I used Cursor+AI Studio, combined with GPT and Gemini, to create a company level Crypto intelligence monitoring and semi-automatic publishing system from scratch. 24-hour uninterrupted scanning of the entire network information through models such as RSS, Grow/GPT/Gemini, automatic cleaning, automatic scoring, and automatic writing, and then automatic distribution based on the "traffic algorithm". Many people ask me, 'With AI programming so popular now, how do you play Vibe Coding?'? ” Today, I will directly release the entire process of this project from requirement alignment, engineering disassembly to final implementation. This is not only the birth of a tool, but also a methodology for ordinary people to master AI. 1、 Why do we need to do this? (Entropy increase of adversarial algorithm) Human beings have their limits. When we sleep, we become emotional and miss information. But AI won't. So I had an idea: could I become an 'AI agent' and complete 90% of the dirty and tiring work for me? It is responsible for 24-hour monitoring (information breadth). It is responsible for cross validation deduplication (information accuracy). It is responsible for preliminary grading and writing (information processing). And I am only responsible for the final 1% of the "decisions" and "finishing touches". This is the prototype of my system - XAgentic Semi Automation. 2、 The core of Vibe Coding: You are not a coder, you are a product manager Many novices fail to use AI to write code because they immediately make the AI "write a greedy snake". The first principle of Vibe Coding is that code is only the result, requirements are the soul. Day 1 (50% of the time): Pure chatting. Compete with GPT on requirements and use "torture" AI to clarify logic. On the morning of the second day (25% of the time): Generate documents and have different LLMs spray each other (cross review). On the afternoon of the second day (25% of the time): It's time to write code (AI Studio writes the front-end, Cursor writes the back-end). Step 1: Violent Alignment Needs (Chat to PRD) I didn't write a single line of code, but treated myself as the client. I threw all the functions I could think of to GPT: I need to monitor RSS and use LLM to scan Twitter. I need to go heavy, I need to rate, and those with insufficient scores should not be added to the database. Those with high scores should be added to the traffic pool Then ask GPT to help me organize it into a * * PRD Requirements Document * *. Step 2: LLM Cross Review (AI Confrontation) This is the most exciting step. I took the requirement document written in GPT and threw it to Claude and Gemini, saying: You are a senior architect. Please launch a crazy attack on this document from the perspectives of project implementation, cost control, and data consistency, and identify its logical vulnerabilities The results really pointed out many pitfalls that I didn't expect: How to prevent AI from hallucinating and shouting orders randomly? ”->Solution: Introduce a validator node and whitelist. What if running multiple models together is too expensive? "->Solution: Discover multiple models at different layers and handle layer convergence. How to prevent being judged as a robot? "->Solution: Simulate the release rhythm and rest time of humans. After this round of "virtual CTO" review, I have obtained a complete PRD. 3、 System architecture: Writing 'traffic thinking' into code The core of this system is not crawlers, but the automated judgment of "traffic" and "value". I divided the output into two pools, which perfectly fit the algorithm mechanism of X: 1. The Traffic Pool Logic: Targeting breaking news and highly discussed topics. Requirement: Speed comes first. Configure Rss information sources and prompt words to obtain information, and then directly determine whether it has traffic value. 2. The Research Pool Logic: Targeting narrative changes, macro analysis, and project depth. Requirement: Depth first. By configuring some VC and investing in research institutions' Rss to obtain in-depth research reports, project analysis, etc., and then automatically entering the traffic pool through scoring. The whole process is like this: Discovery: Grow/GPT/Gemini polling+RSS crawling. Filter: AI automatically removes duplicates (filters out homogeneous junk information). Score: Score based on influence, timeliness, and information density. Route: Discard low scores and put high scores into the pool. Draft: Claude/Emini generates a draft. Review: Every morning, I open the backend like an emperor approving a memorial: click "approve" or click "not interested" (AI will remember my taste and will not push such information again next time). 4、 Results and Delivery In the end, I used AI Studio to handle the high-end frontend (Dashboard), and used cursor to handle the backend logic and database interaction. The final presentation is shown in the video: On the left is the original evidence (guaranteed not to spread rumors). The middle is the logical judgment of AI (why it feels this is important). On the right is a pre written draft of the tweet (I can even edit it directly). The biggest challenge at present is the optimization of information sources. Because AI is too powerful, it will also eat up the garbage of the entire network, so I am currently doing a 'whitelist cleaning'. 5、 For those of you who want to try Vibe Coding During the past period of VibeCoding, I have come to understand one thing: in the era of AI, programming languages are no longer the threshold, imagination, logic, and clear requirement documents are what truly matters. If you also want to make your own tool, my suggestion is: Don't rush to write code. Spend 80% of your time polishing your PRD (requirements document). AI writes code quickly, but if you give junk instructions, it also generates junk code. Make good use of the 'AI judging panel'. Let AI fight against each other, it can help you fill 99% of logical loopholes. Start with 'workflow'. Don't write an app just for the sake of it, think about what action you repeat the most every day and automate it with AI.