Claude Code 实战（十）：Skills 与四种扩展机制

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18

---
name: chenk-blog-write
description: Use when writing new content for chenk.top — bilingual EN/ZH posts, series, tutorials. Covers front matter, voice, matplotlib figures, cover generation, deploy.
---

# Voice
- First person, dry, restrained. No "let's", no exclamations.
- One claim, one example. If the claim has no example, cut the claim.

# Front matter
[exact schema goes here]

# Workflow
1. Read source
2. Write EN
3. Adapt to ZH (not translate)
4. Generate covers
5. Build + deploy

1
2
3
4

# Voice
- First person, practical, no filler
- Code before explanation
- If a claim has no example, cut the claim

1
2
3
4
5

# Front matter format
title: string (required)
date: YYYY-MM-DD HH:MM:SS (required)
tags: list of strings
series: string (optional, must match series name in config)

1
2
3
4
5
6

# Workflow
1. Read the existing content at [path]
2. Generate outline with N sections
3. Write each section
4. Run build to verify
5. Deploy

1
2
3
4
5

# Rules
- Never create documentation files unless explicitly asked
- Never add emojis
- Always use absolute paths
- Keep files clean — no v1/v2 suffixes

1
2
3
4
5
6
7
8

~/.claude/
  skills/
    chenk-blog-write/
      SKILL.md
    update-config/
      SKILL.md
    simplify/
      SKILL.md

1
2
3
4
5
6
7

my-project/
  .claude/
    skills/
      deploy-staging/
        SKILL.md
      code-review/
        SKILL.md

1
2
3
4

---
name: deploy-staging
description: Use when deploying to the staging environment, preparing a staging build, or troubleshooting staging deployment issues. Covers the deploy pipeline, environment config, health checks, and rollback procedures for the acme-api project.
---

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

---
name: deploy-staging
description: Use when deploying to the staging environment, preparing a staging build, or troubleshooting staging deployment issues. Covers the deploy pipeline, environment config, health checks, and rollback procedures for the acme-api project.
---

# Staging environment

- URL: https://staging.acme.dev
- Region: us-east-1
- Cluster: acme-staging-ecs
- Branch: deploys from `main` or any branch with `staging-` prefix

# Pre-deploy checklist

1. All tests pass: `npm run test`
2. TypeScript compiles: `npm run build`
3. No uncommitted changes: `git status` must be clean
4. On correct branch: `main` or `staging-*`

# Deploy command

1
2
3
4

# Health check

After deploy, wait 60 seconds, then verify:

1
2
3
4

# Rollback

If the deploy is bad:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

# Common issues

- **Deploy times out**: Check CloudWatch logs for the new task. Usually a startup crash.
- **Health check fails**: Check if the new code has a different health endpoint. It must be `/health`.
- **ECR push fails**: Run `aws ecr get-login-password` to refresh credentials.

# Rules

- Never deploy directly to production from this skill. Staging only.
- Always run the health check after deploy.
- If the health check fails, roll back immediately — do not debug on staging.

1
2
3
4
5
6

mkdir -p .claude/skills/deploy-staging
# (write the SKILL.md)

# Start Claude and test
claude
> Deploy the latest changes to staging.

1
2
3
4

---
name: python-ml
description: Use when writing or modifying Python ML/AI code — model training, data pipelines, evaluation scripts. Only relevant in Python projects with ML dependencies.
---

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

---
name: strict-review
description: Use when reviewing code changes before commit. Applies strict standards.
---

# When to apply strict standards

Apply these standards ONLY when:
- The file is in `src/core/` or `src/security/`
- The change touches authentication, authorization, or payment logic
- The user explicitly asks for a strict review

For other files, apply standard review practices.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45

START: "I want Claude to do X"
  |
  v
Does X need a tool Claude doesn't have?
(browser, database, external API, etc.)
  |
  ├── YES ──> Build an MCP server.
  |           Only MCP can grant new capabilities.
  |
  └── NO
      |
      v
Should X happen automatically around tool calls?
(block dangerous commands, format on save, log all calls)
  |
  ├── YES ──> Write a hook.
  |           Hooks are the only mechanism that runs
  |           without the model deciding to invoke it.
  |
  └── NO
      |
      v
Is X a compact procedure the user will invoke by name?
(/deploy, /commit, /make-changelog, /run-migrations)
  |
  ├── YES ──> Write a slash command.
  |           Commands are for things you call explicitly.
  |
  └── NO
      |
      v
Is X a body of domain knowledge that should kick in
whenever the topic comes up?
(writing style, deploy procedures, code conventions,
 project-specific knowledge)
  |
  ├── YES ──> Write a skill.
  |           Skills are for things you want Claude to
  |           recognize and apply without being told.
  |
  └── NO
      |
      v
Maybe you don't need an extension.
Just tell Claude in the conversation.

1
2
3
4
5

# BAD:
description: Use for general programming tasks.

# GOOD:
description: Use when writing or modifying Python FastAPI endpoints — covers route structure, dependency injection, Pydantic models, and error handling for the acme-api project.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26

~/.claude/
  settings.json                  # Global deny rules, global hooks
  skills/
    chenk-blog-write/SKILL.md    # Blog writing
    simplify/SKILL.md            # Code review
    update-config/SKILL.md       # Settings management
    claude-api/SKILL.md          # Anthropic SDK usage

my-project/
  .claude/
    settings.json                # Project permissions, env, hooks
    settings.local.json          # My local overrides
    commands/
      deploy.md                  # /deploy command
      release.md                 # /release command
    skills/
      project-conventions/SKILL.md  # Project-specific coding standards
    agents/
      research.md                # Research sub-agent
      test-writer.md             # Test-writing sub-agent
  hooks/
    block-env-read.js            # Secret protection
    bash-blacklist.js            # Dangerous command blocking
    format-on-write.js           # Auto-formatting
    test-on-edit.js              # Auto-testing
  mcp.json                      # MCP server config (browser, DB, etc.)