Hướng dẫn sử dụng Claude Code từ A đến Z dành cho người mới

Summary of Technological Concepts & Features (Claude Code tutorial)

1) What Claude Code is used for

• AI coding agents that help with development by working directly with your environment.
• Compared to a simple web chat, Claude Code is emphasized as being able to:
  • Access local files
  • Run development workflows
• The tutorial also references the broader ecosystem of AI coding agents, but focuses on Claude Code (with earlier mentions of names like Codex/Cal/Special).

---

2) Two ways to use Claude Code (and the tradeoff)

1. Terminal / VS Code terminal approach

   • Runs Claude Code via the VS Code integrated terminal.
   • Claimed benefits:
     • Lighter
     • Less CPU/RAM
     • More control through options (less graphical overhead)

2. IDE panel / Claude Code VS Code extension approach

   • Uses the Claude extension UI panel in VS Code (spark icon / activity bar).
   • Claimed benefits:
     • More beginner-friendly
     • More visual interaction

---

3) Installation & setup

• The tutorial points to Claude Code documentation for detailed install steps.
• For VS Code:
  1. Download VS Code
  2. Install the Claude Code extension
  3. Create a new workspace/project folder in VS Code

---

4) Core prompt files: clot.md / rule system (persistent project behavior)

The tutorial introduces a persistent rules file (shown as clot.md / “clot code.md” in subtitles).

Purpose:

1. Provide consistent guidance so outputs don’t change style across sessions.
2. Preserve a “project identity” so coding decisions stay stable even if session context resets.

Demonstrated behavior:

• Create/update the rules file to enforce formatting (e.g., start responses with smiley icons).
• Rules apply to new work sessions, not necessarily mid-session—unless you start a new session.
• Rules can be overwritten/updated; restarting a session may be necessary.

---

5) Session management & context window

Claude Code uses context, which includes:

• The conversation
• Loaded files
• Tool results All converted into tokens.

Key constraint: context window limit

• Over time, context grows and becomes less efficient.

Tokens

• Treated as the model’s basic processing unit (described with approximate granularity such as ~0.7 words/character-level detail).

As you:

• Ask more questions
• Paste files
• Analyze images/PDFs …context usage rises.

Commands discussed:

• slash clear Clears conversation history to start a fresh session and reduce accumulated context.

• compact / compact focus Performs context compression by retaining only relevant information.

  • Risk: compression may remove needed details.
  • Auto-compression may trigger at high usage (examples mentioned around ~80–90%, sometimes ~96–97%).

---

6) Claude Code autonomy modes (safety vs convenience)

The tutorial presents multiple operation modes:

1. Manual
   • Requires approval for each change/command.
2. Edit automatically
   • Edits automatically, but commands still require approval.
3. Plan
   • Read-only; produces a plan without editing.
4. Auto (LIVE)
   • High autonomy: edit and run commands, but with safety checks/blocking risky actions.
5. Bypass / bypass permissions
   • Minimal prompting for command execution.
   • Strongly warned against for beginners unless confident and in a safe environment.

Related setting:

• VS Code extension toggle shown as “allow dangerously skip permissions” (bypass permission toggle)

Recommendation:

• Use least-privileged modes and verify outputs yourself.

---

7) Efficiency setting: “effort” levels

A selector controls reasoning depth / computation intensity:

• Lower effort
  • Faster
  • Cheaper
  • Less thorough
• Higher effort
  • Slower
  • More tokens/cost
  • More thorough reasoning

Multiple levels are mentioned (e.g., “default” and higher “extra max”-like options).

---

8) Project initialization with init command

The init command generates initial project structure and a starting rules file.

Behavior:

• If the folder is empty:
  • Creates scaffolding and asks what you want to build.
• If the folder already has files:
  • Claude Code analyzes existing files and generates clot.md.

Generated output may include:

• Proposed app structure
• Generated clot.md rules
• Code files and project documentation

---

9) Rewinding/resuming sessions

The tutorial describes “time travel”-like session control:

• /rewind
  • Returns to a previous point within the same session history by creating a new branch/session.
  • Rewinds only backward (not forward).
• Resume
  • Jumps between sessions in a later/other timeline, distinct from rewind.

---

10) Rule-file scaling: splitting clot.md to keep under limits

Practical guidance:

• Recommended clot.md size is around 200 lines (citing an Anthropic suggestion).
• Too-long rules can:
  • Consume excessive tokens when reading
  • Reduce compliance quality (the model may not follow all rules well)

Solution demonstrated: splitting

• If clot.md is very large (example: ~600 lines), Claude Code can:
  • Extract and split clot.md into multiple smaller rule files based on headings (levels 2 and 3)
  • Store them in a hidden rules directory (example: /.clot/rules/-like structure)
  • Create a main clot.md that references sub-rule files

Why the leading dot folder is “hidden”

• It follows a Unix filesystem convention.
• It’s hidden but still accessible.

---

11) “Skills” vs rules

A major distinction is between skills and rules:

• clot.md (rules):
  • Always applied across sessions
• Skills:
  • Activated only when relevant conditions are met (e.g., “when situation X happens”)

Creating skills:

1. Generate them yourself by analyzing the codebase and specifying purpose.
2. Use prebuilt community/official skills.

Skill structure (described):

• Stored under something like /.clot/skills/
• Each skill has a skill.md
• Uses metadata front matter such as:
  • name
  • description
  • execution specification

Activation:

• Automatically, based on request matching
• Manually via a directive (mentioned like a /add <skill> ...-style command)

---

12) Using official “skins” / plugins and installing them

The tutorial discusses Anthropic-provided skill/skins packs and templates.

Installation methods:

• UI:
  • “manage plugins”
  • install plugin/skin
  • restart Claude Code
• Natural language:
  • “Install [skin] for me…” using a pasted plugin link

Demonstration:

• A “frontend design” skin used to redesign a website.

Claimed benefit:

• Skins/templates often improve results and save effort vs building everything from scratch.

---

13) MCP (Model Context Protocol) servers for external tools

MCP = Model Context Protocol, enabling Claude Code to connect to external tools/data sources and automate actions beyond its native environment.

Without MCP limitation (as described):

• Claude Code can reason about and view web content,
• But can’t interact with web UI elements (typing/clicking) without proper tool integration.

Demo scenario:

• Reading VNExpress-like content for interactive steps (search box + screenshot) required MCP.

Example MCP server:

• Playwright MCP (subtitle: “Playdrive” / “PlayDrive MCP”) for browser automation

Installation of an MCP server:

• Done via natural language command to install/connect
• Requires restart/reopen so tools become active

Security warnings:

• Don’t install unknown MCP servers
• Use least privilege for tokens/permissions (read-only when possible)
• Avoid bypass permission modes for risky tasks
• Mentioned: a past malicious MCP example involving subtle data exfiltration (or adding hidden recipients)

---

14) Hooks (automation notifications)

Hooks are automation triggers that notify you at certain times, such as:

• When tasks complete
• Playing sounds / showing popups

Demo:

• Hook triggers audio + a floating notification after the agent finishes tasks.

---

15) Completion conditions for iterative coding/testing

The tutorial shows an iterative workflow where Claude Code keeps working until a completion condition is met:

• build → test → verify → continue until criteria pass

Example:

• A web-app workflow using a “completion” checklist (tests/practical checks) and continuing until all pass.

---

Main speakers/sources (as inferred from the video)

• Primary speaker: the video host/instructor (unnamed in subtitles), teaching “Claude Code from A to Z”
• Referenced organizations/tools:
  • Anthropic
  • Antropic/Antropic GitHub skins
  • VS Code
  • MCP Playwright MCP (Playwright MCP server example)