Mastering Claude Code

From Speed to Strategy

After mastering ten beginner tips, your pattern of using Claude Code begins to shift. You reflexively run /init, develop awareness of context, and form the habit of designing first in plan mode. Yet at some point you hit a wall. Working alone in a single session on a single task has inherent speed limits.

Intermediate tips are strategies to move beyond this ceiling. They cover running multiple agents in parallel, automating repetitive work, and optimizing costs.

Tip 11. Delegating Parallel Work to Subagents

One person doing research, coding, and validation sequentially is not comparable to three people each handling one task simultaneously. The difference in speed is dramatic.

In Claude Code, when you hand off a complex task, you can include an instruction in your prompt: use subagents. The main session analyzes the work, then spawns multiple independent subagents. Each subagent has its own context window and can run its own model.

One agent handles research, another writes validation code, a third explores alternative approaches. All of this happens in parallel. When done, each subagent reports its results back to the main agent. The main thread stays clean, while you effectively get the benefit of a small development team working simultaneously.

Tip 12. Creating Custom Skill Files

If you find yourself rewriting the same type of prompt every time you repeat a task, you need a custom skill file.

Create markdown files in the .claude/skills/ directory. For instance, if you write a detailed technical debt audit procedure in a file called techdebt.md, you can later say run a tech debt audit and that workflow executes automatically.

Store code review standards in codereview.md, and you get consistent-quality reviews without restating your criteria each time.

The real power of skill files lies in team sharing. Commit them to GitHub and your entire team can use the same workflows. Code reviews and pre-deployment checks that once varied by person become standardized through a single skill file. It is much like automating your company's standard operating procedures.

Tip 13. Reducing Subagent Costs with Haiku Models

Once you start running subagents in earnest, costs become visible. There is a strategy to manage this.

If your subagent's task is to read large volumes of text and extract key points, there is no reason to use an Opus-class model. Assign a lightweight, cheaper model like Haiku.

Take a concrete example. You have a research task: read dozens of articles and summarize them. You will process hundreds of thousands of tokens. Running this on Opus is expensive. But setting your subagent to Haiku cuts costs dramatically. The subagent reads and filters; only distilled summaries reach the main agent (Opus). This way, the final analysis stage, where quality matters most, still gets the powerful model.

Do not abandon subagent use because of costs. With strategic model selection, you control expenses while keeping the benefits of parallel work.

Tip 14. Keeping claude.md Current

Your claude.md file is a living document. As your project evolves, you discover new patterns, encounter unexpected pitfalls, and develop new rules. Reflect each of these discoveries in claude.md.

You can instruct Claude directly: Record this pattern I just discovered in claude.md. In your next session, Claude already knows that lesson. You stop repeating the same mistakes, and your understanding of the project deepens.

But there is a catch. claude.md functions as a system prompt. Its contents load with every conversation and consume that much context. So keep this file to 150-200 lines. If you only add and never trim, the file balloons and you run short on context for actual work.

When adding new information, simultaneously prune outdated or less critical items. This balance determines the quality of your claude.md.

Tip 15. Routing from claude.md to External Files

I said keep claude.md to 150-200 lines, but your project may need far more information. Style guides, business context, reference documents, API specs. Putting all of this into claude.md will bloat the file.

The answer is routing. In claude.md, record only where each piece of information lives. See /docs/style-guide.md for style conventions, See /docs/api-spec.md for API specs. Claude reads the files when needed.

The payoff is clear. claude.md stays lean, so you waste less context. At the same time, the total information Claude can access actually grows. You do not need to write your entire project state into the system prompt. Knowing where that state lives is enough.

Tip 16. Escape When You See the Wrong Direction

Sometimes while Claude is generating a response, you can see the direction is wrong. It tries to install a library you did not ask for, or starts writing code with an approach you never intended.

Do not wait for it to finish. Press Escape and generation stops instantly.

The longer Claude moves in the wrong direction, the more tokens you burn. Spent tokens cannot be recovered, and bad context that accumulates in your window will color later responses. Cutting it short, correcting course, and resubmitting is far more economical overall.

Waiting in hope that it will self-correct at the end usually costs you both time and tokens. The quicker you correct course, the better.

Tip 17. Push Back Hard on Output Quality

If Claude's output lands in the fine, I guess category, do not accept it.

Say, This is not good enough. Try a completely different approach. Or, This section is too verbose. Rewrite it more concisely and elegantly. When you set a high bar explicitly, Claude often surprises you with a much better result.

Why does the second attempt work better? Your critique clarifies what not to do. Claude now knows which direction to avoid and explores a different path.

You can go further. Once you get an improved result, ask Claude to record those lessons in claude.md or a skill file. Update the guidelines so you do not repeat this mistake. This structural blocking prevents whole categories of errors from recurring.

Tip 18. Fast Rewind with /rewind

Escape stops a response in progress. But what if the response finished and was the wrong move? What if Claude already modified files and the conversation has moved forward, and now you realize that branching point three steps back was better?

The /rewind command solves this. You return to an earlier point in the conversation. You do not restart the session from scratch. Rewind to where it went wrong, then branch in a different direction.

/rewind paired with Escape becomes a powerful duo. Stop in-progress mistakes with Escape; undo past ones with /rewind. With both tools at hand, there are few irreversible mistakes in a Claude Code session.

Tip 19. Setting Alerts with Hooks

When you start running multiple Claude Code sessions at the same time, new problems emerge. It becomes difficult to determine which session has finished and which is waiting for your input.

Using the /hooks command, you can set up notification hooks. You can give instructions in natural language too. "Play a notification sound when this session finishes." This sets the system to alert you when Claude Code completes its task.

Once this is set up, you don't need to stare at the terminal waiting for work to finish. You can do other things, and when the alert sounds, come back to check the results and give the next instruction.

Imagine running 15 sessions simultaneously. Without notifications, you'd have to switch between terminals one by one to see which has finished. With notifications, you just go to wherever the "ding" sounds. It may seem trivial, but it's essential for managing multiple sessions.

Tip 20. Using Screenshots for Visual Self-Check

Claude can see. This means it doesn't just process text but can analyze images.

The most effective way to use this capability is to create a visual self-check loop. If you're building a website, you'd structure the flow like this.

1. Implement the website design. 2. Take a screenshot. 3. Analyze the screenshot to identify layout issues. 4. Fix the problems. 5. Take and analyze another screenshot.

If you set it up to repeat this loop about three times before showing the first version (V1), the difference in polish between that V1 and one produced without self-checks becomes quite clear.

Screenshot use isn't limited to web design. You can capture error message screens and show them, or photograph other websites' designs to use as inspiration sources. Screenshots are the most efficient way to convey visual information that's hard to explain in text.

Building Strategy on Fundamentals

We've looked at ten intermediate tips. What they have in common is that they're strategies to move beyond the limitation of "doing one thing alone." You parallelize with assistant agents, automate with skill files, optimize costs through model selection, manage multiple sessions with hooks, and improve quality with feedback and iteration.

Once these strategies become second nature, you're ready to take another step forward. Advanced techniques like automating browsers, connecting MCP servers, and having agents create agents themselves are the next level.

Attorney Kim Kyungjin, AI Policy Expert

Specialist in AI legal policy, former member of the National Assembly, author of numerous works

If this book has stayed with you even briefly, please support it so the next story can reach the world.

(Voluntary support account: Nonghyup 302-1096-0948-81 Account holder: Kim Kyungjin)