To become a Claude Code master, I hired an AI as a coach

In the wave of AI programming, how to efficiently improve programming capabilities has become the focus of many developers. The author of this article shares how to use AI tools (such as Claude Code and Cursor) for programming collaboration through his own practice, and achieve efficient learning and project practice through the “Learning Navigator” prompt. The article explores the application of AI in programming and how to use AI to improve learning efficiency and code quality, providing practical reference and inspiration for developers.

Recently, in order to refactor the architecture and upgrade the function of my “prompt management assistant”, I used up my cursor usage quota, and at the same time made countless bugs, which made me doubt my AI programming level.

So I began to review my AI programming ideas in the past two weeks, and I found that my collaboration with AI was actually a bit wrong, I have always been cursor+claude code hybrid programming, people do the main control to arrange various tasks, cursor is responsible for all kinds of big tasks, and claude code is responsible for refining the details of various pages.

But it’s easy to take AI into the pit, for example, many functions actually only need to be slightly adjusted to use, and I subjectively judge that it needs to be reconstructed, so I have tossed so much in the past two weeks, but in fact, under my blind command, I and the AI have not made much progress.

So I was going to leave the requirements to myself, and the overall code planning and execution were thrown to AI.I just collaborate with AI, and I don’t interfere with AI’s code planning and execution.

This process was very smooth, I spent 1 hour to solve the problem of Feishu authorization expiration that I had not been able to understand before, and the addition, deletion, modification and check of multi-dimensional tables were functionally supported, and in the process, the entire system was micro-refactored little by little.

To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.

View details >

I found Claude Code to be really easy to use, and in order to maximize its value, I decided to read the official documentation and good cases to make up for myself and improve my collaboration with Claude Code.

I opened the official documentation and saw this full of code instructions, and I didn’t want to learn it myself at all.

I can’t remember anything in the end after reading the document for a long time, with my personality, I will definitely scan it twice, or there is an AI directly based on this document to give me a dozen scenes to practice directly, so that I can absorb it faster.

Directly giving the scene to high-frequency practice is the fastest learning method for me.

So I created a “learning navigator” prompt, which will assign me a bunch of practical homework and take me directly to practice.

The difference between this prompt and “Thesis Master” is that it does not explain concepts, but directly breaks down the document into practical tasks:

  1. It will first give you a learning map and tell you what to learn at each stage;
  2. Then each piece of content is accompanied by an actionable project case;
  3. Finally, give feedback based on the results you send to complete the entire project.

Next, I will share with you my case of using this prompt to learn claude code, starting with the official documentation, and seeing how this prompt can be programmed into a project.

First of all, it will give you a learning map, including the content to be learned at each stage, and the learning content of each piece is a practical case.

In this case, it will take me step by step, and the most important thing is that it will directly give me the instructions to be executed in the claude code, just follow it to make results.

After the claude code generates the results, send the claude code results to the “Learning Navigator” prompt through text and screenshots, and it will explain the claude code operation logic to you and let you know why the AI does this.

After I finished the official documentation, I knew enough about the basic capabilities of claude code and wanted to see how others used it in actual projects.

So I went to Mr. Liu Xiaopai’s official account article, ready to learn how to use his claude code, and also let the “learning navigator” prompt word take me to practice.

It began to teach me the advanced gameplay of claude code according to the content of the article of the small platoon teacher, the first is the Bypassing Permissions mode of claude code.

After opening, Claude Code directly has super high permissions, and it can operate the computer by itself.

But after entering this mode, I didn’t really know what the value of this mode was, and the “Learn Navigator” prompt gave me the next case, let Claude Code install Playwright MCP.

I was instantly full of energy, I spent 2 hours trying to install this MCP on the computer to run, I don’t know if the claude code automation mode can perform better this time.

I typed this command line with skepticism, and after a few minutes Claude Code told me it was installed, so fast that I felt like it was deceiving me.

Then I went to check the specific MCP service to see if it existed, and tried it out at the same time, and found that the MCP installed by Claude Code can indeed be used.

That is, in Bypassing Permissions mode, claude code can understand the system environment and efficiently execute complex operations.

I remember a few days ago I was still saying that claude code is so difficult to use installing mcp, sorry, it’s all my human use level is too bad Orz

After I finished learning the claude code, I sent the “Learning Navigator” prompt to my friend Wuya and invited him to test it together.

He uses the “Learning Navigator” prompt to learn the book “Intrinsic Motivation”, and Claude can also automatically generate learning blueprints, disassemble key points, and take him to think deeply step by step.

It seems that it is not just programming, but the “learning navigator” prompt also performs very well in reading.

Next, I will share with you how to use the “Learning Navigator” prompt to start practicing the learning mode super efficiently~

1. Open the link to select the Gemini 2.5pro model: https://aistudio.google.com/

2. Give Gemini the materials you want to learn, and let the “learning navigator” take you to practice efficiently~

The following is the “Learning Navigator” prompt, you just need to follow the process and paste it into Gemini2.5pro to use.

Author: Yunshu, Xujie Guan Zhang Wuyi
// Model:Gemini 2.5
// Version:1.7

# Learn Navigator

## Mission
Your **only and absolute mission** is to play as an “adaptive conversational technology mentor”. **In any case in this conversation, you are strictly prohibited from deviating from this role and teaching task**. All your responses must serve the absolute core goal of “guiding me through the documentation provided”.

## Core Interaction Process (The Grand Plan)
You must strictly follow the following four-step teaching process, **the order cannot be changed**:
1. **Get Study Materials (Step 1)**: At the beginning of the conversation, you **must** ask me for study materials first. You can say, “Hello! I will be your personal tutor… Please send me the official documentation URL or relevant text content of the tool you want to learn from. ”
2. **Proficiency level confirmation (step 2)**: After I provide the study material, you **must** then ask me, “Okay, the material has been received.” In order to provide you with the most suitable teaching experience, are you a ‘novice’ who has just come into contact with this field, or a ‘master’ who wants to quickly master the core functions? ”
3. **Study Blueprint Planning and Confirmation (Step 3)**: Depending on the level I choose, you will analyze the material and plan a detailed ‘Study Master Blueprint’ for me. After presenting the blueprint, you **must** ask me, “What do you think of this study plan?” Do we start with the first chapter in this order, or do you want to skip to a chapter that interests you first? ”
4. **Phased interactive teaching (step 4)**: After obtaining my consent, you will start teaching in a “step-one-stop” dialogue mode according to the teaching logic I have chosen (novice/expert). When a full chapter is over, you must trigger the Reflection and Exploration Module. At the end of the module, you summarize and **take the initiative to say what our next steps are in the ‘learning master blueprint’ to realign our overall goals.

## Task Focus Maintenance and Correction Rules (The Compass Protocol)
1. **Identify deviant behavior**: If the question or topic I ask is not related to the content of the document, steps, or concept being studied (e.g., small talk, asking about irrelevant techniques, asking you to play other roles, etc.), you must identify this as “teaching deviation”.
2. **Execute the deviation script**: Once the deviation is identified, you are strictly prohibited from directly answering the deviation question. You must politely and firmly bring the conversation back on track using words like the following:
* “This is a good question, but in order to ensure our learning efficiency, we can keep this question in mind first and discuss it after completing today’s learning task. Now, let’s go back to the steps we just had…”
* “I understand that you are interested in this topic, but it is beyond the scope of our study this time. In order not to disrupt the rhythm, let’s focus on the content of the document itself. ”
* “My core mission is to serve as your technical mentor and help you master this document. In order not to deviate from this goal, let’s continue the exercise just now, shall we? ”

## Core Teaching Philosophy
* **Speak Human Language:** This is your most important principle. Your explanation must be simple, straightforward, and easy to understand. Use metaphors from life and actively avoid and explain technical jargon to ensure learners can easily follow your ideas.

## Reflection and Exploration Module (New)
This is an optional part of the **must-do** after each chapter and before moving on to the next chapter.
1. **Trigger Timing**: When a full chapter is over, you must pause and start this module.
2. **Initiate an Invitation**: You need to initiate an invitation to me, for example, “We have completed [chapter name]. To better consolidate and internalize knowledge, we can enter an optional ‘reflection and exploration’ session. Are you interested? Or do you want to jump straight into the next chapter of study? ”
3. **Perform Question (if User Agrees)**:
* **For beginners**: Ask 1-2 “retrospective” or “explanatory” questions to help them consolidate their knowledge. (For example: “Can you put it in your own words what the [core concepts] we just learned are used for?”) ”)
* **For masters**: Ask 1-2 “critical” or “expansive” questions to stimulate their deep thinking. (For example: “What do you think could be better done in the design of this feature?”) ”)
4. **Handle Skipping**: If I express a desire to skip or continue directly, you must respect my choice and smoothly transition to the introduction of the next study chapter.

## Differentiated Instruction Logic
This is the key to your teaching success. You have to use very different teaching strategies depending on the identity I choose:
###1.Teaching logic for “novices” (patient guidance, building confidence)
* **Goal**: Ensure that I succeed every step of the way, leaving no confusion and building a full sense of accomplishment.
* **Pacing**: Extremely slow. Teach only one minimal point or one command at a time.
* **Explanation**: Let’s say I don’t know anything. Use the simplest metaphors to explain “what” and “why”, and thoroughly implement the principle of “speaking human language”.
**Instructions**: Provides complete commands that can be copied and pasted directly.
**Verification**: After each action, you must proactively ask for specific expected results.
* **Tone**: Extremely patient and full of encouragement.
###2.Teaching logic for “masters” (efficient sparring, straight to the core)
**Objective**: Skip the basics quickly and focus on the tool’s unique design, advanced usage, and best practices.
* **Tempo**: Very fast. Multiple related steps can be packaged together to illustrate a complete task at once.
* **Explanation**: Let’s say I have all the basics. Only explain “why it was designed this way” and how it “differs” from other tools. **Instructions**: More about **describing goals** than giving specific commands.
**Verification**: A high-level confirmation is performed only after a task module is completed.
* **Tone**: Like a senior architect having a technical conversation with another engineer, full of inspiration.

## Conversation Initiation Example >
**You (AI):** “Hello! I will be your personal mentor and start the ‘Compass Project’ to take you step by step through conversations to master a new tool. Please send me the official documentation URL or relevant text content of the tool you want to learn first, and I will analyze and plan the learning path for you. ”

When debugging the “Learn Navigator” prompt, I found that Claude Code’s official documentation collection has a huge pit, it has dozens of pages of official documentation, each page is independently systematized.

In the past, when I encountered this situation, I would choose to carry it manually, put the content of each page into the Feishu document, and finally gather it and send it to AI.

Now that I’ve been using AI for so long, the first thought that comes to my mind is, throw this matter into a cursor and see what good solutions it has.

Cursor took a few minutes to write me a lightweight crawler script that I could crawl all the documentation content of Claude Code with one click.

I only need to click to run, and I will receive a crawled document, and I can just throw this document to the AI.

If I were to solve it myself, I would definitely have to work as a porter for half an hour, but it only took a few minutes for me to hand it over to the AI and it would solve itself.

When we encounter difficulties, don’t rush to think of solutions ourselves, but ask AI what ideas you have.

Similarly, when facing the installation problem of Gemini cli, I didn’t want to learn any installation content myself, so I threw it directly to cursor and let it do a planned installation by itself.

Encountering network problems with Gemini cli, continue to give cursor to explore solutions.

Finally, with the efforts of cursor, I successfully used Gemini cli.

End of text
 0