Module cuối cùng trong series Claude Code tổng hợp 22 advanced capabilities mở rộng core Claude Code: Planning Mode, Extended Thinking, Auto Mode, Background Tasks, Scheduled Tasks, Permission Modes, Headless Mode cho CI/CD, Voice Dictation, Chrome Integration, Remote Control, và nhiều hơn nữa. Đây là post khép lại 9-module series về toàn bộ ecosystem Claude Code.

Claude Code Advanced Features: Toàn cảnh các capabilities

Tổng quan - 22 Advanced Features

1. Planning Mode

Planning mode cho Claude think through complex tasks trước khi implement — tạo detailed plan để bạn review/approve.

Two-Phase Approach

1. Planning Phase — Claude analyze task, tạo detailed plan
2. Implementation Phase — Sau approval, Claude execute

Khi nào dùng?

✅ Use cho: Complex multi-file refactoring, new features, architectural changes, database migrations

❌ Don't use cho: Simple bug fixes, formatting changes, single-file edits, quick queries

Activation

Slash command:

/plan Implement user authentication system

CLI flag:

claude --permission-mode plan

Phím tắt:

• Shift + Tab — Toggle permission modes
• Alt + M — Toggle (Windows/Linux)

Model alias opusplan: Opus cho planning, Sonnet cho execution:

claude --model opusplan "design and implement the new API"

Example Output

User: /plan Add a real-time notification system

Claude: I'll create a plan for implementing real-time notifications.

## Implementation Plan

### Phase 1: Infrastructure Setup
1. Install WebSocket libraries (socket.io)
2. Set up WebSocket server
3. Configure connection management

### Phase 2: Backend Implementation
4. Create notification service
5. Implement notification queue
6. Add database schema for notifications
7. Create API endpoints

### Phase 3: Frontend Integration
8. Set up WebSocket client
9. Create notification UI components
10. Implement notification state management

### Phase 4: Testing & Deployment
11. Write unit tests
12. Write integration tests
13. Update documentation
14. Deploy to staging

Estimated time: 3-4 hours
Files to modify: 12
New files to create: 8

Ready to proceed? (yes/no/modify plan)

2. Extended Thinking

Extended thinking cho Claude dành nhiều thời gian reasoning trước khi đưa solution.

Adaptive Thinking trên Opus 4.6+

Opus 4.6 dùng adaptive thinking — model tự quyết định khi nào và bao nhiêu cần think. Effort levels:

• low — minimal reasoning
• medium — balanced
• high — deep reasoning (default trên Opus 4.6)
• max — Opus 4.6 only, maximum reasoning

Activation

Phím tắt: Alt + T / Option + T — toggle on/off

Effort level:

export CLAUDE_CODE_EFFORT_LEVEL=high
claude --effort high "complex architectural review"

Slash command:

/effort high

Magic word: Thêm "ultrathink" trong prompt → activate deep reasoning mode.

Khi nào dùng?

✅ Architectural decisions, complex problem-solving, edge cases analysis ❌ Simple queries, factual lookups, basic edits

Extended thinking adds latency. Don't enable blanket-style mọi calls — route complex tasks to thinking-enabled, route simpler lookups to standard responses.

3. Auto Mode (Research Preview)

Auto Mode dùng background safety classifier review mọi action trước khi execute. Claude work autonomously nhưng block dangerous operations.

Requirements

• Plan: Team plan trở lên
• Model: Sonnet 4.6 hoặc Opus 4.6
• Classifier: Chạy trên Sonnet 4.6 (extra token cost)

Enable

claude --enable-auto-mode
# Hoặc cycle to auto mode trong REPL với Shift+Tab

Default Blocked Actions

Default Allowed Actions

4. Background Tasks

Long-running operations chạy async không block conversation.

Commands

User: Run tests in background

Claude: Started task bg-1234

/task list           # Show all tasks
/task status bg-1234 # Check progress
/task show bg-1234   # View output
/task cancel bg-1234 # Cancel task

Phím tắt: Ctrl+B — Background currently running task

Configuration

{
  "backgroundTasks": {
    "enabled": true,
    "maxConcurrentTasks": 5,
    "notifyOnCompletion": true,
    "autoCleanup": true,
    "logOutput": true
  }
}

5. Scheduled Tasks

Run prompts tự động trên recurring schedule hoặc one-time reminders.

/loop command

# Explicit interval
/loop 5m check if the deployment finished

# Natural language
/loop check build status every 30 minutes

One-time reminders

remind me at 3pm to push the release branch
in 45 minutes, run the integration tests

Limits & Behavior

• Up to 50 tasks per session
• Session-scoped — cleared khi session end
• Recurring tasks auto-expire after 3 days
• Tasks chỉ fire khi Claude Code running — no catch-up
• Not persisted across restarts

Cloud Scheduled Tasks (/schedule)

Persist across restarts, không cần Claude Code running locally:

/schedule daily at 9am run the test suite and report failures

6. Permission Modes

Cycle với Shift+Tab. Set default qua --permission-mode flag.

7. Headless Mode (claude -p)

Non-interactive mode cho automation và CI/CD.

Examples

# Run specific task
claude -p "Run all tests"

# Process piped content
cat error.log | claude -p "Analyze these errors"

# JSON output
claude -p --output-format json "Analyze code quality"

# Limit autonomous turns
claude -p --max-turns 5 "refactor this module"

# Disable session persistence
claude -p --no-session-persistence "one-off analysis"

CI/CD Integration (GitHub Actions)

name: AI Code Review

on: [pull_request]

jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Install Claude Code
        run: npm install -g @anthropic-ai/claude-code

      - name: Run Claude Code Review
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          claude -p --output-format json \
            --max-turns 3 \
            "Review this PR for code quality, security, performance, and test coverage. Output as JSON" > review.json

      - name: Post Review Comment
        uses: actions/github-script@v7
        with:
          script: |
            const fs = require('fs');
            const review = JSON.parse(fs.readFileSync('review.json', 'utf8'));
            github.rest.issues.createComment({
              issue_number: context.issue.number,
              owner: context.repo.owner,
              repo: context.repo.repo,
              body: JSON.stringify(review, null, 2)
            });

8. Session Management

# Continue last conversation
claude -c

# Resume named session
claude -r "auth-refactor" "finish this PR"

# Rename trong REPL
/rename auth-refactor

# Fork session để thử alternative
/fork

# Hoặc CLI:
claude --resume auth-refactor --fork-session "try OAuth instead"

9. Interactive Features - Complete Keyboard Shortcuts

Vim Mode

/vim hoặc /config để enable. Mode switching: Esc (NORMAL), i/a/o (INSERT).

Bash Mode

Execute shell commands trực tiếp với ! prefix:

! npm test
! git status
! cat src/index.js

Custom Keybindings (/keybindings)

{
  "$schema": "https://www.schemastore.org/claude-code-keybindings.json",
  "bindings": [
    {
      "context": "Chat",
      "bindings": {
        "ctrl+e": "chat:externalEditor",
        "ctrl+u": null,
        "ctrl+k ctrl+s": "chat:stash"
      }
    }
  ]
}

Reserved keys (không rebound được): Ctrl+C, Ctrl+D

10. Voice Dictation

/voice

Push-to-talk voice input với 20 languages. Cần Claude.ai account cho STT.

11. Channels (Research Preview)

MCP servers push messages vào running Claude Code sessions:

claude --channels discord,telegram

Supported: Discord, Telegram. Plugins phải approved qua allowedChannelPlugins setting.

12. Chrome Integration (Beta)

Connect Claude Code tới Chrome/Edge browser cho live web automation:

claude --chrome      # Enable
claude --no-chrome   # Disable

Capabilities

• Live debugging — Read console logs, inspect DOM, debug JS real-time
• Design verification — Compare rendered pages vs design mockups
• Form validation — Test form submissions, input validation
• Web app testing — Interact với authenticated apps
• Data extraction — Scrape + process content từ web pages
• Session recording — Record browser interactions as GIF

Limitations

• Browser support: Chrome + Edge only
• WSL: Không available
• Third-party providers: Không support với Bedrock, Vertex, Foundry

13. Remote Control

Continue locally running Claude Code session từ phone, tablet, browser. Local session vẫn chạy trên máy bạn — không có gì move to cloud.

Start

claude remote-control
claude remote-control --name "Auth Refactor"

Connect từ device khác

3 cách:

1. Session URL — printed lúc start, mở browser
2. QR code — bấm spacebar sau start, scan từ phone
3. Find by name — browse trên claude.ai/code hoặc Claude mobile app

Security

• No inbound ports opened trên máy
• Outbound HTTPS only over TLS
• Scoped credentials — short-lived, narrowly scoped tokens

14. Web Sessions

Run Claude Code trực tiếp trong browser tại claude.ai/code.

# Create từ CLI
claude --remote "implement the new API endpoints"

# Resume Locally
claude --teleport

15. Desktop App

Standalone app với visual diff review, parallel sessions, integrated connectors. macOS + Windows (Pro, Max, Team, Enterprise).

Core Features

• Diff view — File-by-file visual review với inline comments
• App preview — Auto-start dev servers với embedded browser
• PR monitoring — GitHub CLI integration, auto-fix CI failures, auto-merge
• Parallel sessions — Multiple sessions trong sidebar với automatic Git worktree isolation
• Scheduled tasks — Recurring tasks (hourly, daily, weekly)
• Connectors — GitHub, Slack, Linear, Notion, Asana, Calendar

16. Task List

Persistent task tracking survive context compactions.

Toggle: Ctrl+T

Named Task Directories

export CLAUDE_CODE_TASK_LIST_ID=my-project-sprint-3

→ Multiple sessions share same task list (team workflows, multi-session projects).

17. Prompt Suggestions

Smart suggestions dạng grayed-out text dựa trên git history + conversation context.

• Tab — accept suggestion
• Enter — accept + immediately submit

Disable:

export CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=false

18. Git Worktrees

Start Claude Code trong isolated worktree — parallel work trên different branches không cần stash/switch.

claude --worktree
# hoặc
claude -w

Sparse Checkout cho Monorepos

{
  "worktree": {
    "sparsePaths": ["packages/my-package", "shared/"]
  }
}

Auto-Cleanup: No changes → worktree tự động cleanup khi session end.

19. Sandboxing

OS-level filesystem + network isolation cho Bash commands.

claude --sandbox       # Enable
claude --no-sandbox    # Disable

Configuration

{
  "sandbox": {
    "enabled": true,
    "failIfUnavailable": true,
    "filesystem": {
      "allowWrite": ["/Users/me/project"],
      "allowRead": ["/Users/me/project", "/usr/local/lib"],
      "denyRead": ["/Users/me/.ssh", "/Users/me/.aws"]
    },
    "enableWeakerNetworkIsolation": true
  }
}

20. Managed Settings (Enterprise)

Deploy Claude Code config across organization qua platform-native management tools.

Available Settings

• disableBypassPermissionsMode — Ngăn user enable bypass permissions
• availableModels — Restrict models user select được
• allowedChannelPlugins — Control channel plugins permitted
• autoMode.environment — Configure trusted infrastructure cho auto mode

21. Configuration & Settings

Complete Config Example

{
  "permissions": {
    "mode": "default",
    "allowedTools": ["Bash(git log:*)", "Read"],
    "disallowedTools": ["Bash(rm -rf:*)"]
  },

  "hooks": {
    "PreToolUse": [{ "matcher": "Edit", "hooks": ["eslint --fix ${file_path}"] }],
    "PostToolUse": [{ "matcher": "Write", "hooks": ["~/.claude/hooks/security-scan.sh"] }]
  },

  "mcp": {
    "enabled": true,
    "servers": {
      "github": {
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-github"],
        "env": {
          "GITHUB_TOKEN": "${GITHUB_TOKEN}"
        }
      }
    }
  }
}

Environment Variables Reference

# Model selection
export ANTHROPIC_MODEL=claude-opus-4-6

# Thinking
export MAX_THINKING_TOKENS=16000
export CLAUDE_CODE_EFFORT_LEVEL=high

# Feature toggles
export CLAUDE_CODE_DISABLE_AUTO_MEMORY=true
export CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=true
export CLAUDE_CODE_DISABLE_CRON=1
export CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=false
export CLAUDE_CODE_ENABLE_TASKS=true
export CLAUDE_CODE_SIMPLE=true              # --bare flag

# MCP config
export MAX_MCP_OUTPUT_TOKENS=50000
export ENABLE_TOOL_SEARCH=true

# Task management
export CLAUDE_CODE_TASK_LIST_ID=my-project-tasks

# Subagent
export CLAUDE_CODE_SUBAGENT_MODEL=sonnet

22. Workflow ứng dụng cho Blog Project

Cho project blog Next.js cá nhân, đây là combinations có thể dùng:

Workflow 1: Plan Mode + Subagents cho feature mới

1. /plan Add comments system với Disqus alternative (custom backend)
   → Claude tạo detailed plan
2. Review, modify nếu cần
3. /plan approve
4. Claude delegate work qua subagents:
   - implementation-agent: backend API
   - test-engineer: integration tests
   - documentation-writer: API docs
5. Background tasks cho test suite chạy
6. Stop hook verify all tasks completed

Workflow 2: Auto Mode cho daily blog work

# Workflow setup
claude --enable-auto-mode --permission-mode auto

Auto mode block:

• vercel remove, git push --force (production deploys, force push)
• rm -rf lớn (mass deletion)

Auto mode allow:

• File reads/writes trong blog/, components/
• npm install, npm run dev, npm run build
• git add, git commit, git push origin <feature-branch>

Workflow 3: Headless Mode cho CI/CD

.github/workflows/blog-quality-check.yml:

name: Blog Quality Check

on:
  pull_request:
    paths:
      - 'blog/**.md'

jobs:
  audit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npm install -g @anthropic-ai/claude-code

      - name: SEO + Quality Audit
        env:
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
        run: |
          claude -p --output-format json \
            --max-turns 5 \
            "Audit changed MD files for:
            - SEO frontmatter (title 50-60 chars, description 150-160 chars)
            - Image alt texts
            - Code block language tags
            - Internal links validity
            Output as JSON with severity levels" > audit.json

Workflow 4: Scheduled Tasks cho monitoring

# Loop monitoring
/loop 30m check Vercel deployment status. If new deploy succeeded, summarize the changes.

# Reminder
remind me at 5pm to schedule next week's blog post

Workflow 5: Git Worktrees cho experiment redesign

# Trong main project
claude --worktree

# Worktree tự động tạo tại .claude/worktrees/redesign-2026/
# Thử major redesign
> Refactor blog layout với new design system

# Tests pass → merge vào main
# Tests fail → exit worktree, không impact main branch

Workflow 6: Remote Control khi đi cafe

# Tại nhà — start remote session
claude remote-control --name "Blog Sunday Sprint"

# Press spacebar để show QR code

# Tại cafe — scan QR code từ phone
# Continue session trên iPad/phone
# Local tools (filesystem, MCP) vẫn run trên máy ở nhà

Workflow 7: Chrome Integration cho design review

claude --chrome

> Open localhost:3000 và compare blog layout với design ở Figma
> Take screenshots của 5 blog post pages
> Verify dark mode toggles work correctly
> Test contact form submission với invalid email

23. Câu hỏi thường gặp (FAQ)

Q1: Plan Mode vs Extended Thinking khác gì?

→ Plan mode outline steps. Extended thinking reason deeper.

Q2: Auto Mode an toàn không? Có recommend cho personal project không?

Auto Mode cẩn thận với:

• ✅ Hữu ích cho autonomous work trên large refactors
• ✅ Background classifier block dangerous ops
• ⚠️ Vẫn có token cost cho classifier (extra Sonnet 4.6 calls)
• ⚠️ Classifier có thể false negative trên edge cases
• ⚠️ Chỉ available Team plan trở lên

Cho personal project:

• Pro/Max users → dùng acceptEdits mode + careful permissions trong settings.json là đủ

Q3: Khi nào dùng Background Tasks vs Subagents?

→ Background Tasks cho long shell ops. Subagents cho specialized AI tasks.

Q4: Forking session khác Rewind thế nào?

Q5: Sandboxing có chặn được mọi malicious code không?

Sandboxing là 1 layer defense, không phải silver bullet:

• ✅ Block filesystem access ngoài allowed paths
• ✅ Network isolation (không full trên macOS)
• ⚠️ Không protect against legitimate-looking malicious code
• ⚠️ Bash sandbox không cover Python/Node scripts spawned

Defense in depth:

1. Permission rules (allow/deny patterns)
2. Sandbox (filesystem + network)
3. Hooks (PreToolUse validation)
4. Code review

Q6: Channels (Discord/Telegram) có scale lên enterprise không?

Hiện tại là Research Preview — chưa production-ready cho enterprise:

• 2 integrations only (Discord, Telegram)
• Plugin-based architecture
• Managed via allowedChannelPlugins cho team control

Cho enterprise: dùng MCP servers + webhooks pattern thay thế hoặc đợi GA release.

Q7: Remote Control có thay thế Claude Web không?

Không — bổ sung cho nhau:

Q8: Desktop App vs Claude Code CLI khác gì?

→ Desktop App cho integrated workflow. CLI cho scripting + power users.

Q9: Worktrees có conflict với git submodules không?

Worktrees + submodules có thể tricky:

• Worktree có own .git directory
• Submodules share same submodule URLs
• Manual init submodules sau create worktree:

cd .claude/worktrees/<name>
git submodule update --init --recursive

Q10: Managed Settings có override user settings không?

Có — managed settings takes precedence trên user-level:

1. Managed drop-ins (managed-settings.d/*.json) merged alphabetically
2. Managed plist/Registry settings
3. User settings (~/.claude/settings.json)
4. Project settings (.claude/settings.json)

→ Org-level config enforce, user không override được những fields managed.

Q11: Claude Code có offline mode không?

Không — Claude Code requires internet để communicate với Anthropic API.

Local-only options:

• Claude Code's local components (hooks, file operations) work mà không cần network
• MCP servers local-only work nhưng Claude vẫn cần API call
• Cho fully offline: cần on-prem deployment (Bedrock, Vertex, Foundry — Enterprise)

Q12: Sessions persistence storage ở đâu? Cleanup được không?

Storage:

~/.claude/projects/{project}/{session-id}/
├── transcript.jsonl     # Conversation history
├── checkpoints/         # Checkpoint snapshots
├── tasks/               # Task list state
└── subagents/           # Subagent transcripts

Cleanup:

# Manual cleanup old sessions
find ~/.claude/projects -name "transcript.jsonl" -mtime +30 -exec rm -rf {}/.. \;

Disk usage có thể grow với heavy use → periodic cleanup recommended.

Q13: Extended Thinking có affect performance API/Token cost không?

Có:

• Token cost: thinking tokens count toward output budget
• Latency: thêm 5-30s tùy effort level
• API rate limits: thinking tokens count
• Caching: thinking output không cached giữa requests

Cost optimization:

• Default medium cho most tasks
• high/max chỉ cho architectural decisions
• Disable cho simple queries với low hoặc env variable

Q14: Auto Mode classifier có học từ approvals của user không?

Không — classifier là stateless model evaluation, không train theo user behavior.

Mỗi action evaluate independently dựa trên:

• Action content (command, file, target)
• Trusted infrastructure config (autoMode.environment)
• Default rules

Workaround "learning" — dùng auto-adapt-mode.py hook tự add safe rules vào allowlist sau approvals.

24. Kết luận - Hành trình 9 modules Claude Code

Module 09 này là post cuối cùng trong series 9-part về Claude Code. Recap toàn bộ hành trình:

Big picture - Claude Code là gì?

Sau 9 modules, ta thấy Claude Code không phải chỉ là "AI helper trong terminal". Nó là toolkit có 4 tầng abstraction:

Tầng 1 - Interaction: Cách bạn nói với Claude (slash commands, voice, chat)

Tầng 2 - Knowledge: Cách Claude biết về project (memory, skills, MCP)

Tầng 3 - Execution: Cách Claude làm việc (subagents, background tasks, worktrees, plan mode, extended thinking)

Tầng 4 - Control: Cách bạn kiểm soát Claude (hooks, permissions, sandboxing, checkpoints)

Plugins (module 07) là cơ chế đóng gói cả 4 tầng thành package shareable.

Roadmap khuyến nghị toàn series

Tuần 1-2 (Cơ bản):

• Module 01 (Slash Commands)
• Module 02 (Memory)
• Tạo CLAUDE.md đầu tiên, học 5-10 built-in commands

Tuần 3-4 (Tự động hoá nhẹ):

• Module 03 (Skills) — tạo 2-3 custom skills
• Module 06 (Hooks) — thêm 1 PostToolUse format hook

Tháng 2 (Specialized):

• Module 04 (Subagents) — tạo code-reviewer, test-engineer agents
• Module 05 (MCP) — setup GitHub MCP

Tháng 3 (Power user):

• Module 07 (Plugins) — bundle workflow thành plugin
• Module 08 (Checkpoints) — học rewind workflow

Tháng 4+ (Mastery):

• Module 09 (Advanced) — Plan mode, Extended thinking, Auto mode, Headless cho CI/CD

Triết lý cuối cùng

Đừng học hết toàn bộ Claude Code rồi mới bắt đầu dùng. Bắt đầu với cơ chế đơn giản nhất giải quyết được vấn đề bạn đang có. Khi gặp pain point cụ thể, mở module tương ứng học cách giải quyết. Lặp lại.

Không có "ninja master" Claude Code phải biết hết 9 modules. Có "developer hiệu quả" biết chọn đúng tool cho đúng task.

Lời cảm ơn

Cảm ơn bạn đã đọc đến cuối series 9-part về Claude Code. Hy vọng bộ tài liệu này giúp bạn hiểu sâu hơn về toolkit và áp dụng vào workflow hàng ngày.

Claude Code đang phát triển rất nhanh — features mới ra liên tục, best practices đang định hình. Hãy thường xuyên check:

• Official Documentation
• Release Notes — /release-notes trong session
• GitHub Issues — community discussions
• Anthropic Engineering Blog — deep dives

Chúc bạn build được những workflows, plugins, skills, agents tuyệt vời cho project của riêng mình. Happy coding với Claude Code! 🚀

Hãy bắt đầu hành trình bằng cách gõ /help trong Claude Code, rồi từng bước explore. Mỗi tuần làm quen với 1 cơ chế mới, sau 2-3 tháng bạn sẽ có toolkit cá nhân hóa hoàn toàn cho cách làm việc của mình.