"I asked Claude Code, but something different came out..." "It's working, but the result is completely different from what I was looking for!" -- Have you ever experienced something like this?

Claude Code is a terminal-based AI coding assistant provided by Anthropic. As of February 2026, it is available in the Pro ($20 per month) and Max ($100-200 per month) plans, and is rapidly gaining popularity as a useful tool that allows you to write and modify code simply by giving instructions in natural language.

However, if you use it incorrectly, you can easily get stuck in the swamp of ``It's working, but the results aren't what you're looking for.'' In fact, in X, we often hear comments such as ``Claude Code is fast to implement, but if the result is wrong, it's difficult to fix'' and ``Even though I'm paying for it, I can't work properly.''

In this article, we will explain five patterns that beginners of Claude Code tend to fall into and how to give instructions (prompts) to get the desired results.

What is Claude Code? Rough explanation in 30 seconds

Claude Code is an AI coding tool that can be used directly from the terminal (command line). Unlike the web browser version of Claude, the biggest feature is that you can directly read and write your project files.

As of February 2026, Pricing plan is as follows:

  • Pro Plan ($20/month): You can use Claude Code. However, usage is limited
  • Max 5x ($100/month): Approximately 5 times more usable than Pro
  • Max 20x ($200 per month): Approximately 20 times more usable than Pro
  • API pay-as-you-go: Pay only for what you use. For developers

In other words, anyone can use it if they pay a fee. However, ``usable = able to use it well'' does not mean. Here's the catch.

Addicted pattern 1: "Run toss problem" where the instructions are too vague

This is the most common one. This is a pattern that gives instructions that are too general, such as ``create a login function'' and ``refactor it in a nice way.''

Claude Code is excellent, so even if the instructions are vague, it will come up with similar code. But "it seems like that" and "what you're looking for" are two different things. Just like when you ask a human colleague, if you don't know what to do, where to do it, and with what specifications, you won't be able to get the desired product.

Tips for improvement:

  • × “Create a login function”
  • ○ "Add email + password login function to src/auth/. I am using Next.js App Router and use Supabase Auth for authentication. After successful login, redirect to /dashboard"

In short, accuracy can be greatly improved by simply communicating technology stack, file location, and completion conditions.

Addictive pattern ②: “Greedy problem” that makes you do everything at once

“Create the authentication function, create the dashboard, and design all the APIs”—I understand what you mean, but if you throw too many tasks at once, Claude Code tends to lose the context in the middle.

Anthropic's official best practices also recommends breaking tasks into smaller pieces. Specifically, the following steps are valid:

  1. Let them make a plan first: "I want to implement 〇〇. Please come up with a plan for how to proceed."
  2. Check the plan and then execute step by step: “Do Step 1” → Confirm → “Do Step 2”
  3. Use Plan Mode before making major changes: Claude Code's /plan command lets you review your design before implementation

Just by repeating this cycle of "Plan → Execute → Confirm", the problem of "It's working but it's not right" will be greatly reduced.

Addicted pattern 3: "Initial settings skipping problem" of not using CLAUDE.md

Claude Code has a feature called CLAUDE.md that allows you to place a project-specific configuration file. If you create this at the root of your project, it will be automatically loaded in every session, and Claude Code will start with an understanding of the "rules for this project."

According to HumanLayer guide, there are three things you should write in CLAUDE.md:

  • Frequently used commands: Build, test, and deploy commands
  • Coding standards: naming conventions, libraries used, directory structure
  • Things you don't want to do: "Do not leave console.log", "Do not delete existing tests", etc.

The key is to write concisely. A guideline is 60 to 300 lines; if it's too long, the instructions will be obscured. The golden rule is to have clear minimum rules.

Addicted pattern 4: "Loop problem" where the same instructions are repeated when an error occurs

When Claude Code generates an error, do you vaguely repeat "Try again" and "Fix it"? This makes it easy to keep repeating the same mistakes.

As of February 2026, according to actual data posted by a developer on X, the error rate when using Claude Code is approximately 14.6%. In other words,it doesn't workabout once in seven times. The important thing is how to deal with errors when they occur.

Correct response flow in case of error:

  1. Paste the error message as is: Say "This error occurred" in text instead of a screenshot
  2. Specify what you expected: “I was expecting a result of 〇〇, but it turned out to be △△.”
  3. Reset context: If you get lost in a long session, reset the conversation with /clear and restart with new instructions

The third one is especially important. If you have a long conversation, the context may become too large and Claude Code may be drawn to outdated information. If in doubt, reset is the key.

Addicted pattern 5: Not being aware of the limit of the billing plan "Limit exceeding problem"

At X, we heard comments like ``I can't work properly even though I'm paying'' and ``Where did I hear about the 5-hour reset?'' This is a case of not understanding usage restrictions for each plan.

Usage restrictions as of February 2026 Roughly summarized:

  • Pro Plan: Basic usage plan. Heavy use may result in restrictions for several days
  • Max 5x: 5x the usage limit of Pro. In many cases, this is sufficient for individual developers
  • Max 20x: 20x more than Pro. For people who use it all day for work

When the limit is reached, Claude Code's response may become slow or it may become temporarily unusable. The following measures are effective:

  • Check your usage regularly: You can check it on your claude.com dashboard
  • Reducing unnecessary interactions: Accurately convey instructions in one time (→Improving pattern 1 is effective)
  • If you reach the limit, give it some time: It is important to take time to do other work until you reset

Summary: Claude Code is "an excellent colleague who doesn't understand"

Claude Code is a powerful tool that can reduce implementation speed from 2-3 hours to 15 minutes. But that's only when you can give appropriate instructions.

Let's review the main points:

  1. Specific instructions: Clarify technology stack, location, and completion conditions
  2. Divide tasks: Proceed through the cycle of planning → execution → checking
  3. Write CLAUDE.md: Communicate project rules in advance
  4. Add information in case of error: Don't loop, use resets
  5. Understand your plan limits: Check your usage and use it efficiently

Roughly speaking, Claude Code is a "coworker who is extremely talented, but doesn't understand my intentions." If you ask politely, they will deliver results that exceed your expectations, and if you ask roughly, they will say, ``Something is wrong.'' It's the same with human relationships.

FAQ

Can I use Claude Code for free?

As of February 2026, a paid plan of Pro ($20 per month) or higher or API pay-as-you-go is required. Not available with the free plan Free. For more information, please check the official pricing page.

Can I use Claude Code even if I am a programming beginner?

It can be used if you have basic knowledge of terminal operations. However, you do need the ability to determine whether the generated code is correct. We recommend practicing with a small project first.

Where should I put CLAUDE.md?

Place the file name CLAUDE.md in the root directory of the project. Claude Code will automatically load it at the start of your session. A concise content of 60 to 300 lines is recommended.

What should I do if I am restricted from using Claude Code?

Please wait for it to reset over time or consider upgrading to the Max plan. If you switch to API pay-as-you-go billing, you can use it without any restrictions and pay only for what you use.

References