Q: How do you assess whether someone is very good at using AI? A: Look at how much they spend on AI; if they can break $300 a month. - Breaking $300 indicates they have subscribed to quite a few AI products, meaning they have at least experienced them and found them effective. - Those who can continuously spend money on AI show that they are gaining higher value from it. - If AI downtime affects their normal work, it means their work patterns have already changed, not to mention they would cancel subscriptions for useless AI. - $300 suggests they might have activated some premium subscriptions, which is a signal of high demand from a geek. I don't believe someone can claim they are good at using AI while only using a $20 or even free API plan, saying they have a lot of practical experience. This person might understand one aspect, but it is absolutely impossible for them to have a comprehensive understanding. Honestly, just running an agent to manage a bill would cost a few hundred dollars a month.

Teaching chatGPT, midjourney, and stable diffusion in 2023 Teaching comfyui, dify, and coze in 2024 Teaching deepseek, office vibe, and coding in 2025 I took a look, and 60% of the teachers are the same people 😂 The content changes, the character labels and titles change, but the formula is the same. Platforms invest 199 to turn small into big, or just publish a book to lead a course. Anyway, the biggest hurdle for teaching others to get started might just be swallowing their pride. There are plenty of 40-point teachers everywhere, and courses selling for thousands aren't bad either. If we say people aren't FOMOing over AI, I think that's hard to argue. But it seems there aren't many who can actually use it 😂

47 Claude Code Tips Just by saving them, you will know them.

This script and voice model are so powerful. Is there anything else about Guo Degang talking about the Ming Dynasty...

Don't ask me how to vibe with coding Here's a set of prompts for you, no need to thank me.

To create something that can go viral, it must hit the following points: - Show off: I share this thing to let others understand me and think I'm impressive and knowledgeable. - Fun: I share this thing to bring joy to others. - Social recognition: I share this thing because it has value to others, and I can gain social recognition. - Actual value: I share this thing to make money or gain attention. - Herd mentality: Everyone else is sharing, and if I don't share, I seem out of place. Whether it's content or a product, the more elements it hits, the more likely it is to go viral. If it can also form a profit chain, then it can spread even further. The above is from the perspective of sharing motivation, but sharing is a funnel; others must first see it, then understand it, before they will share it. So, exposure must come first. However, there are some differences between wanting exposure and wanting to share. Exposure mostly relies on comments, and the logic of stimulating comments is different from that of stimulating shares.

I added Opus 4.1 on Readdit to look for posts and found mentions of two MCPs that enhance ClaudeCode's capabilities: Zen and Serena. However, opinions are completely different; some really like it, while others think it's completely useless. I believe that's how AI is: when it can deliver great results, people love it, but when the results are zero, people hate it (because it wastes their time). If an AI product has a small group of people saying it's amazing, then it's worth a try. Has anyone tried these two MCPs? Please comment on your experience!

Anyone can do a VibeCoding mini project, like this -- a Reddit content listener. It can actually be used to find study materials, translate content, and locate trending posts in specific subreddits, which helps you gradually get familiar with the style of that subreddit, making it easier to write popular posts. I personally use it as an information source, and it can even be used to learn English. Agencies use it to find content inspiration, brands use it to monitor events related to other brands, and I mainly work on it to prepare for content marketing on Reddit, building an automated prompt information source for my Reddit Agent. Limited beta test code: RDDT-TDGIA-DIZ [Real beta test, already deleted the database once, haha]

Here's another idea for unemployed workers to make money. Use vibecoding to create iOS client applications, specifically skeletons. What is a skeleton? For example, a categorized image gallery or a categorized audio listening app. It's just a list that categorizes the content, making it easier for users to find. Once you have a skeleton, you just need to keep the content interchangeable. Now, vibecoding makes this very fast, and it's easy to adjust for you. It's also relatively hard to get flagged by the app store, essentially like AI skinning. All you need to do is organize the content and package it all together. Then, use something like Qimai to check ASO keywords. Just lay it all out there. The product is free, filled with ads, and an 8-dollar permanent membership removes ads. You can create something like "Friends" or "Modern Times," or even some celebrity courses. Don't feel guilty about piracy; it's fine. You can even use Li Kui and Li Gui to ride on others' brand names. It doesn't matter if the downloads are junk; let them uninstall as they please. Previously, someone used vibecoding to generate tools for appconnect, continuously creating content for listing. This process doesn't take much time. On your first day, you might take 3 days to list one app, but later you can list a new application in just 3 hours. If you keep doing this for a hundred or so applications, you'll always find long-tail keywords that rank. This thing adds up over time, and the core point is that you've packaged everything locally, with no resource consumption, creating zero-cost assets. Find what can generate volume, and then invest in serious optimization. In the AI era, there are just too many arbitrage opportunities like this. You can learn while you work; there's nothing you can't handle. The most important thing is to believe in yourself and in AI. Don't magnify the difficulties for yourself. If you do 20 of these, you'll naturally find the way and gain experience.

I took a look at the Prompt101 shared by Anthropic. There are two points that I think are worth sharing to enhance usage: The first is called Pre-fill Response. >What it does: It provides a starting point for the model's response directly in the API call. >Why: This is an effective technique to force the model to output a specific structure (especially JSON). >Example: In the "assistant" section of the API request, directly fill in a left curly brace { to guide the model to complete a full JSON object. This pre-fill makes it easier to guide the model to respond according to your output pattern because LLMs are like a game of word association. The second is Prompt Debugging, utilizing the Extended Thinking / Scratchpad feature. >Function: This is a powerful debugging tool. When this feature is enabled, the model will explicitly show its "thinking process" or "scratchpad." >How to use: By analyzing the model's thought records, you can understand where it encountered difficulties or misunderstandings. Then, you can turn these insights into clearer step-by-step instructions, solidifying them into your prompts, making the prompts themselves stronger and more efficient. This is actually an idea shared by Yulong in 2023, which is that when the output of the prompt does not meet expectations, it is necessary to communicate with the model to confirm which part of the prompt the model misunderstood semantically. Printing thoughts is also a variant of this idea. That's what I've learned; the video may have other information for learning 👇