Claude-Code on Chen Kai Blog

Claude Code Hands-On (10): Skills, and When to Reach for Each Extension Mechanism

Mon, 27 Apr 2026 09:00:00 +0000

Claude Code now has four extension mechanisms: slash commands, MCP servers, hooks, and Skills. They overlap. The first time you have a “Claude should know how to do X” thought, the question is which of the four to use.

This is the closing chapter of the series. Let me lay out the decision tree.

Claude Code Hands-On (9): settings.json, the Three-Layer Permission Model, and Env

Sun, 26 Apr 2026 09:00:00 +0000

Hooks let you interact with Claude Code, while settings.json specifies what it can access. This file also confuses many with its precedence rules.

This chapter is the missing reference.

The three layers#

There are three settings.json files Claude Code reads, in order:

User settings — ~/.claude/settings.json. Applies to every project on your machine.
Project settings — <repo>/.claude/settings.json. Committed to git. Applies to anyone working in this repo.
Local settings — <repo>/.claude/settings.local.json. Gitignored. Your private overrides for this repo.

The merge rule: later layers override earlier ones, key by key. Permissions are additive for allow, subtractive for deny — once any layer denies something, no other layer can re-allow it. This asymmetry is what makes the system safe.

Claude Code Hands-On (8): Sub-Agents, Worktrees, and Plan Mode

Sat, 25 Apr 2026 09:00:00 +0000

After hooks, the next thing that changes how Claude Code operates is concurrency control. Not in the threading sense, but in terms of how many tasks it handles, their isolation, and the level of oversight.

Three features, in increasing order of trust required.

Plan mode — the airlock#

Plan mode is the cheapest of the three. Press Shift+Tab until the indicator says plan. The model now plans without taking any actions. It will read, think, propose, and stop. You read the plan. You either approve, edit, or kill it. Only then does it execute.

Claude Code Hands-On (7): Ten Hooks I Actually Use, with the Code

Fri, 24 Apr 2026 09:00:00 +0000

Chapter 5 provided a conceptual tour of hooks. This chapter is the field guide. From the 100-script reference repo, ten scripts earn their place in every serious project I run. I’ll walk through these ten with code.

All examples assume Node 18+, save scripts to ./hooks/, mark them chmod +x, and wire them in .claude/settings.json like:

1
2
3
4
5
6
7


{
 "hooks": {
 "PreToolUse": [
 { "matcher": "Read|Grep", "hooks": [{ "type": "command", "command": "node ./hooks/block-env-read.js" }] }
 ]
 }
}

Before we dive in, here’s the hook lifecycle to make the following code clear:

Claude Code Hands-On (6): The SDK, GitHub Integration, and Claude in CI

Thu, 23 Apr 2026 09:00:00 +0000

The CLI is the obvious surface. The SDK is the interesting one. The GitHub integration is where it pays off.

The SDK in one paragraph#

@anthropic-ai/claude-code is the npm package. It exposes the same Claude Code engine the CLI uses, with the same tools and permissions, as a programmatic interface. You give it a prompt; you get an async iterable of conversation events. Plug it into anything — a script, a service, a CI step.

Claude Code Hands-On (5): Hooks, or How to Stop Worrying About Yolo Mode

Wed, 22 Apr 2026 09:00:00 +0000

If MCP is how Claude reaches out, hooks are how you reach in. They enforce the rules you care about, not just what you hope for.

The model#

A hook is a shell command that Claude Code runs at specific moments. The two you’ll use most are:

Claude Code Hands-On (4): MCP Servers, or How Claude Talks to Anything

Tue, 21 Apr 2026 09:00:00 +0000

If you only learn one extension mechanism in Claude Code, learn MCP. It is the difference between an autocomplete and a platform.

The 60-second pitch#

MCP is Model Context Protocol — a small open spec that lets Claude Code call tools and read resources from external servers. The “server” is any process that speaks MCP over stdio or HTTP. Claude treats MCP tools the same as its built-in tools: the model decides when to call them, you confirm or auto-approve, results come back as text.

Claude Code Hands-On (3): Custom Slash Commands and Conversation Control

Mon, 20 Apr 2026 09:00:00 +0000

Built-in slash commands like /clear and /init are the visible part of the iceberg. The main point of the system is for you to write your own commands, which live in your repo.

What a slash command is#

A file at .claude/commands/<name>.md. Contents are a Markdown prompt. Filename becomes the command. After creating the command, you need to restart Claude Code (one of the few places it isn’t hot-reloaded).

Claude Code Hands-On (2): Shortcuts, the Four-State Toggle, and Thinking Modes

Sun, 19 Apr 2026 09:00:00 +0000

The shortcuts are not in the help screen for a reason — they’re discoverable through use, not documentation. Here they are anyway.

`Shift+Tab` — the four-state cycle#

Most people think Shift+Tab toggles auto-accept on and off. It doesn’t. It cycles four states in order:

1

Normal → Accept edits → Plan mode → Bypass permissions → Normal

Each state changes what Claude does without asking:

Claude Code Hands-On (1): Install, the Three-Layer Config, and the # @ /init Trio

Sat, 18 Apr 2026 09:00:00 +0000

This is the first in a ten-part field guide to Claude Code. The order is deliberate: each piece unlocks the next. By the end, you’ll be using features that 90% of users never touch.

Install#

There is one supported install path, and it’s the right one.