AI BOARD
AI Board
AI tools, policy moves, and public decisions are tracked in plain language.
Complete Mastery of Claude Code
Complete Mastery of Claude Code
Beginner -> Practice -> Advanced, a systematic master guide
Original video: Silvel Developer YouTube | Organized and supplemented on: 2026-03-26
https://www.threads.com/@choi.openai/post/DWROhpiDx7M?xmt=AQF0-nUMjKeESuRnT-IRv-2ogXzl_yjdNtOJ7O30-VnqoQ
Part 1: Beginner Section, Build a Solid Foundation
1. What Is Claude Code?
Claude Code is an AI coding assistant made by Anthropic. It lives inside your computer terminal (command window), understands the file structure of your project, and writes actual code when you talk to it in natural language, whether Korean or English. Since its official release in May 2025, it has become one of the most widely adopted AI development tools.
🎒 Analogy Even Elementary Students Can Understand
Imagine a very smart friend who helps with homework. If you say, "Solve math problem 5," the friend first reads through the textbook, makes a solution plan, and writes the answer. Claude Code is just like that friend, except it creates computer programs instead of doing homework.
📰 Real Case: Vibe Coding by a Liberal Arts Reporter
According to ZDNet Korea, a reporter who had studied only liberal arts from high school through college completed a news monitoring program by doing nothing but conversing with AI, without knowing a single line of code. You do not memorize code. You say what you want done.
🏢 Enterprise Adoption Results
Verified results have been reported by global companies, including Fountain (50% faster development), Rakuten (7 hours of autonomous coding), CRED (2x development speed), and TELUS (500,000 hours saved).
2. Start from the Project Root Directory
The most important rule when starting Claude Code is to run it from the top-level folder of the project (root directory). Claude searches files based on the current location, so if you run it from a subfolder such as apps/api, it cannot understand the whole project structure and works inefficiently.
🎒 Analogy Even Elementary Students Can Understand
When looking for a book in a library, if you start at the first-floor information desk, you can see the map of the entire library. But if you start in the science corner on the third floor, you see only the books in that corner. Root directory = the first-floor information desk.
3. CLAUDE.md: The Project Manual for AI
claude init creates a CLAUDE.md file. This file is a 'manual' containing all project context, and Claude automatically reads it at the start of every conversation. One blogger described it this way: "Thanks to CLAUDE.md, I no longer have to repeat the same explanation to Claude every time."
Six things you must put in CLAUDE.md:
•Absolute rules : "Never query the production DB," "Do not commit .env files"
•Architecture : Organize the project structure in tree format
•Tech stack : Languages, frameworks, and libraries used
•Build/test commands : npm run build, pytest, and so on
•Coding conventions : "Use only ES modules," "Functional components + Hooks," and so on
•Domain context : Business logic (for example, product color/size/stock rules)
🎒 Analogy Even Elementary Students Can Understand
Imagine a new cook came to your house. If you hand over a memo listing where the refrigerator is, forbidden ingredients (allergies), and favorite flavors, that cook can make good food right away without asking every time. CLAUDE.md is that memo.
💡 Practical Case: E-Commerce Project
The lecture used an e-commerce platform as an example. Business rules such as "If a product's stock is 0, show a sold-out badge" and "Standardize payment API errors with the AppError class" were written in CLAUDE.md, and Claude coded while following those rules without asking every time.
Key tip: Keep CLAUDE.md within 300 lines. If it grows, split it by folder (for example, CLAUDE.md for API, CLAUDE.md for web, CLAUDE.md for DB). You do not need to update it manually. If you command, "Add the pattern we just decided to CLAUDE.md," it will be reflected automatically.
4. Trigger Keywords: Repeat Work with One Word
If you register trigger keywords in CLAUDE.md, several steps run automatically when you enter a specific word.
💡 Practical Case: The Magic of Saying "Deploy it"
When you type "Deploy it": ① run tests -> ② stop/report if they fail -> ③ if they pass, git add + commit -> ④ push + print summary. The four commands you used to type every time are finished with one phrase. Set this up when the repeated work of committing and pushing feels tedious, and productivity rises sharply.
5. Team Sharing: Compound Engineering
At Anthropic, the practice of an entire team sharing the same CLAUDE.md and using Claude under the same rules is called compound engineering. When one team member updates the architecture, Claude automatically has the latest information when another team member runs git pull.
🎒 Analogy Even Elementary Students Can Understand
If a whole class puts the same rule sheet on the wall, everyone can see a new rule as soon as one person posts it. Sharing CLAUDE.md through Git works the same way.
Part 2: Keyboard Shortcuts and Navigation
6. Shift+Tab: Plan Mode vs Edit Mode (The Most Important Shortcut)
Claude Code has two modes. In Edit Mode, Claude immediately creates and edits files. In Plan Mode, it does not write code. It only plans how to do the work and asks for review. Shift+Tab switches between them.
Why start with Plan Mode first? Three reasons:
①Token optimization : If wrong code is generated and then fixed, tokens are wasted 2 to 3 times over. Planning first prevents unnecessary code generation.
②Preserve the context window : If coding begins after all exploration information has piled up, the bag is already full. With planning only, you can start from a clean state.
③Improve execution accuracy : You get a chance to review the plan and adjust direction. This is Anthropic's official recommendation.
🎒 Analogy Even Elementary Students Can Understand
If you start a drawing directly with paint, mistakes are hard to fix. If you draw a pencil sketch first, you can revise it until you like it and then add color. Plan Mode = sketch, Edit Mode = coloring.
⚠️ Failure Case: What Happens When You Run Without a Plan
The lecture warns: "If you run it immediately without a plan, Claude may mass-edit the code in the wrong direction, causing a disaster." Context is wasted, the usage limit is wasted, and time, money, and resources are wasted.
7. Escape: Stop Work and Rewind
Escape Press once = immediately stop the current task ("Interrupted"). Escape Press twice = activate the Rewind UI to go back to the previous prompt. If it is heading in the wrong direction, you can return at any time.
🎒 Analogy Even Elementary Students Can Understand
It is like the "pause" button in a game (Escape once) and "return to the previous save point" (Escape twice). If you are attacking the boss with the wrong strategy, you can stop, go back to the earlier point, and start again.
8. Other Useful Shortcuts and Tips
•Drag and drop screenshots : Dragging an image into the terminal is more accurate and saves more tokens than describing UI design in words. Architecture diagrams are also best given as images.
•Exclamation mark (!) command : You can run Bash commands such as !git status, !npm run build without leaving the Claude terminal. The instructor calls this "my favorite tip."
•iTerm2 window split — Cmd+D (vertical split), Cmd+Shift+D (horizontal split) to run multiple Claude sessions at the same time. Use Cmd+[/Cmd+] to switch windows.
Part 3: Complete Guide to Essential Slash Commands
| Command | Function | Easy analogy | Practical tip |
|---|---|---|---|
| /clear | Reset the entire context | Wipe the board clean | Use it whenever one feature is finished |
| /compact | Clean up while leaving only the summary | Compress the notes, keeping only the essentials | When cleaning up while preserving important context |
| /context | Check token usage | Check remaining space in the bag | Consider /compact if it exceeds 60% |
| /model | Switch AI models | Change teachers by difficulty level | Sonnet recommended for parallel work |
| /resume | Restore a previous session | Continue yesterday's game | claude --resume also works outside the terminal |
| /mcp | Manage MCP servers | Check the toolbox | Save tokens by disabling unused MCPs |
| /export | Export conversation | Back up a KakaoTalk conversation | Useful when asking another AI for review |
| /output-style | Change output style | Switch between explanation/training/default modes | Use training mode if you want to learn coding |
| /memory | Automatic memory management | AI's personal diary | Automatically save and load from MEMORY.md |
| /voice | Voice input mode | Give commands by speaking | Supports 20+ languages, including Korean |
9. Model Selection Strategy: How to Save Your Bullets
Opus 4.6 = genius professor. Smartest but most expensive. Good for complex architecture design.
Sonnet = capable practitioner. Good for everyday coding work. Recommended for parallel work.
Haiku = sprinter. Fastest and cheapest. Use it for code exploration and basic tests.
The instructor's strategy: use Opus for the main work, and use Sonnet/Haiku when delegating parallel work to sub-agents. "Tokens are like bullets. If you use them all, you cannot do the next task," so they must be distributed strategically.
10. Create Custom Slash Commands
.claude/commands/ folder, create ship.md, and you can run tests, commit, and push all at once with /ship. It is more stable than trigger keywords because Claude recognizes the structure, and it also supports tab completion.
💡 Examples of Popular Custom Commands in the Community
/feature-spec : Automatically generate user stories and test strategies
/refactor-plan : Identify code smells and create a refactoring plan
/api-endpoint : Generate routes, controllers, validation, tests, and documentation at once
/review : Automatically apply a code review checklist
Part 4: Practice Section, Context Management and Workflows
11. Context Is King: Why Claude Becomes Dumb
🎒 Analogy Even Elementary Students Can Understand: Practice Notebook (Scratchpad)
Imagine solving a complex math problem. All the formulas you learned are in your head (AI's training data), but when solving the actual problem, you write numbers and steps in a practice notebook. This notebook is the 'context window.' When the notebook is full, you must erase what you wrote earlier, and you cannot remember what you erased.
Claude's context window is 200,000 tokens (about 100,000 Korean characters). It looks like a lot, but in practice it fills quickly. Industry-tested thresholds:
•0~50% : Work freely (comfortable state)
•50~70% : Caution needed (check /context periodically)
•70~90% : /compact required (precision begins to drop)
•90%+ : /clear required (hallucinations and errors surge)
12. Second Brain and /memory
The practice of saving patterns, solutions, and reasons for decisions learned while working with Claude is the second brain. /memory manages it automatically, and it is loaded automatically at the start of each session. One user reported: "The context catch-up that took 45 minutes every morning dropped to 5 minutes."
Practical formula: personal memory -> /memory (MEMORY.md) | team shared knowledge -> CLAUDE.md
13. Layered Context Strategy (Lazy Loading)
❌ Bad Example vs ✅ Good Example
Bad example: Put all 50 API endpoints and 40 DB table schemas into CLAUDE.md -> thousands of tokens wasted every time (less than 5% is actually needed)
Good example: At the bottom of CLAUDE.md, record only paths such as "See: docs/api_spec.md, docs/db_schema.md" -> Claude reads the relevant file only when needed
🎒 Analogy Even Elementary Students Can Understand
If you carried every library book in your bag, your back would hurt. Instead, carry only the book list. When you think, "I need the math book!" go and take it out then. That is Lazy Loading: load only when needed.
14. Golden Rule: One Session = One Feature
The most practical rule: One feature, one task per session. Break work down into units such as "Implement the Stripe webhook handler," not "Build the entire payment system."
15. Offload Heavy Work to Scripts
💡 Practical Case: Migrating a 100,000-Row CSV
If Claude directly processes a 100,000-row CSV inside the conversation, the context gets polluted. Instead: ① ask Claude, "Write a CSV parsing script" -> ② run the script -> ③ pass only the result summary (JSON) to Claude -> ④ Claude proceeds with the next task using only the summary. Claude does not need to read 100,000 rows and can keep a clean context.
16. Pro Tips: Habits That Make a Difference in Practice
•Read Thinking : Do not ignore Claude's thinking process. If an assumption such as "This function seems to do X" is wrong, stop immediately with Escape. Code built on a wrong assumption is all useless.
•Get critique from another AI : Export the conversation with /export and ask ChatGPT or Gemini, "Point out anything I am missing or any wrong approach." Each AI finds different problems from a different angle.
•Paste the entire error log : Do not interpret it. Send the original text as is. Claude is excellent at stack trace analysis. If you interpret it yourself, information may be omitted or distorted.
•Make TDD a habit : Repeat small change -> write test -> confirm pass -> commit. Verified numbers: applying TDD reduces production bugs by 70% and reaches 90% test coverage (40% without it).
•Local TODO file : Manage todo.md from the beginning to the end of the project. Even when sessions change, AI can understand the progress.
•Mermaid diagrams : If you organize architecture as diagrams rather than text, Claude understands it much faster. Keep them organized by feature in docs/architecture.md.
•Ban "do you remember" : Do not start with "Do you remember?" Claude can access the full context inside the session. Refer directly, as in "Continue refactoring auth.ts."
17. WAT Framework: AI Project Management Methodology
Workflow + Agents + Tools. The key is to separate AI reasoning from code execution.
💡 Practical Case: Adding a Blog Comment Feature (Applying WAT)
W (Workflow): "1. Design comments table schema -> 2. Implement API endpoint -> 3. Frontend UI -> 4. Check tests at each step"
A (Agent): While the API is being implemented, another sub-agent designs tests in advance. On error, self-healing automatically recovers.
T (Tool): DB migration with scripts/migrate.sh, PR creation with MCP GitHub, and automatic tests on every commit with hooks.
Part 5: Advanced Section, Skills, Sub-agents, and Hooks
18. Skills: Work Manuals for AI
🎒 Analogy Even Elementary Students Can Understand
When a new part-time worker joins a company, explaining "Do this this way, do that that way" by mouth every time leads to uneven results. But if you create a work manual, the quality stays the same no matter who does it. Skills = AI's work manuals. Write it in one skill.md file, and that is enough.
Key difference between prompts and skills:
•Prompt = one-time instruction (entered every time, uneven results, disposable)
•Skill = reusable work system (automatic trigger, consistent quality, shareable with the team)
The secret weapon of skills: two-stage loading:
Normally, only the name and one-line description (50 to 100 bytes) are loaded into context, and the full content is loaded only when called. Unlike CLAUDE.md, which is fully loaded every time, even if you create 20 skills, the context burden is almost zero. This is why the instructor says, "Honestly, the real power of Claude Code comes not from CLAUDE.md, but from the skill system."
💡 Practical Demo: PPT Generation Skill
In the lecture, the instructor actually created a PPT skill. When you type "Create a PPT," the skill is triggered automatically and completes everything at once: enter presentation topic, audience, and duration -> plan slides -> create a PPTX file with python-pptx -> check the validation checklist. It does not need to be perfect at first. The key is to refine it gradually as you use it.
📰 Non-Developer Use Case: Marketer Automation
One marketer used skills to build an automation that handles a Tally form submission -> add to the Stibee address book -> add Solapi contact -> automatically classify job group, all at once. They reportedly completed it by asking Claude step by step about API connection order and field mapping.
19. Sub-agents: Specialist Helpers with Separate Context
🎒 Analogy Even Elementary Students Can Understand
Imagine a school group project. If the class president does everything alone, it takes a long time. But if one person handles the drawings, one person handles the writing, and one person prepares the presentation, it finishes three times faster. On top of that, each person works in their own notebook, a separate context, so the class president's notebook does not get messy.
Four key advantages:
①Parallel processing - Run multiple tasks at the same time. Industry reports suggest a 35-40% reduction in work time.
②Context protection - Even after searching 100 files, only a summary is passed back to the main session
③Specialization - Agents dedicated to testing, frontend, and backend work
④Reuse - Once created, it can be used across multiple projects and shared with a team
Four practical patterns:
•Isolation pattern - Perform heavy analysis in a separate space and return only the result
•Parallel pattern - Three agents run security checks + performance measurement + test coverage at the same time
•Chain pattern - Pass work sequentially: analysis -> refactor -> test -> review
•Resume pattern - Continue work that was interrupted
💡 Decision Rule: Main vs. Subagent
A plain analogy is, "Will you do it yourself, or send someone on an errand?" Decide with two questions: ① Is the task independent? -> If Yes, use a subagent. ② Are you worried about context pollution? -> If Yes, use a subagent. If both are No, handling it directly in the main session is faster.
20. Agent Teams: An AI Team Whose Members Communicate with Each Other
This is an evolved form of subagents. Subagents communicate only with the main agent, but in Agent Teams, team members can send messages directly to each other. Like a real development team, frontend, backend, and testing roles can work at the same time while communicating, for example, "The API spec changed, please update the frontend."
21. Hooks: The Automation Engine
🎒 An Analogy Even an Elementary Student Can Understand
If you hang a bell on the front door, it automatically goes "ding-dong!" whenever someone opens the door. Even if nobody presses a separate button, the event "the door opens" triggers it automatically. Hooks are exactly this. Unlike skills, which are probabilistic, hooks run 100% of the time with no exception.
Event flow:
Pre(input validation) -> tool execution -> Post(lint/format post-processing) -> Notification -> Stop(report generation)
💡 Practical Demo: Work Completion Notification Hook
In the lecture, we directly built "a hook that plays a macOS notification and sound effect when Claude enters a waiting-for-response state." If you tell Claude, "Make this kind of hook," it is automatically registered in settings.json. After that, whenever Claude finishes a long task, you get a notification and can come right back even while doing something else. Integration with Slack or Google Chat is also possible.
Caution: Claude stops and waits while a hook runs. Put heavy work in the background, and always set a timeout.
22. Skills vs. MCP vs. Subagents: When to Use What?
| Item | Burden when not in use | Efficiency during execution | Suitable situations | Easy analogy |
|---|---|---|---|---|
| Skill | Lightest | Normal | Repeated tasks, enforcing format | Cooking recipe book |
| MCP | Heavy | AI calls it automatically | External service integration | Universal remote control |
| Subagent | Heavy | Most efficient | Independent parallel work | Specialist errand-runner |
Part 6: Advanced Tips and Productivity Gains
23. Multi-Instance + Worktree
Open four terminal tabs, assign each a color and name for frontend (green), backend (blue), DB (purple), and testing (red), and work on them at the same time. To prevent conflicts in the same directory, use Worktree to create separate folders by branch, so you can keep multiple branches open at once without checking out back and forth.
24. Voice Input: The Era of Coding by Speaking
/voiceto activate it. Hold the spacebar, speak, and release to convert speech into text. It supports 20+ languages, including Korean. At the lecturer's company, "coding with AI by speaking" is already the standard way of working, and this trend is spreading.
25. Creating a Custom MCP: Gemini Plan Review
Off-the-shelf MCPs, such as Notion and Linear, take up a large amount of context just from their tool descriptions. In the lecture, we directly create a 'plan review MCP' using the Gemini API: Claude sends its plan to Gemini for review, then combines both AIs' perspectives to create a sturdier plan.
Part 7: Practical Integration Demo - Processing 10 Pieces of Feedback in 30 Minutes
This is the highlight of the lecture. In a real project, a Coupang Rocket Growth settlement analysis app, 10 pieces of client feedback are processed at once by combining every feature:
Step 1.Voice inputdelivers 10 pieces of feedback in 30 seconds
Step 2.Plan Modesets up the work plan
Step 3.Custom MCPsends the plan to Gemini for review
Step 4.Three subagents run in parallel (backend + frontend + testing)
Step 5.Hooksends a work-completion notification -> you come back after coffee
Step 6.Code review plugin (runs five agents in parallel internally)
Step 7.Custom skillautomatically creates the client report
Result: half a day (4-5 hours) -> about 30 minutes
From code edits to the client report, everything is handled in one flow. The lecturer stresses, "This is the true full workflow of a solo developer."
Key Summary: Transitioning into an Agentic Engineer
Beginner section: CLAUDE.md setup + shortcuts (Plan first with Shift+Tab!) + essential slash commands
Practical section: context management (Lazy Loading, one session = one feature, script offloading) + workflow (Plan -> TDD -> TODO)
Advanced section: skills (reusable work systems) + subagents (parallelization and isolation) + hooks (event automation) + custom MCPs (tool expansion)
The core message of this series: expand AI's capabilities with skills, parallelize work with subagents, and automate workflows with hooks. This is not just using tools well; it is changing the paradigm of development itself. In an era when a solo developer can produce team-level productivity, you have now taken your first step as an agentic engineer.
📚 References (Sources) - Click to expand•Official Claude Code Best Practices
•Claude Code Tips: 10 Real Productivity Workflows (2026)
•A Practical Guide to the Claude Code Context Window Size
•How to Use Claude Code in 2026: From Installation to Practical Use (for Non-Developers)
•How Non-Developers Can Code with Claude Code - Retention
•A Humanities Journalist's Vibe Coding Challenge - ZDNet Korea
•How a Non-Developer Marketer Uses Claude Code
•Writing a Good CLAUDE.md — HumanLayer
•Claude Code Ultimate Guide — GitHub
•Custom Subagents — Claude Code Docs
•Extend Claude with Skills — Claude Code Docs
•Claude Code Anniversary: 5 Real-World Use Cases (2026)
•5 Powerful Claude Code Use Cases — The Prompt Warrior
•Claude Code Templates: 1000+ Agents, Commands, Skills & MCP