[Red Envelope] Sharing my Vibe Coding experience—seeking advice from the pros

Internet services Chit-chat
1

Hey, I wanted to chat with everyone about something—I’ve been way too reliant on AI lately, spending tons of time and money on Vibe Coding. It all started when I stumbled on a Douyin video called Open clow. The video talked about someone installing open clow for other people right when it came out, earning hundreds of thousands in a month. That got me so pumped, I just couldn’t resist diving into this whole world of Agents and AI stuff.

I don’t have any programming background at all; at first, I only had a vague idea of how things worked and could barely be considered computer literate—then I fell deep into this rabbit hole. At the beginning, I didn’t even know how to connect a DeepSeek API. But now, I can use top-tier Agents like cloud code and codex, plus the most advanced models like fable5 and even GPT 5.5.

But you know, I realized the tools themselves aren’t what matter most—it’s the people who use them. I used to think that with Vibe Code, I could make decent tools and decent software without any programming foundation. Turns out, I was wrong. I’ve now burned through almost a hundred billion tokens and spent about two or three thousand yuan on AI subscriptions. I’ve tried all the major models out there—MiniMax M3, DeepSeek V4 PRO, the top domestic models, and some of the best from overseas. They’re all really powerful, but for someone with no programming basics, they’re just toys—nothing close to real productivity tools.

I’m currently trying to develop a product selection tool for my family’s e-commerce business—basically a program to find hot, trending items or breakout products that get a lot of traffic. But the further I got, the more I realized I was out of my depth. The UI is a mess; I can’t figure out deployment at all. And the hardest thing of all is getting information sources. If I go the official API route, just checking a few thousand products would bankrupt me. But if I try using web scrapers, the data just isn’t up-to-date or complete enough, and the whole thing falls short of being a decent product. So, I had to put it on the back burner.

Now I’m using codex to make a horror game—ugh, but honestly, trying to do any of this without a programming background is just too hard. Like, basic issues crop up everywhere—characters drifting weirdly, camera in the wrong spot—all the simple stuff. I really feel helpless right now. I want to buckle down and learn systematically, but I don’t even know where to start. I’m just realizing now that AI’s power is still pretty limited; maybe that dream of just saying something and having a finished product pop out is still a few years away.

1 Like
2

Take a look

3

I’ve been using Claude’s Design system recently, and I think you could check it out.

If you use the Design system to design the frontend first, you can give it feedback. Once the frontend fully meets your requirements, you send it to Claude Code and let Claude Code handle the backend implementation. This workflow is really effective, especially for users like you who don’t know how to code but want to create a product.

4 Likes
4

Thank you for sharing.

6

Indeed, basic programming skills are the bottleneck. I suggest starting with Python basics and HTML/CSS — once you understand the fundamental logic, using AI together will be much more efficient.

1 Like
7

Emmm, I have the same view. Outsiders often say that with a powerful model, you’ll be invincible and can easily bring any of your ideas to life. But honestly, that’s just hype to attract traffic—most people who say this are running proxies or giving courses. Of course, you can still achieve good results, but first you need a fully-formed, closed-loop mindset. That’s really important. When directing AI to work, you can first use cc or cx to have a dialogue and turn your idea into a perfect plan or checklist document, then execute and refine it step by step. I think this approach will greatly improve your efficiency.

And then there’s another saying out there: “Humanities majors + AI = invincible.” Personally, I don’t quite agree with this. Humanities majors may be good with words and descriptions, but if their ideas don’t form a closed loop in their minds, their collaboration with AI will only have minimal effect.

Next, about domestic models: their performance is still pretty poor at the moment. Some can’t even handle much context—after a line or two, they start compressing, and if they compress the text multiple times, it leads to hallucinations. That’s frustrating. Of course, this isn’t to put down domestic models—I still hope they’ll rise to the top and dominate the world. But for now, only the latest GLM model can compete with cc and cx. The results are still slightly worse, but it’s good enough for everyday use.

If you want to learn, I recommend you look into system architecture. The type of code doesn’t matter—the model can already handle it. What matters most is your thinking and whether your ideas form a closed loop. When interacting with AI, if you can use AI to refine and implement the ideas and diagrams in your mind, that’s a huge win. As long as you get to that stage, you’re in great shape. Good luck! If you have ideas or difficulties, we can always discuss them, or chat with NL community friends to build a perfect tech forum together~

8

试试写思维导图 导出md

我上个月才开始接触 最开始就纯用chatbox 硬聊 代码编辑工具甚至是sublime这种原始人
后面接触了 aider 了解到架构师模式 一个主模型调用弱模型和编辑器模型 三个模型一起工作

然而连这其实也都是三年前的老物件了 我还觉得效率老高了 实际现在的体验是依托

后面接触到vscode 用cline的插件一段时间 换了kilo code 对vscode自带的copilot agent模式也体验了一段时间 昨天因为实在太缺token了 还去用了zcode

zcode我体验不深 直观感受就是 有种git代码版本管理不是很突出的感觉 没有代码直接编辑的区域 貌似只能让ai去编辑

但是code工具这么多 先plan一次 再去code 确实效率和幻觉都能降低不少 这是有目共睹的

kilo在plan完之后有选项去新对话 可以减少幻觉

vscode聊天里的plan好像没有去新对话的这个功能 对plan生成的清单也 缺乏掌握 老是不进行plan清单的下一步

特别是遇到gpt5.4 那种老爱回复:“如果你需要的话 我可以xxxxx” 的模型体验就更加弱智 就好像是抽一鞭子 才走一步路的感觉,体验甚至不如ds

然后是代码工具的请求格式 我用axonhub网关转出的 /message格式
/chat这种openai兼容的格式 感觉和我在用chatbox 硬聊一样 不如/message 格式写代码好用

我过去的一个月 假如能早点知道这些事情 我也不用浪费那么多token在aider上了 那个真的体验太差了

9

学习

10

b8c54057a84f2afda38b3358192f1f62
b8c54057a84f2afda38b3358192f1f621080×1080 96.3 KB

11

学习了，感谢楼主！

12

你试了 MiniMax、DeepSeek、GPT-5.5、Claude code、Codex……其实主流模型今天的能力，做一个选品软件早就绰绰有余了。**做不出成品，99% 不是模型不行，而是工程方法不对。**你试了 MiniMax、DeepSeek、GPT-5.5、Claude code、Codex……其实主流模型今天的能力，做一个选品软件早就绰绰有余了。做不出成品，99% 不是模型不行，而是工程方法不对。

我分享几个我vb的心得：

  1. 先让 AI 写"计划"，再写代码：我想做 X，请先列出实现步骤和需要的技术，不要写代码。
  2. 报错原文直接整段粘贴： 把完整红色报错 + 你执行的命令 + 相关代码一起丢给它，加一句"先解释错误原因，再给修复方案"。
  3. 让它解释每一段代码： 这是没有基础的人积累判断力最快的方式，比单独看教程有效。
  4. 善用"重开一个对话": 当一个对话越聊越乱、AI 开始鬼打墙时，别死磕。新开对话，把当前能跑的代码和目标重新讲一遍
  5. 让 AI 帮你写注释和 README:做到一半经常忘了之前在干嘛。让它给代码加中文注释、生成一个简单的"这个项目怎么跑起来"说明
  6. 用 AI 搭"脚手架"，自己填业务逻辑:环境配置、目录结构、依赖安装这类繁琐又标准的活，全交给 AI；你把精力留给"我到底想要什么"。

踩坑：

  1. 一次让它改太多，崩了不知道哪错
  2. 盲目相信 AI 给的代码能直接跑
  3. 迷信"换个更强的模型就能成"
  4. 让 AI 一次性"全自动部署上线"
13

Thank you for your guidance.

14

Thank you for your guidance. Indeed, having a closed-loop way of thinking is the most important thing. A lot of times, I actually have a complete closed-loop approach when I work on projects, but what’s really frustrating is that I always hit a particularly tough bottleneck halfway through, and then I get completely stuck and can’t figure out how to keep going. It’s really tough.

15

你遇到的问题和编程/vibecoding没关系啊，你是现实业务/资源有缺失，比如商品数据。

16

有一种东西叫审美，AI能给你生成小说、文章、视频、音乐、代码，但到底好不好，还是需要你人去判断。审美这东西是需要培养的，你懂吧。你问AI，什么音乐好听，哪个songwriter厉害，AI回答了你某某某，你去搜了搜，听了听，感觉上确实厉害，但你说不上为啥厉害，你也不知道和其他歌曲差别在哪，这就是问题所在。你设计APP，设计网站，你知道有种特效很好看，但你说不上来。你脑子里能构建出你希望的网页长啥样，但你不知道怎么给AI描述（摊手）。AI确实很厉害，但专业性的知识还是得学。