[AI Library] Chapter 29: Advanced Hacks: 7 Pro Tips
Mastering Claude Code
Chapter 29: Advanced Hacks: 7 Pro Tips
Kim Kyung-jin
Mastering Claude Code
People Who Push the Boundaries
Five terminal windows are open at the same time. On the left monitor, three Claude Code sessions are each building different features, and on the right monitor, a browser opens automatically, filling in the form for a newly created web app and clicking buttons. Each time an alert sounds, you check the completed session and assign the next task.
This is not science fiction. By combining Claude Code's advanced features, this workflow is genuinely possible. The seven tips covered in this chapter are for those who want to push Claude Code to its limits.
Tip 21. Validating Features with Chrome DevTools
Screenshots catch visual problems. But a screenshot alone cannot tell you whether buttons actually work when pressed or whether form submissions are processed correctly.
Claude Code can directly use Chrome DevTools. It opens a browser, interacts with the app, checks console errors, and verifies that network requests are flowing properly.
This is an extension of the screenshot loop. If a screenshot is a "visual inspection," then DevTools use is a "hands-on inspection." By combining these two for frontend work, you can ensure both the visual polish and functional stability of Claude's output.
One more thing. DevTools validation is also useful when interacting with services that do not provide explicit APIs. You can delegate to Claude the work of reading data through a browser in a logged-in state, filling out forms, or clicking buttons.
Tip 22. Chaining MCP Connection Servers
MCP (Model Context Protocol) servers are a powerful way to connect external functions to Claude Code. Connect a Notion MCP and you can read and write Notion databases; connect a GitHub MCP and you can manage pull requests.
But there is no reason to use only one MCP server. When you connect multiple MCP servers simultaneously, Claude moves between different services. A single session can read planning documents from Notion, create issues on GitHub based on that content, and register deadlines in Google Calendar.
There is one caveat. Each MCP server loads tool definitions into the context window. The more servers you connect, the more tokens the base model consumes. You need to manage this by activating only the servers you need for the current task and disabling unused ones.
Using the /context command you learned in Tip 5, you can check how many tokens each MCP server is consuming.
Connecting all MCPs is not the answer. If your project truly needs only one or two functions, hardcoding API endpoints directly instead of using MCP might be more efficient in terms of token usage. Judgment must fit the situation.
Tip 23. Letting an Agent Create Other Agents
Auxiliary agents are whom the main agent delegates work to. But you can go one step further. You instruct Claude: "Design an agent to perform this task."
This is not delegating a task but having Claude design the system itself to perform the task. Claude determines the agents' roles, which models to use, the execution order, and validation methods, then creates the agents according to that design.
Using Agent Teams makes this approach even more effective. Ordinary auxiliary agents cannot communicate with each other. They work independently and only report results to the main agent. Agent Teams are different. Agents within a team share task lists, communicate with each other, and are aware of whether others have completed their work. They can assign tasks to other agents.
The cost is higher and execution takes longer. But large projects produce far more cohesive results. Forcing together code snippets that agents created while looking in different directions is different from integrating code that agents created while coordinating with each other. The consistency of the outcome differs.
Tip 24. Running Multiple Claude Code Sessions in Parallel
If you open multiple terminals in one project folder and run Claude Code in each, there is a risk of conflicts from modifying the same files simultaneously.
Combining this with Git Worktree solves the problem. Each session works on a separate branch in a separate directory, so they do not interfere with each other. In Claude Code, you can easily create a worktree using the --worktree flag.
This creates three independent workspaces, each running a Claude Code session. You can run four or five simultaneously. When work is done, merge each branch into main.
Some people push this approach to the extreme. One of Claude Code's core developers says he always runs five agents in the terminal and operates about ten more on the web.
Tip 25. Creating Custom Slash Commands
Beyond Claude Code's built-in slash commands (/init, /compact, /rewind, etc.), you can create your own commands.
Similar to skill files, but slash commands offer a more direct invocation method. You can bundle frequently performed workflows into a single command and start complex procedures with one-line input.
For example, suppose you create a custom command called /deploy-check. With this single command, you can configure it to run validation, lint checks, build verification, and deployment to staging sequentially. Instead of typing "run validation, check lint, build..." every time, you just type /deploy-check.
Creating custom commands that the entire team uses contributes to workflow standardization. When new team members join, you can guide them to use these commands, and they can immediately begin workflows of the same quality as experienced members.
Tip 26. Browser Automation
One step beyond validating features with Chrome DevTools takes you into full-fledged browser automation.
Claude Code opens a browser, navigates websites, fills out forms, clicks buttons, and reads data. It can be used to collect data from services that do not provide explicit APIs or to automate repetitive web tasks.
It is also possible to borrow the design of a website you admire. Take a screenshot of a site you like or extract its HTML styles and pass them to Claude, and it can recreate that design pattern. Of course, it is better to use it as a source of inspiration rather than copy it directly, and add your own elements.
The scope of browser automation is broad. Still, the most stable form is performing navigation, clicking, and input tasks in an already logged-in state. When determining the scope of automation, it is good to decide in advance what you will delegate to Claude and where human involvement begins.
Tip 27. Security Considerations: Managing API Keys
The more you actively use advanced features, the more important security becomes. When you connect multiple MCP servers, run browser automation, and operate multiple sessions, API keys and authentication credentials can be exposed in multiple places.
You must follow several principles.
Never put API keys directly in code. Manage them through environment variables or .env files. Always add .env files to .gitignore so they do not get uploaded to GitHub. If a key ever enters commit history, consider it already exposed and regenerate it immediately.
The same applies to Claude Code sessions. If Claude generates a response containing an API key, that content remains in the context. You should design the structure so that key values themselves appear in conversation as little as possible.
Security verification is also necessary when installing VS Code extensions or MCP connection servers. Check in advance the developer's identity, whether outbound network calls are made, whether data is collected, and whether shell commands can be executed. Installing without verification could lead to irreversible credential leaks.
Those Who Know the Depth of Tools
Seven advanced tips reveal the depth of Claude Code as a tool. Controlling the browser, letting agents design other agents, managing dozens of sessions at once while keeping security intact. This is the level of properly using the tool.
To use these techniques safely, you must understand how Git worktrees function and Claude Code's permission system. Worktrees provide the safety net for parallel work, while the permission system that balances autonomy and control forms the foundation.
Gyeong-jin Kim, Attorney, AI Policy Expert
AI legal policy specialist · former National Assembly member · author of multiple works
If this book has stayed with you, even for a moment, please support the next story so it can reach the world.
(Voluntary support requested. Account: Nonghyup 302-1096-0948-81, Account holder: Gyeong-jin Kim)
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.