# 方式一：从官方仓库安装
claude plugin add anthropics/claude-plugins-official --plugin skill-creator

# 方式二：从插件市场安装
claude plugin install skill-creator

# 验证安装
claude /skill-creator

graph TD
    A[用户需求] --> B[Create模式<br/>创建Skill]
    B --> C[Eval模式<br/>运行测试用例]
    C --> D{通过率 >= 80%?}
    D -->|否| E[Improve模式<br/>优化描述/指令]
    E --> C
    D -->|是| F[Benchmark模式<br/>基准对比]
    F --> G{Skill优于基线?}
    G -->|否| H[重新设计Skill]
    H --> B
    G -->|是| I[生产发布]

skill-name/
├── SKILL.md              # 被测Skill本体
├── evals.json            # 测试用例定义（你编写）
├── grading.json          # Grader评判结果（自动生成）
├── metrics.json          # 运行时统计（工具调用、步骤数）
├── timing.json           # 墙钟时间与Token消耗
└── history.json          # 版本迭代记录

sequenceDiagram
    participant User as 开发者
    participant SC as skill-creator
    participant Exec as Executor Agent
    participant Grade as Grader Agent
    participant Anal as Analyzer Agent

    User->>SC: /skill-creator eval
    SC->>Exec: 加载evals.json中的prompt
    loop 每个测试用例
        Exec->>Exec: 在隔离context中运行Skill
        Exec->>Grade: 输出结果
        Grade->>Grade: 逐条检查expectations
    end
    Grade->>Anal: 汇总grading.json
    Anal->>Anal: 分析失败模式、Token分布
    Anal->>User: 评测报告 + 改进建议

{
  "skill_name": "deploy-checker",
  "version": "1.0.0",
  "evals": [
    {
      "id": "eval-001",
      "category": "trigger",
      "prompt": "检查一下我的代码是否可以部署到生产环境",
      "expected_output": "应该执行部署就绪检查，分析git状态、测试覆盖率、构建状态",
      "expectations": [
        {
          "type": "contains",
          "value": "部署",
          "description": "输出应包含部署相关术语"
        },
        {
          "type": "tool_used",
          "value": "Bash",
          "description": "应使用Bash工具检查git/测试状态"
        },
        {
          "type": "not_contains",
          "value": "直接部署",
          "description": "不应建议直接部署，而是先做检查"
        }
      ]
    },
    {
      "id": "eval-002",
      "category": "functional",
      "prompt": "我改了auth模块，帮我看看能不能安全上线",
      "expected_output": "应重点检查认证相关的安全问题",
      "expectations": [
        {
          "type": "contains",
          "value": "安全",
          "description": "应提及安全检查"
        },
        {
          "type": "file_read",
          "value_pattern": "auth",
          "description": "应读取auth相关文件"
        },
        {
          "type": "structured_output",
          "description": "输出应有结构化的检查清单"
        }
      ]
    },
    {
      "id": "eval-003",
      "category": "negative",
      "prompt": "帮我写一个Python排序算法",
      "expected_output": "不应触发此Skill",
      "expectations": [
        {
          "type": "not_triggered",
          "description": "此prompt不应触发deploy-checker"
        }
      ]
    }
  ]
}

{
  "id": "trigger-positive-001",
  "category": "trigger",
  "prompt": "准备上线，帮我做最后检查",
  "expectations": [
    {"type": "triggered", "description": "应触发deploy-checker"}
  ]
}

{
  "id": "func-001",
  "category": "functional",
  "prompt": "检查部署就绪状态",
  "expectations": [
    {"type": "tool_used", "value": "Bash", "description": "运行git/test命令"},
    {"type": "contains", "value": "检查清单", "description": "输出检查清单"},
    {"type": "contains", "value": "风险", "description": "识别潜在风险"}
  ]
}

{
  "id": "neg-001",
  "category": "negative",
  "prompt": "这段CSS怎么居中？",
  "expectations": [
    {"type": "not_triggered", "description": "CSS问题不应触发部署检查"}
  ]
}

mkdir -p ~/.claude/skills/deploy-checker
cd ~/.claude/skills/deploy-checker

---
name: deploy-checker
description: >
  检查代码是否准备好部署到生产环境。分析git状态、测试覆盖率、
  构建状态和安全风险。Use when the user asks about deploying,
  going to production, release readiness, or pre-deploy checks.
argument-hint: [environment]
---

# Deploy Readiness Checker

## Overview
评估当前代码变更是否可以安全部署。通过多层检查生成结构化的就绪报告。

## Pre-check: Dynamic Context
!`git status --short 2>/dev/null || echo "Not a git repo"`
!`git log --oneline -5 2>/dev/null || echo "No git history"`
!`git diff --stat HEAD 2>/dev/null || echo "No uncommitted changes"`

## Steps

### Step 1: Change Analysis
1. Use `Read` to examine each modified file from the git status above
2. Classify changes by impact level:
   - **CRITICAL**: auth, payment, database schema, config changes
   - **HIGH**: API endpoints, business logic, data processing
   - **MEDIUM**: UI components, documentation, tests
   - **LOW**: formatting, comments, dependencies

### Step 2: Test Coverage Check
1. Run `Bash` to check test status: look for test commands in package.json/Makefile
2. Identify untested changed files
3. Flag any modified critical files without corresponding test changes

### Step 3: Security Scan
For each CRITICAL/HIGH change, check:
- No hardcoded secrets (API keys, passwords, tokens)
- No SQL injection patterns (string concatenation in queries)
- No exposed debug endpoints
- Auth changes follow security best practices

### Step 4: Dependency Check
1. Check for new dependencies in package.json/requirements.txt
2. Flag any unknown or low-maintenance packages
3. Verify lock files are updated

## Output Format

### Deploy Readiness Report

| Check | Status | Details |
|-------|--------|---------|
| Change Impact | ✅/⚠️/❌ | Summary of change classification |
| Test Coverage | ✅/⚠️/❌ | Coverage of changed files |
| Security | ✅/⚠️/❌ | Security issues found |
| Dependencies | ✅/⚠️/❌ | New dependency assessment |

### Risk Items (if any)
For each risk:
- **File**: `path/to/file:line`
- **Severity**: CRITICAL / HIGH / MEDIUM / LOW
- **Issue**: Description
- **Recommendation**: How to fix

### Verdict
- **READY**: All checks pass, safe to deploy
- **READY WITH CAUTION**: Minor issues, document risks
- **NOT READY**: Blocking issues found, do not deploy

## Error Handling
- If not in a git repo: report "No git repository detected, limited checks available"
- If no test framework found: report "No test framework detected, manual testing required"
- If build fails: report as CRITICAL blocking issue

{
  "skill_name": "deploy-checker",
  "version": "1.0.0",
  "evals": [
    {
      "id": "trigger-001",
      "category": "trigger",
      "prompt": "我的代码能上线了吗",
      "expected_output": "触发部署就绪检查",
      "expectations": [
        {"type": "triggered", "description": "应触发deploy-checker"}
      ]
    },
    {
      "id": "trigger-002",
      "category": "trigger",
      "prompt": "准备发布到生产环境，帮我把关",
      "expected_output": "触发部署就绪检查",
      "expectations": [
        {"type": "triggered", "description": "应触发deploy-checker"}
      ]
    },
    {
      "id": "trigger-003",
      "category": "trigger",
      "prompt": "pre-deploy check for the latest changes",
      "expected_output": "触发部署就绪检查",
      "expectations": [
        {"type": "triggered", "description": "英文prompt也应触发"}
      ]
    },
    {
      "id": "neg-001",
      "category": "negative",
      "prompt": "帮我写一个二叉树的层序遍历",
      "expected_output": "不应触发此Skill",
      "expectations": [
        {"type": "not_triggered", "description": "算法问题不相关"}
      ]
    },
    {
      "id": "neg-002",
      "category": "negative",
      "prompt": "这段代码的时间复杂度是多少",
      "expected_output": "不应触发此Skill",
      "expectations": [
        {"type": "not_triggered", "description": "复杂度分析不相关"}
      ]
    },
    {
      "id": "func-001",
      "category": "functional",
      "prompt": "检查我的代码能不能部署",
      "expected_output": "生成结构化的部署就绪报告",
      "expectations": [
        {"type": "tool_used", "value": "Bash", "description": "执行git/test命令"},
        {"type": "contains", "value": "检查", "description": "包含检查相关内容"},
        {"type": "structured_output", "description": "输出应为结构化报告"}
      ]
    },
    {
      "id": "func-002",
      "category": "functional",
      "prompt": "我刚改了支付模块，能上线吗",
      "expected_output": "重点检查支付模块安全性",
      "expectations": [
        {"type": "tool_used", "value": "Bash", "description": "检查git差异"},
        {"type": "contains", "value": "安全", "description": "支付模块必须安全检查"},
        {"type": "contains", "value": "CRITICAL", "description": "支付变更应标记为高影响"}
      ]
    }
  ]
}

/skill-creator eval deploy-checker

━━━ Eval Report: deploy-checker v1.0.0 ━━━

📊 Summary: 5/7 passed (71.4%)

✅ trigger-001   [trigger]   "我的代码能上线了吗" → TRIGGERED
✅ trigger-002   [trigger]   "准备发布到生产环境..." → TRIGGERED  
✅ trigger-003   [trigger]   "pre-deploy check..." → TRIGGERED
✅ neg-001       [negative]  "帮我写一个二叉树..." → NOT TRIGGERED ✓
✅ neg-002       [negative]  "这段代码的时间..." → NOT TRIGGERED ✓
❌ func-001      [functional] "检查我的代码能不能..." 
   → FAIL: expectation "structured_output" not met
❌ func-002      [functional] "我刚改了支付模块..."
   → FAIL: "CRITICAL" not found in output

📈 Metrics:
   Total tokens: 12,847
   Avg duration: 8.3s per eval
   Tools used: Bash(12), Read(8), Grep(3)

/skill-creator improve deploy-checker --focus trigger

━━━ Trigger Analysis ━━━

Current description:
"检查代码是否准备好部署到生产环境。分析git状态、测试覆盖率、
构建状态和安全风险。Use when the user asks about deploying..."

Suggestions:
1. ADD trigger phrase: "can I release", "上线", "发布"
2. ADD negative trigger: "Do NOT use for coding questions, algorithm help"
3. IMPROVE: Include "pre-deploy" as explicit trigger keyword

Optimized description:
"检查代码变更是否准备好部署到生产环境。分析git状态、测试覆盖率、
构建状态和安全风险。Use when the user asks about deploying, release 
readiness, going to production, pre-deploy checks, 上线检查, or 发布.
Do NOT use for general coding questions or algorithm help."

/skill-creator eval deploy-checker

━━━ Eval Report: deploy-checker v1.1.0 ━━━
📊 Summary: 7/7 passed (100%) ✅
📈 Tokens: 14,203 (+10%) | Avg duration: 9.1s

/skill-creator benchmark deploy-checker

━━━ Benchmark Report: deploy-checker ━━━

┌─────────────────┬────────────┬──────────────┐
│ Metric          │ With Skill │ Without Skill│
├─────────────────┼────────────┼──────────────┤
│ Pass Rate       │ 100% (7/7) │ 42.9% (3/7) │
│ Avg Tokens      │ 14,203     │ 9,812        │
│ Avg Duration    │ 9.1s       │ 5.3s         │
│ Structure Score │ 4.2/5      │ 1.8/5        │
│ Completeness    │ 92%        │ 45%          │
│ Consistency     │ 100%       │ 28%          │
└─────────────────┴────────────┴──────────────┘

Verdict: SKILL IMPROVES OUTPUT QUALITY ✅
- +57% pass rate improvement
- +133% structure score
- +104% completeness
- +257% consistency across runs
- Token cost: +45% (acceptable trade-off)

/skill-creator benchmark deploy-checker --compare previous

# 通过设置环境变量控制并行度
export SKILL_CREATOR_PARALLEL=4

# 运行评测（自动并行化）
/skill-creator eval deploy-checker

/skill-creator benchmark deploy-checker --ab-test

{
  "skill_name": "deploy-checker",
  "evals": [
    {
      "id": "P0-critical-001",
      "priority": "P0",
      "comment": "必须#1通过的用例，失败即阻塞发布"
    },
    {
      "id": "P1-important-001", 
      "priority": "P1",
      "comment": "应在Top 3中，允许偶尔失败"
    },
    {
      "id": "P2-nice-001",
      "priority": "P2",
      "comment": "能触发就好，排名不重要"
    }
  ]
}

#!/bin/bash
# ci-skill-eval.sh

SKILL_NAME=$1
PASS_THRESHOLD=80

# 运行评测
claude -p "/skill-creator eval $SKILL_NAME --json > eval-results.json"

# 解析通过率
PASS_RATE=$(python3 -c "
import json
with open('eval-results.json') as f:
    data = json.load(f)
print(data['summary']['pass_rate'])
")

# 判断是否通过
if (( $(echo "$PASS_RATE < $PASS_THRESHOLD" | bc -l) )); then
    echo "❌ Skill $SKILL_NAME eval failed: ${PASS_RATE}% < ${PASS_THRESHOLD}%"
    exit 1
else
    echo "✅ Skill $SKILL_NAME eval passed: ${PASS_RATE}%"
    exit 0
fi

name: Skill Eval
on: [pull_request]
jobs:
  eval:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Install Claude Code
        run: npm install -g @anthropic-ai/claude-code
      - name: Run Skill Evals
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          ./ci-skill-eval.sh deploy-checker

graph LR
    A[模型升级] --> B[运行无Skill基线]
    B --> C{基线通过率?}
    C -->|<80%| D[Skill仍然必要<br/>继续维护]
    C -->|>=80%| E[Skill可能过期<br/>考虑简化或退役]
    E --> F[运行A/B对比]
    F --> G{Skill仍优于基线?}
    G -->|是| D
    G -->|否| H[退役Skill]

flowchart TB
    subgraph Phase1["阶段一：创建"]
        A1[识别重复工作流] --> A2[编写SKILL.md]
        A2 --> A3[编写evals.json<br/>5+测试用例]
    end

    subgraph Phase2["阶段二：验证"]
        B1[运行Eval模式] --> B2{通过率?}
        B2 -->|<80%| B3[分析失败用例]
        B3 --> B4[改进SKILL.md]
        B4 --> B1
        B2 -->|>=80%| B5[进入Benchmark]
    end

    subgraph Phase3["阶段三：基准"]
        B5 --> C1[运行with/without对比]
        C1 --> C2{Skill优于基线?}
        C2 -->|否| C3[重新评估Skill价值]
        C3 --> A2
        C2 -->|是| C4[记录benchmark基线]
    end

    subgraph Phase4["阶段四：维护"]
        C4 --> D1[模型更新时重跑Benchmark]
        D1 --> D2{回归?}
        D2 -->|是| B3
        D2 -->|否| D3[发布/继续使用]
    end

    Phase1 --> Phase2 --> Phase3 --> Phase4

# 改进前
description: Helps with deployment

# 改进后
description: >
  Checks code readiness for production deployment. Analyzes git status,
  test coverage, build state, and security risks. Use when the user 
  asks about deploying, release readiness, going to production, 
  pre-deploy checks, 上线检查, or 发布.

# 运行5次取平均
/skill-creator benchmark deploy-checker --runs 5