OpenClaw 开发系列:第10篇-技能(Skills)— 用 Markdown 驾驭 Agent 行为

{  "id": "my-plugin",  "skills": "skills"  // 技能目录相对于插件根目录}

my-plugin/├── openclaw.plugin.json├── index.ts└── skills/    ├── code-review/    │   └── SKILL.md    ├── deploy/    │   └── SKILL.md    └── testing/        └── SKILL.md

---name: deploy-workflowdescription: Deploy application to staging or production environmenttrigger: when user asks to deploy, or mentions deploymenttools: [exec, read, browser]---# Deployment Workflow## Steps### Step 1: Pre-flight CheckUse `exec` to run tests:exec: npm testtextIf tests fail, STOP and report the failures. Do not proceed with deployment.### Step 2: BuildUse `exec` to build:exec: npm run buildtext### Step 3: Verify BuildUse `read` to check the build output:read: dist/index.js (first 50 lines)text### Step 4: DeployIf target is staging:exec: npm run deploy:stagingtextIf target is production:- First ask user for confirmation- Then run:exec: npm run deploy:productiontext### Step 5: VerifyUse `browser` to verify the deployment:browser: navigate to https://app.example.com/healthtextCheck that the health endpoint returns 200 OK.## Error Handling- If any step fails, report the error and suggest how to fix it- Never skip steps, even if user asks to "just deploy"- If the user asks to deploy to production, always double-check their intent

---name: api-debuggerdescription: Debug API issues by analyzing logs, testing endpoints, and checking configurationstrigger: when user reports API errors, 500 errors, or unexpected API behaviortools: [read, exec, browser, web_search]---# API Debugger## Decision Tree### Is the error reproducible?- **Yes** → Use `exec` to reproduce and capture the full error:exec: curl -v https://api.example.com/problematic-endpointtextAnalyze the response status code, headers, and body.- **No** → Check recent logs:read: logs/app.log (last 100 lines)textSearch for error patterns.### Is it a client error (4xx)?- Check authentication configuration- Verify request format and parameters- Review API documentation with `web_search`### Is it a server error (5xx)?- Check server logs for stack traces- Verify database connectivity- Check recent deployments for regressions## Reporting FormatWhen reporting findings, structure as:1. **Symptom**: What the user is experiencing2. **Root Cause**: What is actually wrong3. **Evidence**: Logs, error messages, test results4. **Solution**: Specific steps to fix the issue5. **Prevention**: How to avoid this in the future

# 不好的描述description: Help with code# 好的描述description: Perform code review focusing on security vulnerabilities,  performance issues, and code correctness. Activate when user  shares code and asks for review or feedback.

# 不好的触发条件trigger: when needed# 好的触发条件trigger: activate when user asks to "review code", "check for bugs",  "is this secure?", or shares a pull request URL

## Edge Cases- If the file is larger than 1000 lines, use `exec` with grep first  to narrow down the relevant section- If the API returns a redirect (3xx), follow it and check the final response- If the user is not authenticated, ask for credentials before proceeding

# 不好：过于僵化Always run step 1, then step 2, then step 3. Never skip any step.# 好：有合理的灵活性Generally follow the order: 1 → 2 → 3. However, if you alreadyhave the information from step 1 (e.g., from a previous analysis),you may skip directly to step 3.