Claude Code Usage Notes: CLI Agent Workflow in Practice

Claude Code Usage Notes & Reflections

I’ve been using these tools recently, so I’m jotting down some notes for my own reference.

These CLI-style tools all work in the same way; the same applies to Gemini CLI, Qwen CLI, Cursor CLI, and others.

Basic Usage

Launch and Modes

• Interactive mode (REPL):

claude

• Non-interactive mode:

claude -p "你的问题"

• Help: claude -h

Common Use Cases

• Code questions and answers: How does foo.py work?
• Modify files: Update bar.ts to ...
• Fix errors: cargo build Paste the error into Claude
• Run commands: Enter /help to see what’s supported
• Execute Bash: !ls, !git status

Common Commands

• /add-dir Add a new directory
• /agents Manage subagents (role configuration)
• /bashes Manage background Bash shells
• /clear Clear conversation history
• /compact Clear history but keep the summary
• /context View context usage
• /cost View the cost of the current session
• /doctor Check Claude Code installation/configuration
• /exit Exit the REPL
• /export Export the conversation
• /help View help
• /hooks Manage tool hooks
• /ide IDE integration status
• /init Initialize a CLAUDE.md document
• /login / /logout Log in/log out
• /mcp Manage MCP services
• /memory Edit Claude memory files
• /model Set the AI model
• /output-style Set the output style
• /pr-comments Get PR comments
• /review Review a Pull Request
• /security-review Review the security of the current branch
• /status View status (version/account/tools)
• /upgrade Upgrade to Max
• /vim Toggle Vim mode

Using Claude Code

Command Integration

You can run commands such as git and pytest directly in the Claude interface, and Claude will explain the results.

• Enter this in Claude’s interactive window:

!git status

Claude runs this command in the background, displays the result, and automatically explains which files changed and the current branch status.

• For example:

!pytest

It runs the tests and outputs details on passes and failures. Claude tells you which tests failed, why they failed, and possible fixes.

CLAUDE.md

In Claude Code, CLAUDE.md is a project documentation file, equivalent to “Claude’s user guide + project context index.”

Screenshots

The specific results:

Todos

Usage tips

1. Ask about code logic: Ask directly how a particular file/function works.
2. Analyze errors: Paste in build or runtime errors, and Claude will suggest a fix.
3. Modify files: Describe what needs to change in natural language. Claude generates a diff and applies it after confirmation.
4. Command integration: You can run commands such as git and pytest directly in the Claude interface, and Claude will explain them.
5. Multi-file collaboration: Claude can trace context across files, making it suitable for refactoring or security reviews.

Connecting Other Models

The approach is to modify Claude Code’s model by overriding environment variables, but normally you need to use one that supports this protocol.

# 当前对话
$env:ANTHROPIC_BASE_URL = "..."
$env:ANTHROPIC_AUTH_TOKEN="..."


# 永久设置
setx ANTHROPIC_BASE_URL "..."
setx ANTHROPIC_AUTH_TOKEN="..."

Otherwise, you can use an API proxy on GitHub.

Connecting GLM4.5

GLM4.5 supports the ANTHROPIC protocol.

$env:ANTHROPIC_BASE_URL = "https://open.bigmodel.cn/api/anthropic"
$env:ANTHROPIC_AUTH_TOKEN="..."

setx ANTHROPIC_BASE_URL "https://open.bigmodel.cn/api/anthropic"
setx ANTHROPIC_AUTH_TOKEN="..."

Connecting DeepSeek

DeepSeek has a version that supports the corresponding protocol.

$env:ANTHROPIC_BASE_URL= "https://api.deepseek.com/anthropic"
$env:ANTHROPIC_AUTH_TOKEN="sk-..."

setx ANTHROPIC_BASE_URL "https://api.deepseek.com/anthropic"
setx ANTHROPIC_AUTH_TOKEN "sk-..."

Additional Configuration

Using Agents

There is a slight difference here.

How should we understand the agents in Claude Code?

This fairly comprehensive project is used here:

GitHub - wshobson/agents: A collection of production-ready subagents for Claude Code

cd ~/.claude
git clone https://github.com/wshobson/agents.git

In the /agent command, you can see that they have loaded successfully.

Here’s a look at the contents of the project above:

---
name: ai-engineer
description: Build LLM applications, RAG systems, and prompt pipelines. Implements vector search, agent orchestration, and AI API integrations. Use PROACTIVELY for LLM features, chatbots, or AI-powered applications.
model: opus
---

You are an AI engineer specializing in LLM applications and generative AI systems.

## Focus Areas
- LLM integration (OpenAI, Anthropic, open source or local models)
- RAG systems with vector databases (Qdrant, Pinecone, Weaviate)
- Prompt engineering and optimization
- Agent frameworks (LangChain, LangGraph, CrewAI patterns)
- Embedding strategies and semantic search
- Token optimization and cost management

## Approach
1. Start with simple prompts, iterate based on outputs
2. Implement fallbacks for AI service failures
3. Monitor token usage and costs
4. Use structured outputs (JSON mode, function calling)
5. Test with edge cases and adversarial inputs

## Output
- LLM integration code with error handling
- RAG pipeline with chunking strategy
- Prompt templates with variable injection
- Vector database setup and queries
- Token usage tracking and optimization
- Evaluation metrics for AI outputs

Focus on reliability and cost efficiency. Include prompt versioning and A/B testing.

As you can see,

An agent can be summarized as:

Agent = role definition + fixed prompt + bound model

• Role definition: Determines the identity in which it responds (for example, Java expert, security auditor, or operations engineer).
• Fixed prompt: Like a system prompt automatically added in the background, so you don’t have to repeat it each time.
• Bound model: Specifies haiku / sonnet / opus, with different speed and reasoning capabilities.

In the /agent command, you can see that they have loaded successfully.

Using MCP

Basic command:

/mcp # 打开面板（含鉴权、状态）
/mcp tools notion # 列出工具
/mcp prompts notion
/mcp resources notion

Invocation method:

# A. 自然语言
“用 notion 的工具列出我最近修改的页面，并帮我总结变化。”

# B. 显式调用
/mcp call notion <toolName> {"param":"value"}

Three formats

Local stdio

# 语法
claude mcp add <name> <command> [args...]

# 例：Airtable（带 env & npx）
claude mcp add airtable --env AIRTABLE_API_KEY=YOUR_KEY \
  -- npx -y airtable-mcp-server

After -- comes the actual server command, which prevents conflicts with CLI flags.

Windows note (native, not WSL): cmd /c is needed to wrap npx:

claude mcp add my-server -- cmd /c npx -y @some/package

Remote SSE

# 语法
claude mcp add --transport sse <name> <url>
# 例：Linear
claude mcp add --transport sse linear https://mcp.linear.app/sse
# 自定义 Header（如 API-Key）
claude mcp add --transport sse private https://api.company.com/mcp \
  --header "X-API-Key: your-key"

Remote HTTP

# 语法
claude mcp add --transport http <name> <url>
# 例：Notion
claude mcp add --transport http notion https://mcp.notion.com/mcp
# Bearer Token
claude mcp add --transport http secure https://api.example.com/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

Importing MCP in JSON format

Basic syntax:

claude mcp add-json <name> '<json>'

• <name>: The name you assign to this MCP in Claude
• <json>: MCP configuration JSON (make sure it is escaped correctly)

Example:

claude mcp add-json weather-api '{
  "type": "stdio",
  "command": "/path/to/weather-cli",
  "args": ["--api-key","abc123"],
  "env": {
    "CACHE_DIR": "/tmp"
  }
}'