评估框架技能¶
作为 Claude Code 会话的官方评估框架,实现评估驱动开发(EDD)原则。
激活时机¶
- 在 AI 辅助工作流中设置评估驱动开发(EDD)时
- 定义 Claude Code 任务完成的通过/不通过标准时
- 使用 pass@k 指标衡量智能体可靠性时
- 为提示词或智能体变更创建回归测试套件时
- 在不同模型版本之间对智能体性能进行基准测试时
理念¶
评估驱动开发将评估视为“AI 开发的单元测试”: - 在实现之前定义预期行为 - 在开发过程中持续运行评估 - 每次变更时跟踪回归 - 使用 pass@k 指标衡量可靠性
评估类型¶
功能评估¶
测试 Claude 是否能完成以前无法完成的事情:
[CAPABILITY EVAL: feature-name]
Task: Description of what Claude should accomplish
Success Criteria:
- [ ] Criterion 1
- [ ] Criterion 2
- [ ] Criterion 3
Expected Output: Description of expected result
回归评估¶
确认变更不会破坏现有功能:
[REGRESSION EVAL: feature-name]
Baseline: SHA or checkpoint name
Tests:
- existing-test-1: PASS/FAIL
- existing-test-2: PASS/FAIL
- existing-test-3: PASS/FAIL
Result: X/Y passed (previously Y/Y)
评分器类型¶
1. 基于代码的评分器¶
使用代码进行确定性检查:
# Check if file contains expected pattern
grep -q "export function handleAuth" src/auth.ts && echo "PASS" || echo "FAIL"
# Check if tests pass
npm test -- --testPathPattern="auth" && echo "PASS" || echo "FAIL"
# Check if build succeeds
npm run build && echo "PASS" || echo "FAIL"
2. 基于模型的评分器¶
使用 Claude 评估开放式输出:
[MODEL GRADER PROMPT]
Evaluate the following code change:
1. Does it solve the stated problem?
2. Is it well-structured?
3. Are edge cases handled?
4. Is error handling appropriate?
Score: 1-5 (1=poor, 5=excellent)
Reasoning: [explanation]
3. 人工评分器¶
人工审查标志:
[HUMAN REVIEW REQUIRED]
Change: Description of what changed
Reason: Why human review is needed
Risk Level: LOW/MEDIUM/HIGH
指标¶
pass@k¶
“k 次尝试中至少成功一次” - pass@1:首次尝试成功率 - pass@3:3 次尝试内成功 - 常见目标:pass@3 > 90%
pass^k¶
“k 次尝试全部成功” - 对可靠性提出更高标准 - pass^3:连续 3 次成功 - 用于关键路径
评估工作流¶
1. 定义(编码前)¶
## EVAL DEFINITION: feature-xyz
### Capability Evals
1. Can create new user account
2. Can validate email format
3. Can hash password securely
### Regression Evals
1. Existing login still works
2. Session management unchanged
3. Logout flow intact
### Success Metrics
- pass@3 > 90% for capability evals
- pass^3 = 100% for regression evals
2. 实现¶
编写代码以通过已定义的评估。
3. 评估¶
# Run capability evals
[Run each capability eval, record PASS/FAIL]
# Run regression evals
npm test -- --testPathPattern="existing"
# Generate report
4. 报告¶
EVAL REPORT: feature-xyz
========================
Capability Evals:
create-user: PASS (pass@1)
validate-email: PASS (pass@2)
hash-password: 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
集成模式¶
实现前¶
在.claude/evals/feature-name.md 中创建评估定义文件
实现中¶
运行当前评估并报告状态实现后¶
生成完整评估报告评估仓库¶
在项目中存储评估:
最佳实践¶
- 在编码前定义评估 - 强制对成功标准进行清晰思考
- 频繁运行评估 - 尽早发现回归
- 随时间跟踪 pass@k - 监控可靠性趋势
- 尽可能使用代码评分器 - 确定性 > 概率性
- 安全事项需人工审查 - 不要完全自动化安全检查
- 保持评估快速 - 缓慢的评估不会被执行
- 将评估与代码一起进行版本管理 - 评估是一等产出物
示例:添加认证¶
## EVAL: add-authentication
### Phase 1: 정의 (10분)
Capability Evals:
- [ ] User can register with email/password
- [ ] User can login with valid credentials
- [ ] Invalid credentials rejected with proper error
- [ ] Sessions persist across page reloads
- [ ] Logout clears session
Regression Evals:
- [ ] Public routes still accessible
- [ ] API responses unchanged
- [ ] Database schema compatible
### Phase 2: 구현 (가변)
[Write code]
### Phase 3: 평가
Run: /eval check add-authentication
### Phase 4: 보고서
EVAL REPORT: add-authentication
==============================
Capability: 5/5 passed (pass@3: 100%)
Regression: 3/3 passed (pass^3: 100%)
Status: SHIP IT
产品评估(v1.8)¶
当仅靠单元测试无法捕捉行为质量时,请使用产品评估。
评分器类型¶
- 代码评分器(确定性断言)
- 规则评分器(正则表达式/模式约束)
- 模型评分器(LLM 评审员评分细则)
- 人工评分器(对模糊输出进行人工判定)
pass@k 指南¶
pass@1:直接可靠性pass@3:受控重试下的实用可靠性pass^3:稳定性测试(3 次全部通过)
推荐阈值: - 功能评估:pass@3 >= 0.90 - 回归评估:发布关键路径 pass^3 = 1.00
评估反模式¶
- 针对已知评估示例过度拟合提示词
- 仅测量正常路径输出
- 追求通过率而忽略成本和延迟波动
- 在发布门禁中允许不稳定的评分器
最小评估产出物布局¶
.claude/evals/<feature>.md定义.claude/evals/<feature>.log运行历史docs/releases/<version>/eval-summary.md发布快照
本页原文 Markdown:在 AtomGit 查看·内容源自开源项目 ev/ECC