templates (the evaluation prompt the judge model runs).
  • Designing structured criteria for multi-dimensional scored judges.
  • Linking a judge to a specific prompt for automatic feedback.
  • Troubleshooting judges that aren't producing expected evaluations.

Execution Sequence

Follow these steps in order. Load reference files only when needed for the current step.

Step 1: Understand the Evaluation Goal

Ask the user what they want to evaluate and what "good" looks like. Key questions:

  • What kind of LLM output are you evaluating? (chat response, extracted data, generated code, etc.)
  • What does a good output look like? What makes an output bad?
  • Is this a hard pass/fail check or a quality spectrum?
  • Should the judge run on all LLM spans or only spans from a specific prompt?

Step 2: Choose Evaluation Type

Based on the user's answers:

  • Binary -- use when the criterion is unambiguous (safety, format compliance, factual correctness). The judge returns pass or fail.
  • Scored -- use when quality is a spectrum (helpfulness, tone, completeness). The judge returns a numeric score.
  • Scored with criteria -- use when multiple quality dimensions matter. The judge scores each criterion independently plus an overall score.

Read references/evaluation-patterns.md for concrete template examples by use case.

Step 3: Write the Judge Template

The template is the evaluation prompt the judge model uses. Read references/judge-creation-guide.md for template conventions and writing guidance.

Key rules:

  • Keep templates focused on one evaluation goal.
  • Describe quality criteria, not output format.
  • For scored judges, define what each score level means.
  • For multi-criteria judges, list each criterion with a clear description.

Step 4: Configure the Judge

Read references/defaults-and-api-reference.md for all configurable fields.

Decide:

  • Score range (scored only): default 0-10 but can be customized.
  • Pass threshold (scored only): defaults to 70% of scale.
  • Sample rate: what percentage of spans to evaluate (1.0 = all).
  • Temperature: default 0.0 (deterministic). Keep at 0.0 for evaluation consistency.
  • Target prompt: link to a specific prompt to scope evaluations and enable auto-feedback.

Step 5: Create the Judge

Two options:

Dashboard: Navigate to the project, open Judges, click "Create Judge", fill in the fields.

API: POST /signals/projects/{project_id}/automations with the payload from the defaults reference.

Step 6: Link to a Prompt (Optional)

If the judge should only evaluate a specific prompt's outputs and feed back into prompt optimization:

POST /signals/projects/{project_id}/prompts/{prompt_id}/judges/{automation_id}/link

Step 7: Validate

After creating the judge, confirm it's working:

  • Check the Judges section in the dashboard for new evaluations.
  • Verify the evaluation count is increasing as new spans arrive.
  • Review a few evaluations to confirm the judge is scoring as expected.
  • If the judge is linked to a prompt, verify feedback appears in the prompt's feedback tab.

Key Principles

  • One goal per judge: a judge that tries to evaluate safety AND helpfulness will do neither well. Create separate judges.
  • Criteria over format: the template should describe what "good" means, not what the output should look like.
  • Start binary, graduate to scored: binary judges are easier to validate. Add scored judges once you understand the quality dimensions.
  • Link to prompts: prompt-linked judges enable the optimization loop. Always link when using ze.prompt.
  • Low temperature: keep at 0.0 for reproducible evaluations.

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

34.53%
按下载量换算26

Claude

32.36%
按下载量换算24

Cursor

19.83%
按下载量换算15

Gemini CLI

9.79%
按下载量换算7

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

需要联网

该 Skill 可能需要联网访问来源站点、仓库或外部 API;具体网络访问范围需要结合源码和 README 复核。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills