Eval Harness Skill
SkillMonitoring & opsFormal evaluation framework implementing eval-driven development (EDD) principles. Used to define pass/fail criteria, measure pass@k metrics, and create regression test suites.
Use Eval Harness Skill in Claude, ChatGPT or Ahel Desktop
Free. Sign in, add Eval Harness Skill and connect your AI. About a minute.
Also: Claude Code · Cursor · Codex
Then ask your AI: use the Eval Harness Skill
Details
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; ahel provides instructions and does not run this skill.
No other account needed.
Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.
What this skill tells your AI
The instructions your AI receives, as published by hashgraph-online/awesome-codex-plugins in plugins/Colin4k1024/tsp/skills/eval-harness/SKILL.md and read by ahel’s review.
一个用于 Claude Code 会话的正式评估框架,实现 eval-driven development (EDD) 原则。
何时激活
- 为 AI 辅助工作流设置 EDD
- 定义 Claude Code 任务完成的 pass/fail 标准
- 使用 pass@k 指标测量 agent 可靠性
- 为 prompt 或 agent 更改创建回归测试套件
- 跨模型版本基准测试 agent 性能
理念
Eval-Driven Development 将评估视为"AI 开发的单元测试":
- 在实现前定义预期行为
- 在开发期间持续运行评估
- 用每次变更追踪回归
- 使用 pass@k 指标测量可靠性
评估类型
Capability Evals
测试 Claude 能否做以前不能做的事:
[CAPABILITY EVAL: points-calculation]
Task: 计算用户积分并确定等级
Success Criteria:
- [ ] 积分正确累加
- [ ] 等级边界正确
- [ ] 权益解锁逻辑正确
Expected Output: 用户总积分 = 1500,等级 = L3
Regression Evals
确保变更不破坏现有功能:
[REGRESSION EVAL: login-flow]
Baseline: sha-abc123
Tests:
- existing-login: PASS
- session-management: PASS
- logout-flow: PASS
Result: 3/3 passed (previously 3/3)
Grader 类型
1. Code-Based Grader
使用代码的确定性检查:
# 检查文件是否包含预期模式
grep -q "export function handlePoints" src/points.ts && echo "PASS" || echo "FAIL"
# 检查测试是否通过
npm test -- --testPathPattern="points" && echo "PASS" || echo "FAIL"
2. Model-Based Grader
使用 Claude 评估开放式输出:
[MODEL GRADER PROMPT]
评估以下代码变更:
1. 它是否解决了陈述的问题?
2. 结构是否良好?
3. 边界情况是否处理?
4. 错误处理是否适当?
Score: 1-5 (1=差, 5=优秀)
Reasoning: [解释]
3. Human Grader
标记为手动审查:
[HUMAN REVIEW REQUIRED]
Change: 描述变更内容
Reason: 为什么需要人工审查
Risk Level: LOW/MEDIUM/HIGH
指标
pass@k
"k 次尝试中至少一次成功"
- pass@1: 首次尝试成功率
- pass@3: 3 次内成功
- 典型目标: pass@3 > 90%
pass^k
"所有 k 次试验都成功"
- 更高可靠性标准
- 用于关键路径
评估工作流
1. 定义(编码前)
## EVAL DEFINITION: points-system
### Capability Evals
1. 可以计算用户积分
2. 可以确定用户等级
3. 可以解锁权益
### Regression Evals
1. 现有登录仍然有效
2. 会话管理未改变
3. 登出流程完整
### Success Metrics
- pass@3 > 90% for capability evals
- pass^3 = 100% for regression evals
2. 实现
编写代码通过定义的评估。
3. 评估
# 运行 capability evals
[Run each capability eval, record PASS/FAIL]
# 运行 regression evals
npm test -- --testPathPattern="existing"
# 生成报告
4. 报告
EVAL REPORT: points-system
==========================
Capability Evals:
calculate-points: PASS (pass@1)
determine-level: PASS (pass@2)
unlock-benefits: PASS (pass@1)
Overall: 3/3 passed
Regression Evals:
login-flow: PASS
session-mgmt: PASS
logout-flow: PASS
Overall: 3/3 passed
Metrics:
pass@1: 67% (2/3)
pass@3: 100% (3/3)
Status: READY FOR REVIEW
评估存储
在项目中存储评估:
.claude/
evals/
points-system.md # 评估定义
points-system.log # 评估运行历史
baseline.json # 回归基线
最佳实践
- 编码前定义评估 - 强制清晰思考成功标准
- 频繁运行评估 - 尽早捕获回归
- 追踪 pass@k 趋势 - 监控可靠性趋势
- 尽可能使用 code graders - 确定性 > 概率性
- 人工审查安全 - 永远不完全自动化安全检查
- 保持评估快速 - 慢评估不会被运行
- 版本评估与代码 - 评估是一级 artifacts
与 Error Experience Library 的关系
- Error Experience Library 记录已解决的错误
- Eval Harness 验证新能力是否正确实现
- 两者都可用于回归测试
Signals
- GitHub stars
- 1k
- Forks
- 316
- Last commit
- Oct 2026
Advanced
- Item type
- skill
- Key
eval-harness-hashgraph-online- Source
- github.com/hashgraph-online/awesome-codex-plugins
github.com/hashgraph-online/awesome-codex-plugins
Related picks
Skill · mattpocock
The pick for TypeScripttypescript-pro
Skill · jeffallan
The pick for TypeScriptlark-markdown
Skill · larksuite
The pick for Markdownmarkdown-formatter
Skill · nvidia
The pick for Markdowninternal-comms
Skill · anthropics
More in Monitoring & opsagent-eval
Skill · affaan-m
More in Monitoring & ops