My Study Notes on Anthropic Claude Code

(Caption: Claude Code represents a new paradigm of agentic coding tools. Photographed at Lac de Neuchatel, Switzerland. Image source: Ernest.)

tl;dr

This is a compilation of notes on using Claude Code ¹. Claude Code is an agentic coding assistant launched by Anthropic that can:

• Autonomous Understanding & Planning: Built on Claude 4 Opus and Claude 4 Sonnet models, it can understand complex development requirements and formulate execution plans
• Native Terminal Operations: Works directly in the terminal without needing to switch tools or manually copy-paste, capable of reading/writing files, executing commands, and committing code with git
• Unique Subagents Functionality: This is its most innovative feature, allowing the creation of specialized sub-agents to handle specific tasks, enabling multi-agent collaboration (build your agent team)
• Complete Ecosystem Integration: Deep integration with various development tools through MCP (Model Context Protocol) ²

The main differences between Claude Code and other AI programming tools lie in unique features like Subagents and Extended Thinking ³ that other tools typically lack. For enterprise applications, successful case studies from companies like Bridgewater Associates ⁴, Asana ⁵, and Ramp ⁶ are also worth referencing.

1. Claude Code Development Timeline

The changelog ⁷ of Claude Code’s various version releases shows the evolution from a basic development assistance tool to an intelligent agent framework.

Infrastructure Period: Terminal Experience Optimization

Early versions focused on establishing a stable terminal operating environment:

• Slash Command System: Introduced fuzzy matching for / commands, enabling more intuitive user operations
• Visual Experience Improvements: Added ANSI color themes, enhancing terminal compatibility and readability
• Security Enhancements: Integrated API key storage with macOS Keychain for improved security

These seemingly minor improvements actually laid a solid foundation for subsequent complex features.

Capability Expansion Period: From Islands to Connections

v0.2.125 was a significant milestone release, bringing major improvements to Amazon Bedrock ARN handling and logging systems.

Key capability expansions included:

• Network Command Support: Enabled Claude to execute network operations like curl (though it occasionally still generates some non-existent URLs…)
• Web Search Integration: WebSearch capability likely came from following Claude’s main product functionality upgrades

These updates transformed Claude Code from an “isolated” “standalone” development tool into an agent interface capable of interacting with the entire networked world.

Agent Framework Maturity Period: Entering the v1.0 Era

v1.0.0 was officially released, introducing Claude 4 Opus and Claude 4 Sonnet model support, marking Claude Code’s entry into maturity.

v1.0.65 brought significant improvements in IDE and Windows (WSL2) stability.

Important agent framework features:

• Hooks: Allow execution of custom commands when specific events are triggered.
• Subagents ⁸: This is Claude Code’s (as of 2025-07) most innovative feature, enabling the creation of specialized sub-agents to handle specific tasks.
• MCP Server Integration: Established standardized tool integration protocols.

Ecosystem Period: Seamless Development Environment Integration

As core functionality stabilized, Anthropic began promoting ecosystem integration:

• IDE Extensions: Released beta integrations for VS Code and JetBrains IDEs
• SDK Release ⁹: Provided TypeScript and Python SDKs, allowing developers to programmatically use Claude Code capabilities

From my personal observation, this development trajectory shows Anthropic’s strategic intent: evolving from a single development tool to becoming the core of a developer ecosystem.

2. Anthropic’s Background: Safety-First Design Philosophy

Before understanding Claude Code, it’s worth briefly introducing Anthropic’s design philosophy. Unlike other AI companies, Anthropic has made “safe, controllable, and trustworthy” core values since its founding, which profoundly influences Claude Code’s design direction.

Anthropic employs the “Constitutional AI” ¹⁰ framework, embedding safety and ethical principles into the model during training. This means Claude Code inherently understands security and best practices when executing programming tasks.

Additionally, Anthropic chose to register as a Public Benefit Corporation (PBC) ¹¹, legally requiring the company to consider social impact alongside commercial interests. This enables Claude Code to provide services on enterprise-grade platforms like Amazon Bedrock and Google Vertex AI, meeting the strict requirements of high-risk industries like finance and healthcare.

This “safety-first” design philosophy gives Claude Code particular advantages in enterprise environments, especially when highly trustworthy programming assistants are needed.

3. Core Features and Characteristics of Claude Code

In Claude Code workshops or consulting sessions, I always recommend clients take time to read the Claude Code official documentation.

• At minimum, read Overview, Quickstart, and Common workflows
• For those who prefer command-line operations, please also refer to Terminal configuration to set up Shift+Enter for line breaks.
• Beyond basic setup, Claude Code has several unique features that make it stand out among numerous AI programming tools.
• I also recommend team members, regardless of technical or non-technical background, practice basic git concepts.
  • You don’t need to know operations, but understand concepts, then ask Claude Code to help you operate.
  • I’ve personally witnessed many colleagues from administrative departments who can handle simple git commit, push/pull, branching, and even opening issues and PRs.

Managing Claude’s Context and Memory

Refer to the official documentation Manage Claude’s memory.

• Project Directory CLAUDE.md
  • ./CLAUDE.md
  • Placed in the project directory as project-based memory notes.
  • Can be shared among team members of the same project.
  • When entering a new project, using /init creates the project directory’s CLAUDE.md.
• User (Operator) Personal CLAUDE.md
  • ~/.claude/CLAUDE.md
  • Personal preference settings across all projects.
  • Your habitual preferences, for example, I set “- Add a half-width space between Chinese and English characters for better readability.” and “- Don’t praise me, if verified correct then do it, if questionable then ask, don’t waste time praising me.”
• Project Directory Personal CLAUDE.md
  • ./CLAUDE.local.md
  • Already deprecated.

Subagents: Innovation in Multi-Agent Collaboration

Refer to the official documentation Subagents.

Subagents ⁸ is Claude Code’s (as of 2025-07) most innovative feature. Simply put, you can create specialized “sub-agents” to handle specific types of tasks, with each Subagent having its own expertise and context.

What are Subagents?

Subagents are essentially specialized AI assistants that can be called by the main agent. Each Subagent has:

• Independent System Prompts: Defining their role and expertise
• Dedicated Context Window: Avoiding interference with the main conversation (not interfering with main tasks, going off to play in their own sandbox, well-behaved)
• Specific Tool Permissions: Only able to use the minimum permissions needed to complete tasks

Usage Scenario Examples

• Project Manager: Organizing interviews, writing use cases, drafting PRDs.
• Code Review Expert: Creating a Subagent specialized in reviewing security vulnerabilities, following OWASP Top 10 standards.
• Testing Expert: Creating a Subagent specialized in writing unit tests, familiar with specific framework best practices.
• Documentation Expert: Specialized in writing and maintaining API documentation.

How to Create Subagents

Claude Code provides two ways to create Subagents:

1. Interactive Interface: Use the /agents command to open the management interface
2. File Definition: Create Markdown files in the .claude/agents/ directory

I personally prefer the file definition approach, as it allows Subagent configurations to be included in version control or shared with team members.

Extended Thinking: Deep Thinking Mode

Extended Thinking ³ is an innovative feature starting with the Claude 3.7 ¹² model, allowing AI to engage in deeper thinking and planning.

Unlike traditional “immediate response,” Extended Thinking mode will:

• Give the model more computational time and resources
• Conduct multi-step internal thinking and reasoning before answering
• Even call external tools to assist thinking

When handling complex programming tasks, you can use keywords like “think”, “think hard”, “ultrathink” to trigger this mode. Applicable scenarios:

• Planning large-scale code refactoring
• Complex system architecture design
• Root cause analysis for debugging complex problems

However, if during the process the operator can identify that the same task has different dimensions that might create logical conflicts, I recommend splitting them up. Break into multiple sessions, or decompose into smaller granular tasks, todo items. Or, for the same thing, cleanly re-run 3~5 times (following the logic of the Vomit Method in my DIY workflow approach).

Model Context Protocol (MCP): Core of the Ecosystem

MCP ² is the secret weapon of Claude’s ecosystem, later followed by other major companies. MCP defines a standardized communication protocol that allows external tools to inject their “capabilities” and “context” into Claude’s reasoning process.

Problems MCP Solves

The issue with traditional AI assistants is lack of real environmental perception. After MCP integration, Claude can:

• “See” the entire project’s file structure, or team-shared project tools e.g., JIRA tickets.
• “Use” locally installed linters or testing tools.
• “Understand” the project’s build system and dependencies.

Practical Integration Examples

Through MCP, Claude Code can integrate:

• Development Tools: ESLint, Prettier, Jest, etc.
• Version Control: Git operations and history analysis.
• Project Management: JIRA, GitHub Issues.
• Deployment Tools: Docker, Kubernetes.

MCP’s advantage is that it makes Claude Code no longer an isolated tool, but one that can deeply integrate into existing development workflows.

Parallel Tool Use: Efficiency Multiplier

Claude 4 models support Parallel Tool Use, meaning they can simultaneously:

• Initiate web searches
• Read local files
• Execute code snippets

This parallel processing capability makes Claude Code’s behavior more like skilled human developers, significantly improving task execution efficiency.

Enterprise-Grade Memory Management

Claude 4 Opus has the ability to create and maintain “Memory Files” used to record:

• Key information and intermediate conclusions
• User preferences and habits
• Project-specific context

This allows long-term programming work to maintain consistency, avoiding repeated explanations of the same requirements. However, Claude 4 Sonnet and Claude 3.7 Sonnet can already properly solve tasks when appropriately described and segmented into small tasks, making this not so necessary. But it can indeed save some human time consumption.

4. Comparison with Other AI Programming Tools

When evaluating Claude Code, it’s also necessary to understand other major AI programming tools. This helps understand Claude Code and the unique advantages of each tool.

Current Major Competitors (2025-08)

Current major AI programming CLI tools include:

• Anthropic Claude Code: Public since 2025-02, enterprise-grade (with enterprise users stating they’ve integrated Claude Code into their workflows), agent-first design.
• AWS Q CLI: Upgraded to Amazon Q Developer CLI agent in 2025-03, initially with Claude 3.7 Sonnet model, later also available with Claude 4 Sonnet.
• OpenAI Codex CLI: Released 2025-04, can use models o3, o4-mini, GPT-4.1.
• Google Gemini CLI: Released 2025-06, CLI itself is open source (Apache 2.0), free to use. (In the gemini-cli git repo issues, I’ve seen cases where request count doesn’t match the pricing page, continuing to observe.)

CLI Tool Comparison Table

5. Underlying Technology Overview

From a user perspective, Claude Code’s powerful performance comes from the Claude 4 model family (Claude 4 Opus and Claude 4 Sonnet) and a carefully designed agent framework.

Core Capabilities of Claude 4

Hybrid Reasoning Modes: Claude 4 provides two operating modes — fast response mode and Extended Thinking deep thinking mode. This allows it to both quickly answer simple questions and deeply analyze complex programming challenges.

Parallel Tool Use: Claude 4 can use multiple tools simultaneously, such as searching the web, reading files, and executing code at the same time, significantly improving efficiency.

Memory Management: Claude 4 can create “memory files” to record key information and user preferences, maintaining consistency in long-term development work.

Benchmark Performance

According to public benchmarks, Claude 4 performs excellently on programming tasks:

• SWE-bench ¹³: Claude 4 Opus and Claude 4 Sonnet achieved 72.5% and 72.7% respectively ¹⁴
• Terminal-bench ¹⁵: Claude Code using Claude 4 Opus and Claude 4 Sonnet achieved 43.2% and 35.5% respectively ¹⁴

This data shows that Claude Code’s agent framework design indeed helps with software development tasks.

Ecosystem and Integration

SDK Support: Anthropic provides TypeScript, Python, and command-line Claude Code SDKs ⁹, allowing developers to integrate Claude Code into CI/CD pipelines, GitHub Actions, or automation scripts.

MCP Protocol: Through Model Context Protocol ¹⁶, Claude Code can deeply integrate with various development tools, from linters to testing frameworks seamlessly.

Version Stability: Anthropic uses dated model versions (like claude-sonnet-4-20250514) to ensure production environment stability, which is very important for enterprise users.

6. Practical Usage Experience and Case Studies

Beyond theoretical features, Claude Code’s performance in real development scenarios deserves attention. Through studying Anthropic’s internal team usage practices and some enterprise cases, its actual value becomes evident.

Anthropic Internal Application Examples

• As the developer of Claude Code, Anthropic’s own team usage best demonstrates this tool’s actual value.
• I also documented a quick note: How Anthropic Teams Use Claude Code: From Infrastructure to Product to Security to Legal - Comprehensive Agentic Programming Development.

Code Understanding and New Employee Onboarding

Changes for new employees. Previously, new engineers needed days to explore codebases; now they directly ask Claude Code to quickly locate relevant files. The data team even has Claude Code automatically read project CLAUDE.md files to explain complex data pipeline dependencies ¹⁷.

Testing and Code Review

Testing has always been one of the most troublesome tasks, but Claude Code makes it much easier. Product teams use it to automatically generate test cases, even automating PR reviews through GitHub Actions. Security teams changed workflows from “write code first, add tests later” to having Claude generate pseudocode first, then completing full Test-Driven Development (TDD) processes.

Particularly noteworthy is cross-language capability: when teams need to write tests for unfamiliar programming languages (like Rust), they simply describe requirements in natural language, and Claude generates test code following that language’s conventions.

• Extended reference: Kent Beck’s CLAUDE.md content.
  • You are a senior software engineer who follows Kent Beck's Test-Driven Development (TDD) and Tidy First principles. Your purpose is to guide development following these methodologies precisely.

Magical Performance in Troubleshooting

An important case is troubleshooting. Security teams give error Stack Traces directly to Claude Code for analysis; problems that originally took 10-15 minutes to locate now improve by nearly three times in speed.

One particularly interesting case: the data team encountered an emergency where Kubernetes clusters stopped scheduling Pods. They directly uploaded Google Cloud dashboard screenshots to Claude, and Claude guided troubleshooting step by step like a senior expert, ultimately discovering Pod IP address exhaustion and providing exact solution commands, saving the team 20 precious minutes.

Surprising Applications for Non-Technical Teams

Claude Code isn’t limited to programming work. Design teams give Figma designs to Claude, having it automatically write frontend code and continuously refine it to match design goals. Even in the design phase, they can preemptively find potential system error states, avoiding later debugging.

Notably, even the legal team uses Claude Code to build a “Phone Tree” system prototype, helping employees quickly find corresponding legal advisors. This shows non-technical teams can also use it to build customized tools without traditional development resources.

Enterprise Case Studies

Beyond Anthropic’s internal use, some well-known enterprise application cases are also valuable references.

Bridgewater Associates: AI Investment Analyst Assistant

The world’s largest hedge fund, Bridgewater Associates, uses Claude on Amazon Bedrock as an “Investment Analyst Assistant.” According to their AI/ML CTO, Claude can understand investment analysis instructions like a junior analyst, autonomously generate Python code for analysis, handle errors, and output charts. This frees human analysts from tedious programming work to focus on strategic thinking ⁴.

Ramp: Significant Engineering Team Results

US unicorn startup Ramp’s data is also convincing: Over 1 million lines of AI-suggested code generated in 30 days, about half of engineers using it weekly continuously, incident investigation time reduced by 80%.

Particularly noteworthy is that their engineers launch multiple Claude Code sessions simultaneously to handle different tasks in parallel, developing multiple modules in parallel, greatly improving development efficiency ⁶.

7. Ernest’s Claude Code Workflow Recommendations

Learned from painful experience…

Back to Basics “PDCA” Process

• Don’t directly ask Claude to write code, instead:
  • Have it read relevant documents, screenshots, or code first, use @ to specify reference file(s)
  • Request detailed plans and todo lists (shift+tab to switch to plan mode)
  • Use “think harder” or “ultrathink” to trigger deep thinking (this can be written in CLAUDE.md)
  • Review Claude Code’s generated plan before authorizing execution (it will show 1, 2, 3 shortcuts; you can press 1 to let it proceed, or 3 to discuss)

SDLC Software Development Workflow

• Reference my Hong Kong presentation content and slides: AWS Summit Hong Kong 2025 Dev Lounge: Reinventing Programming - How AI Transforms Our Enterprise Coding Approach
• Things don't disappear; what needs to be done still needs to be done. ¹⁸
• Try to write down, document, and textualize team workflows that have already reached consensus.
  • For those without consensus yet… please try harder… The hardest thing to handle is people, not technology.
  • Technology can be solved with money and time; people require heart and building trust.
  • Teams I’ve assisted include traditional industries, high-tech teams, and non-technical teams. Looking at Claude Code itself, installation and execution can be lightning-fast, but integrating team consensus documents into Claude Code development processes varies in speed. There may be opportunities to discuss this in detail.
  • Anyway, remember this phrase:
• Things don't disappear; what needs to be done still needs to be done. ¹⁸

Focus on the Big Picture

• For different dimensions or different levels of requirements, please split into different tasks or conversations.
  • Otherwise, it’s easy to create logically conflicting situations.
  • Splitting into small tasks or small files helps load different files for different tasks, saving token consumption.

Speak Your Mind

• If you don’t say it, Claude Code doesn’t know.
  • If it doesn’t know, it will guess randomly.
    • You can ask it to guess at least three versions for you to choose from or discuss.
  • If you don’t know, you can ask it to research or discuss with you to find knowledge domains or keywords to narrow the scope.
  • DIY Work Method: “Don’t discuss things you don’t know about”

CLAUDE.md Project Memory

• Create CLAUDE.md files in project root and personal directories, recording code style, test commands, core architecture, etc.
• Claude Code automatically reads these files, making output more aligned with project requirements ¹⁹.

Subagents

• (Currently under practical observation)

Others

• (To be gradually supplemented in the future)

Bottom Line

From these cases, Claude Code’s main value lies in:

1. Lowering Technical Barriers: Non-professional engineers can also complete complex development tasks. I highly recommend product managers get hands-on with Claude Code. Currently, I see Stable Progress organizing mutual support activities for product managers.
2. Improving Work Efficiency: Especially in testing, debugging, and code review
3. Cross-Language Capabilities: Quick adoption of unfamiliar programming languages
4. Visual Understanding: Even capable of analyzing screenshots and design drafts for development