Write 20,000 lines of code on the high-speed train in 2 hours? The author personally tested Claude Code as a “command line partner”: a 5-step process (project memory→PRD→ planning→ coding → testing) halves bugs, doubles efficiency, and comes with a command cheat sheet, which even novices can get started with in 10 minutes.
In the past two days, I went to Shanghai to participate in Feishu’s Future Infinite Conference, and in order to make the time on the road pass faster, I have been writing code in claude code.
I just recently reorganized the logic of using claude code, and finally I can work efficiently with it.
This time on the high-speed rail, I used two claude codes to work together, and the efficiency was more than doubled compared to my previous mixed development with cursor+claude code, and the development of the 2.0 version of the “prompt word management assistant” is quite smooth, and it is expected to be officially released next week.
In the past two days, many friends have asked me how to use claude code, and it just so happens that I used the case of “Prompt Management Assistant” version 2.0 to share with you the methodology of my collaboration with claude coding, how to make claude code an AI partner for efficient development.
When using cursor for AI programming before, the overall process was as follows: 1. Write a prd document to make it easier for the AI to understand all the content to be developed this time2. AI writes code, based on the prd document and the content developed by AI together (mainly click accept)3. Test feedback, give it the product test feedback problems developed by AI and let it fix it
Using this process for product development is actually very comfortable when writing some scripts and low amount of code, AI will come up with the product for you in a while, and the effect is very good.
When I first started working as a prompt management assistant, I made the first version in more than an hour, and AI was developed very quickly in a low-code state.
But once the project becomes larger, the bug rate of AI writing code has greatly increased, and the main reason I analyzed is that AI doesn’t understand your project at all, it is really “learning by doing”.
It’s like finding a programmer to help us optimize the existing code, it doesn’t understand what this code warehouse is, it doesn’t know where each file is stored, it gets a request and immediately starts to get ready.
It’s strange that the quality of the code can be used…
Therefore, when I develop in cursor, I first let the AI locate the files involved in this requirement, and then let it perform function upgrades or iterations based on these files, which slightly reduces the failure rate of cursor code writing.
It’s a bit useless, people have to keep an eye on all the projects at any time, and AI can’t do the work independently, or the bug is really flying all over the sky.
When it came time to work with Claude Code, I upgraded the process thanks to its own feature upgrades:
1. Let AI write project documents, claude.md documents are its overall memory, and the AI’s understanding of the project is present throughout the development process and will not be lost
2. Write a PRD document to facilitate the AI to understand all the content to be developed this time
3. Make a specific development plan, precise to which documents to optimize, this optimization is divided into several stages
4. AI writes code, based on the PRD document and the content developed by AI together
5. Test feedback, give it the product test feedback questions developed by AI and let it fix it
The main purpose is to add two processes, project documentation and development plan, both to give AI more contextual information, thereby reducing the bug rate of AI code writing.
Next, I will demonstrate the development effect of this process through the case of “Prompt Management Assistant”, and take you step by step to see how to work efficiently with Claude Code.
In order to facilitate everyone’s better understanding of the instructions used in Claude Code, I have compiled a Claude Code Instruction Quick Check Guide (https://t16jzwqrzjx.feishu.cn/wiki/E4Rbw3vpBikNLbknNKOc7GtYnqd?from=from_copylink with AI, which explains the meaning of commonly used commands, which can help you quickly understand claude codeThe meaning of various instructions.
Step 1: Let AI write project documents
The first step to enter the project is to let the AI write the project file, using the /init command, just confirm that it will browse the project by itself and generate a claude.md document.
I’ll take a module for you to see how detailed claude.md comb, it will disassemble and record your core architecture, and it knows the function and storage location of each folder.
claude.md documents are applied to claude code in the global world, just like cursor’s system rules.
Its role is to give AI more context and let AI better understand the project for project development.
/init is not only used for initialization, but also after completing a version iteration, /init can also be run, at which time it will automatically update the claude.md document and add the developed content.
My own development habit is: every time I complete a version of development, I do a GitHub save, and then run /init once.
Step 2: Write a PRD document
The second step is to write a PRD document, which previously listed prompts and cases in the “Xiaobai AI Programming Guide”, and I will not talk about it in this article.
Here I mainly share with you how to efficiently discuss your needs with Claude Code.
By creating claude.md documentation, Claude Code has a good understanding of the project, and it already knows the core logic, file structure, and module functions of my project, and it is a very efficient thing to discuss with it.
The red part is the request I put forward, we can see that the AI is based on claude.md documentation, and then read the overall system, and finally give me a thought of its own.
It first poured cold water on me, and then gave me three different progressive schemes; Finally, it thought for a while and told me to start with Option 1, and gave me detailed logic and reasons.
Claude code plays a crucial role in discussing requirements and understanding the project.
Step 3: Make a specific development plan
The third step is to make a plan with Claude Code, here we will first talk about the three modes of operation of Claude Code.
Claude code supports three job modes, namely default (default mode), plan (plan mode), Auto-accept (work mode), the conversation between me and claude code just now is the default mode, just hold down shift+tab to switch modes.
The plan displayed in the lower left corner means that we have switched to planning mode, so let’s make a development plan along the previous plan 1.
The first step in Claude code planning is still to look at the code and learn more about what the existing codebase looks like.
It then gives the user a detailed plan to develop the requirement.
When I saw this plan, I was full of confidence, at least the AI thought of every piece, then its work effect would naturally be very good.
If you want to accept the plan, select yes, if you don’t want to accept it, choose no, and continue discussing the plan.
Step 4: AI writes code
After confirming the plan, we entered the fourth step, let the AI work to write the code, and we can hang up directly in this link, just wait for the AI to finish writing and cut back to take a look.
You don’t even need to click on accept for this, the AI will do it all by itself, and humans only need to test the feedback.
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
Step 5: Test the feedback questions
The fifth step is to test the feedback problem, and one thing to note about this is that because claude code uses command-line tools, we can’t map them in.
Therefore, we need to save the screenshot to the folder first, and then use the @ image name to feedback the bug to the claude code, which is a little more troublesome.
After reading these five steps, you should be able to understand why I said that using claude code can double the programming efficiency.
Claude code gives AI more context, and based on this ability, I optimize my process of collaborating with AI to improve efficiency.
The ability to collaborate with AI is undoubtedly the most important thing under the AI wave.
After talking so much about improving efficiency with claude code, I would like to end the article by talking a little about the fear of using claude code.
Recently, many friends have been taken by me to use claude code, they actually have a question with me before using claude code, it is a command line tool, can I really use it?
When the claude code first came out, Bai Shu shared with me excitedly, he and the teacher Xiao Pai quickly used the claude code to write something, and my first reaction at that time was that this was a command-line tool, and I might not understand it.
So it took more than a month to actually use Claude Code, and I was also afraid to experience a new product.
After using it, I found out that this is not a command-line version of the chatbot, and the biggest difference between it and the web version of Claude is that it is placed in the terminal.
If we forget all the instructions and the command line, we will find that it is a chatbot, but this chatbot can write code, and the functions are more and more powerful.
When I told everyone that this was a command-line tool, they were naturally very resistant; But when we treat it as a chatbot, everyone is willing to get started and experience it, and it is quite fun.
So I think you don’t treat it as a very professional programmer developer tool, you can just use it as a large bean bag AI programming.
Finally, I would like to say: don’t overcomplicate Claude Code and don’t keep yourself out the door first.
