Mastering Claude Code

Introduction

Four screens are running in parallel. One schedules the day's agenda, another writes a LinkedIn post, the third monitors the team's project progress, and the fourth creates visual materials for a YouTube video. This is the scene an AI automation consultant in the United States encounters every morning when arriving at the office. Running four agents in parallel takes about two minutes. Handling the same tasks by hand takes over twenty-five minutes.

The left panel of this consultant's project folder is packed with dozens of files and folders: context files, decision logs, skill folders, and agent configurations. First-time viewers are easily overwhelmed. Yet this structure was not built overnight. It is the product of building gradually, day by day, through four stages.

Those four stages are called Home, Life, Hands, and Growth.

Stage 1 'Home': Project Folder Structure and CLAUDE.md Design

Just as you prepare land before building a house, creating an AI executive assistant requires first setting up the space where it will operate. Here, 'home' means a single project folder.

Open Visual Studio Code and create a new folder anywhere you like (desktop or elsewhere). You can name it freely, but remember that this folder becomes the entire workspace for the assistant. When you click 'Open Folder' in the left navigator of VS Code and open the folder you just created, a blank canvas unfolds. There are no files yet.

[Figure 21-1] Screen layout when opening a blank project folder in VS Code

Now you must place the first file in this blank space. Click 'New File' in the left panel and enter the filename as CLAUDE.md. This file serves as the brain of the assistant. Before the agent reads the user's message, it loads this file first every time.

What if you cram all information into CLAUDE.md? Every conversation loads vast text, causing the model to waste tokens (the units it counts). So this file contains only core rules and path guidance that says 'what is located where.' Actual detailed information goes in separate files, and CLAUDE.md is designed to function as a router.

[Figure 21-2] Routing architecture diagram of CLAUDE.md

In the starting stage, CLAUDE.md needs only ten to twenty lines. This file will evolve as the assistant grows, so there is no need to perfect it right away. About eighty-seven lines is appropriate, and it should not exceed 150 to 200 lines generously.

When you tell the agent, "This folder is the space where you will work as my executive assistant. Put a brief description in CLAUDE.md," the agent writes the initial content on its own. This is how the foundation of the home is laid.

The agent can also generate the folder structure automatically. Typically, a framework like the following is created.

[Figure 21-3] Standard folder structure for an executive assistant project

You can infer the role of each folder from its name. Context holds information about the user and work, decisions contains important decision records, and projects holds subfolders for each ongoing project. You do not need to memorize this structure. If you ask the agent 'What does this folder do?' it will explain.

Here is one practical tip. If you upload this project to GitHub, you can pull the repository and use the assistant immediately on any device. Backup, rollback, and version control all follow automatically. Of course, you can start with just Git commits on your own computer. GitHub integration can be done later.

Stage 2 'Life': Building Context Through Interview

Stage 2 is the process of bringing furniture into an empty house, filling the refrigerator, and hanging photos on the wall. It is the time when the agent learns about the user.

This process proceeds in interview format. When you provide a pre-prepared onboarding prompt to the agent, it begins asking questions. 'What is your name?' 'What is your role?' 'What is your time zone?' Starting with basic questions, it digs progressively into business details, team composition, current goals, and communication style.

The core principle here is 'honest and complete.' The more thoroughly you answer the agent's questions, the fewer times you will repeat the same context later saying, 'Oh, I should have mentioned this too.' For items you do not know the answer to, you can say 'skip.' You can fill them in later.

The agent's questions cover roughly six areas.

1. Personal information: name, role, time zone, work style. 2. Business and work: company name, business field, main products or services. 3. Team: key people to work with, role assignments. 4. Priorities and goals: this quarter's objectives, urgent tasks, annual direction. 5. Communication preferences: whether you prefer concise answers or detailed explanations. 6. Desired automation tasks: repetitive work, tasks you want to delegate.

When the interview concludes, the agent generates four core files based on your answers.

[Figure 21-4] Interview-based context file generation flowchart

These files are saved inside the context folder. CLAUDE.md is automatically updated to provide path guidance like 'If you need user information, read context/me.md.'

Additional files are also created. In decisions/log.md, major decisions are recorded with date, decision content, reasoning, and context. Each project folder contains a description file for that project. Communication style rules are stored in the .claude/rules/ folder: detailed guidelines like 'do not use em dashes' and 'keep internal communication casual.'

When all these processes are complete, the agent performs an initial Git commit. The first snapshot is saved.

One thing to remember: these files are not fixed documents. If business direction changes, you update work.md. If a new team member joins, you update team.md. When you tell the agent 'Remember this,' the information is reflected in the appropriate file.

Stage 3 'Hands': Creating Your First Skill

Now that you have built the home and breathed life into it, it is time to give the agent hands. This is the stage where the agent can actually do things, not just think.

A skill is reusable instruction. Once written, you can call it anytime with a slash command or natural language, and because it follows the same process each time, results are more consistent. We covered the concept of workflow earlier, and you can think of a skill as another name for workflow. Just as workflows had tools, skills have Python scripts. The essence is the same.

What should you build as your first skill? Integration with project management tools is a practical choice. You put the API key of whatever tool you use (ClickUp, Notion, Asana) in the .env file and configure the skill so the agent can communicate with that tool.

Let us examine the process of creating a skill using a research skill as an example. In plan mode, you tell the agent this: 'Create a research skill for me. I am going to use the Perplexity API. First, create the .env file, and the skill's work is not simple web search but deep research that reflects my business context.'

The agent explores the project structure, and if needed, launches a subagent to analyze the structure, then returns with a plan. That plan includes the following.

[Figure 21-5] Basic structure of a skill: YAML frontmatter and step-by-step instructions

The structure of a skill file is clear. There is a YAML (YAML Ain't Markup Language) frontmatter at the top, followed by step-by-step instructions written in Markdown. The frontmatter contains the skill name and description, and the agent decides which skill to call by reading only this part. The full instructions are loaded only after the skill is selected.

This approach is called progressive context loading, and it is a key mechanism for preventing token waste in the model.

Once a skill is complete, you must test it. When you enter 'Use the research skill to investigate ice cream events in Portland,' the agent reads the skill, calls the Perplexity API, organizes the results, and saves them as a report in the research/ folder. The report includes source links.

Using subagents can reduce costs. The main agent uses the Opus model, but when delegating research to a subagent, you can specify a lighter model like Haiku. The Perplexity API guarantees research quality, and the lighter model handles only organization and summarization, so the structure cuts costs without sacrificing quality.

Subagents are saved as Markdown files in the .claude/agents/ folder. Each has its own context window and operates independently.

Stage Four 'Growth': Evolving the Assistant Through Daily Use

Stage Four requires no separate setup. Using it every day is Stage Four itself.

The AI automation consultant's assistant mentioned earlier did not begin with dozens of skills and rich context baked in. On day one, it had only an empty folder and one line of CLAUDE.md. A month later, that same project looked completely different. Documents had multiplied, decision logs had accumulated, new skills had been added, and the agent's response accuracy had noticeably improved.

The secret to growth lies in the feedback cycle. You call a skill, observe the agent working, and tell it, "This part was good, that part needs fixing." The agent then revises the skill file. In the first few iterations, the output may still feel AI-generated. But after ten, twenty cycles, your preferences and work context become deeply embedded in the skill.

Here are concrete growth strategies.

Migrate existing tools. If you have system prompts saved in ChatGPT custom GPTs or Claude projects, bring those instructions over and convert them into skills. Simply tell the agent, "Turn this prompt into a skill."

Spot repetitive work. If you find yourself giving the same instruction twice or more, it is a candidate for a skill. If you keep telling the agent not to use em dashes, that rule is better captured once in a skill or rule file.

Observe the agent's work process. Watch the agent execute skills the first few times. If it is making unnecessary API calls repeatedly, hardcode frequently used ID values in the skill file to save the model time counting tokens. This loop of observation and improvement accelerates growth.

Add new folders freely. Create a Brand Asset folder and place your logos, fonts, and brand guidelines inside. Your content creation skills can then maintain brand consistency. After adding the folder, tell the agent, "I have added the brand-assets folder. Update CLAUDE.md for me."

[Figure 21-6] The assistant's evolution curve over time: Day 1 vs. Day 30]

You can also use the auto-memory feature. When you say, "Remember that I always prefer X," the agent saves that to an appropriate file and automatically applies it in future conversations.

How It Differs from ChatGPT and Custom GPTs

Saving memory and setting custom prompts in ChatGPT or Claude's web interface does raise productivity. Yet there are limits. The moment you think, "I wish it understood this context too," you are typing out additional explanations. You reach 50% quickly, but getting to 90% is hard.

Let's look at where the differences emerge.

Access to the entire file system. A custom GPT works only within the chat window. A code-based executive assistant, however, can read and write every file in your project folder. me.md, work.md, priorities.md, research reports, decision logs,the agent opens the files it needs, when it needs them. You do not have to paste all your context into the chat window.

Tool execution capability. A custom GPT generates text. An executive assistant runs Python scripts, calls APIs, creates files, and spawns helper agents. Checking your calendar and auto-blocking time, pulling task status from ClickUp,all these actions happen with a single line of conversation.

Persistence and structure of context. As conversations grow long, you hit the context window limit. An executive assistant separates important information into files and stores them permanently. Even if you reset the conversation, the files remain, so the next session picks up where the previous one left off. It is a secretary that never forgets.

Growth potential. A custom GPT's system prompt is static. You must edit it manually. An executive assistant grows richer as you use it,files multiply, skills become more refined, and decision history accumulates. When you tell the agent, "Record this experience," it becomes the foundation for better judgment the next time a similar situation arises.

[Figure 21-7] Comparison table: Traditional ChatGPT approach vs. executive assistant approach]

Closing

Follow the four stages in order, but adjust the pace as you wish. Building the house takes ten minutes; the interview that breathes life into it deserves time and care, and the more you invest, the easier the future becomes. Start your first skill small, then grow it by giving feedback a little at a time, every day.

The real power of this framework emerges when the first skill you built in Stage Three begins to actually run every morning. How that skill,the routine called "Morning Coffee," which reads your calendar, checks project status, and tells you what to do today,transforms your day becomes clear in the story that follows.

Kim Kyung-jin, Attorney and AI Expert

Specialist in AI Law and Policy. Former Member of Congress. Author of numerous works.

If this book has remained at your side for even a moment, please support the next story coming into the world.

(Voluntary support account: NH Bank 302-1096-0948-81 Account holder: Kim Kyung-jin)