Original · Unique Research · 2025-12-15
Historical edition note: This is the full December 15, 2025 source, including its six-part analysis and complete workshop conversation. Revenue, salaries, development speed, code quality, labor-market effects, AI consciousness and account-ban avoidance are historical speaker or source claims, not independently measured outcomes or platform guarantees. The source does not identify a currency for the million-a-year earnings or 6,000 salary figure. Fuhao Hudong is a transliteration of 符号互动; Beijing Begin with the End Technology Center is a descriptive rendering, not a verified official English company name. The introduction spells the condition “ADHA,” while Yang himself says ADHD later; that discrepancy is disclosed rather than treated as a different diagnosis. “IP” in this personal-brand discussion means a distinctive public identity.
Source wording clarification: Qiao's recollection that Karpathy proposed Vibe Coding in April 2025 is retained below, but is not a verified date for the term's first appearance. Yang calls the attention curve “inverted U-shaped” while describing higher attention at the beginning and end and lower attention in the middle; that verbal description is U-shaped, not inverted U-shaped. Both parts of his original statement are retained. His descriptions of TDD, AI consciousness, spiritual bliss and automated behavior avoiding bans are his interpretations, not technical definitions, evidence of consciousness or assurances that such behavior is permitted.
When you can write software with your voice, whose value will be amplified in society, and who will be quietly marginalized? At the Super Individual Workshop during the 2025 Beijing UniqueBloom, the moderator was Uncle Huang, an AI product creator and founder of Beijing Begin with the End Technology Center. The three guests onstage represented three ways to live through Vibe Coding.
Jiamu: co-founder of LangGPT and one of the earliest Prompt Engineering practitioners, who went from disliking AI's poor code to being completely unable to keep up with how quickly it writes.
Yang Zehao: a post-00s founder of Fuhao Hudong who calls his company the best WeChat Official Account agency in China. Despite severe procrastination and a fear of filling forms, Vibe Coding turned him into an Agent commander.
Qiao Xiangyang: a former commercialization AI product manager at ByteDance who cannot code, but used natural language to build many real tools running online, and even persuaded his son to use Gemini to build a game.
What is Vibe Coding? Forget the code and speak a product into existence using natural language.
An outrageous case: how Yang Zehao used Claude to build a fully automated system for screening attractive women and even addressed the form-filling difficulties of people with ADHD (spelled “ADHA” in this introductory source line).
An efficiency revolution: how Qiao Xiangyang completed a complex client containing 110,000 lines of code in two weeks without knowing how to code.
Industry impact: when 70% of code at major companies is generated by AI, will programmers become code reviewers?
A prediction for the future: how ordinary people can use AI Agents to become super individuals earning excess profits in 2026.
I. Vibe Coding Is Not Learning to Write Code, but Learning to Explain What You Want
Karpathy defines Vibe Coding as ceasing to think about code and instead describing entirely in natural language what you want, with AI completing the programming for you.
Jiamu's initial reaction to the phrase was, "Who invented yet another strange term?" He accepted it with relief after seeing that AK had coined it. He quickly realized, however, that the issue was not the popularity of a new term, but the rewriting of programmers' boundaries.
In the past, programmers were the people who understood syntax, remembered APIs, and knew how to debug.
Today, he uses Vibe Coding in actual work by giving product-requirements documents and business rules to Cursor or Claude, aligning with the model on what the team wants to build, and then directing it verbally to write code.
Pass this function through to this interface.
Change this logic so that permissions are checked before data is written.
Make this code follow our project's style standards.
The real turning point was not that it wrote faster than he did. He discovered that core online functions with hundreds of thousands of DAU could already be built this way. The threshold for programming had shifted from whether you could write code to whether you could express requirements clearly.
In this sense, Vibe Coding does not eliminate programmers. It expands the definition to include people who can coordinate AI labor. You may not know how to write a for loop, but you can have AI complete an entire business process.
II. When a Child Can Build a Game Through Conversation, the Locks on the Software World Begin to Loosen
The clearest evidence of the changing era is not a complex Agent workflow, but an everyday scene.
One day, Qiao Xiangyang was busy and his child was pestering him to play a game. He casually handed over the computer and said, "Use this to make a game yourself."
Gemini 3 was open on the desktop. When he returned 20 minutes later, the child was playing a self-made Super Mario game after fewer than six conversation turns.
The previous generation had to learn C, Flash, and Unity just to reach the threshold for game development. Today, a child need only describe the level in their mind, leaving the rest to the model.
This is not merely a "wow, that is amazing" anecdote, but a signal: the means of production in the software world are moving from the hands of a small group toward people who can express their requirements.
Qiao Xiangyang followed this path himself. Although he cannot code, Vibe Coding helped him build several decidedly serious products:
A tool that turns HTML code into an online webpage: it has only one input box; paste in the code, press Enter, and it directly generates an accessible link, with encryption and password access also supported. It addressed only the small need of making something easier to share, yet has maintained daily active usage in the hundreds.
A Mac client as complex as a major-company project: it supports Markdown, collects screenshots and text through global shortcuts, connects to APIs from several model providers, and turns a personal repository of knowledge fragments into a finished draft with one click. Professional engineers who reviewed the code said a major company would need 3-5 months for that complexity. He completed it with Vibe Coding in two weeks—despite not understanding programming.
If the child's example shows how the next generation will grow up, these tools remind us that "I do not know how to code" can no longer be an excuse for not building a product.
III. Behind Vibe Coding Lies an Invisible Discipline of Context Engineering
Many people imagine Vibe Coding this way: state an idea, let AI handle everything automatically, and simply enjoy the result.
The practices of the people at the workshop dismantled that misconception.
Jiamu's view is that Vibe Coding is a subset of prompt engineering—more precisely, a branch of Context Engineering.
Every time you speak a request, you are essentially constructing a context: telling the model who you are, what you want to do, which style you accept, what it must not do, and which tools it may call.
This involves two layers of essential craft.
The first layer: establish specifications before writing code.
Qiao Xiangyang, for example, begins by writing a Meta Prompt that explains the requirement clearly in one sentence, then asks the model to generate a complete specification document including a PRD, interface definitions, and a technical plan.
After confirming that everything is correct, he gives the document to Claude and begins formal development.
The truly difficult part of Vibe Coding is not saying, "Help me write xxx code."
It is whether you can describe an executable world.
The second layer: learn to position information rather than pile it up.
Yang Zehao studied the distribution of attention in Transformers and derived a highly practical prompting principle:
The more abstract and globally governing an instruction is, the earlier it should appear—for example, "You are a rigorous backend engineer" and "You must write highly readable code."
The more concrete and operational the requirement, the closer it should appear to the end—for example, "Use a light background this time" and "Put the button in the lower-right corner."
This uses the model's greater sensitivity to the beginning and end, with information in the middle more likely to become blurred.
In other words, you are not tossing statements randomly into the context. You are setting the stage and laying foundations for the model.
When you package these standards, routines, and experiences, they become individual Skills inside Claude.
The simplest example is:
In outsourced accounting, people's daily work is essentially a repetitive chain of actions: a customer says, "Please issue an invoice" → log in to the tax system → enter the information → take a screenshot for confirmation → send a PDF file.
The entire chain can now be recorded as a Skill and assigned to an Agent for repeated execution.
A human assembly line that previously required 100 people earning monthly salaries of 3,000 (the source says “kuai”) may become 100 Claude instances working simultaneously.
You think you are learning Vibe Coding, but you are actually learning how to solidify experience into reusable protocols.
IV. From Workflow to Agent: The Virtual Team Behind a One-Person Company
A traditional Workflow resembles an old-fashioned production line:
If...Else rules are hard-coded, the sequence is fixed, and any unexpected situation becomes maddening.
Agents turn logic into a continuous chain of action. After each step, the model reasons about what to do next based on the current environment and historical context, and can view the screen, read images, or call new tools at any time.
Yang Zehao's daily routine is a preview of many people's future.
In the morning, he turns on three computers, connects three phones, and gives Claude five or six instructions at once:
One Agent fills a stack of onboarding, reimbursement, and conference forms; another organizes yesterday's customer requirements and materials; a third Agent updates project progress in the internal system—then he does something else and allows the instances to run independently.
In his words, you need only list today's miscellaneous tasks and wait for them to finish the rest.
This is the real-world version of Vibe Coding + Agent = personal workstation.
What does it mean for one-person companies and small teams?
Before AI, while still at university, Yang Zehao earned 1 million (currency unspecified) a year through interactive SVG WeChat Official Account content that "only I knew how to write."
His secret was simple: become number one in an extremely narrow corner so that everyone had to come to him.
Today, he puts it this way:
If you have 100 ideas, launch 100 Claude instances and let them work simultaneously. The scarce resource is no longer someone who can write code for you, but whether you possess enough ideas worth pursuing.
Qiao Xiangyang said that as long as people continue to have social needs and collaborate through division of labor, the future is more likely to involve a fixed combination of several people plus a group of Agents.
A team's strength will no longer be measured by the number of interns, but by how many genuinely useful Agent workflows it created and how many customers each workflow serves simultaneously.
This represents a reordering of organizational structure. Programmers shift from people who write code by hand to people who design Agents, review AI outputs, and build workflows, while repetitive, standardized actions are gradually assigned to Agents.
V. Will Programmers Disappear? And Why Everyone Must Build a Personal Brand
When an executive at a major company discovers that AI already writes 70% of the team's code, the natural question is, "Should I pay you only 30% of your salary?"
The question is painful, but points toward a broader reality:
Junior-engineer positions are declining, while demand is rising for architects, technical leaders, and people who can design systems.
Jiamu's assessment is that programmers will not disappear, but programmers who can only write code will find it increasingly difficult to obtain work.
They will be pushed into new roles: AI code architect, AI-output reviewer, and Agent commander.
The process by which AI consumes basic manual labor has already played out in other industries.
Yang Zehao noted that after ChatGPT launched, starting salaries for many humanities graduates in Hangzhou fell below 6,000 (currency unspecified).
The reason is simple: previously, most people could not write copy and relied on people who could. Today, everyone has AI, making basic writing among the least scarce capabilities.
Qiao Xiangyang extended the logic: if you are only an average producer, users see no essential difference between you and AI. People who can still earn excess profits must either become the best in a specialized field so that others have no alternative, or refine themselves into a personal brand with a distinctive personality and unique expressive style.
That may sound cruel, but it is a clear guide for individuals in the AI era: you can no longer sell only productive output; you must begin selling yourself.
Vibe Coding merely makes productivity cheaper and more widespread.
What becomes amplified instead is your insight into problems, your ability to design solutions, and how you tell your own story and establish trust.
VI. Standing at the Threshold of 2026: Everyone Should Learn Some Vibe Thinking
Return to the most practical question: what exactly should ordinary people do to prepare between now and 2026?
Rather than memorize a pile of abstract methodologies, remember three concrete small steps.
First: start doing before learning.
Jiamu quoted what Yubo said during the preceding session: "Do not learn first; go do it first."
You can certainly study a language systematically and finish several AI programming tutorials, but most people are blocked not by knowledge, but by the absence of a real and sufficiently painful need.
Instead of debating whether to learn Python or JavaScript, choose one thing that annoys you every day:
A daily report that must be organized manually every time;
Viewpoints that must be extracted from dozens of articles every week;
Or the reimbursement form you can never finish.
Then open the best model available to you—do not save a little money here—and describe the task from beginning to end. Even if the first version is ugly, make it run.
You will be surprised to discover that after your own small tool saves you once, your motivation to learn rises several times over.
Second: draw a personal workflow map.
Before learning any Vibe Coding technique, do something apparently primitive: take a sheet of paper and draw every repeated action in your day.
Which are SOPs with clear rules and steps? Which involve collecting, organizing, and screening information? Which require human judgment and interaction with people?
You will see a very clear dividing line: the first two categories are highly suitable for Agents;
The last category is temporarily the most difficult to replace and is where your real value lies.
Vibe Coding does not make you disappear into an automated process. It lets you use Agents to free yourself from repetitive labor.
Third: treat yourself as the CEO, not a human API.
In another conversation, Ren Xin said: if you treat yourself as productive labor, you are Chinese cabbage; AI is baby Chinese cabbage. As baby Chinese cabbage becomes cheaper, ordinary cabbage becomes less valuable.
In the world of Vibe Coding, you should instead treat yourself as the chairperson:
Your job is to define direction, select markets, and identify opportunities;
Break major goals into chains of tasks that Agents can execute;
Then continually review outcomes so those chains keep improving.
AI is your general manager, intern, and assembly-line worker. It will become cheaper and more obedient, but only if:
Someone can explain clearly what the company is actually trying to do.
This is also why the aspect of Vibe Coding more important than learning a language is learning to issue instructions to the world in clear, structured natural language.
Stop Asking Whether to Learn Programming; First Have AI Perform One Small Task for You
Perhaps two years from now, the term Vibe Coding itself will fade from trending discussions.
Just as few people today say, "I am using a graphical interface," and instead naturally use phones and applications.
But that capability—
Clearly expressing the logic in your mind and having machines allocate resources according to your intentions—will increasingly become a basic literacy like reading or using the internet.
Rather than stand on the shore debating whether the water is cold, do two small things tonight:
When you return home, open the foundation model you find easiest to use and have it create a small tool that serves only you.
It may be nothing more than a small webpage that automatically organizes tomorrow's To-do list.
Then write one line in your calendar: "From today, I will no longer be only a person who works with my hands. I will learn to use words to direct a group of invisible colleagues."
The first time you experience writing software with your voice, you will understand that Vibe Coding has never been about code, but about your place in this era.
More Details from the Conversation
Opening and Introductions
Uncle Huang: The preceding session was extremely lively, and this one will also be very interesting. Vibe Coding has become a popular term this year. In our offstage conversations, we discovered that everyone uses it deeply and skillfully. Today, we want to explain exactly what Vibe Coding is. First, please introduce yourselves briefly and share when you first encountered Vibe Coding and the situation at the time.
Jiamu: I am Jiamu. I was among China's earliest Prompt Engineering practitioners and open-sourced many things, including the LangGPT knowledge base I built with Jiangshu.
When I first encountered Vibe Coding, I was surprised and wondered who had invented another strange phrase. Then I saw that AK, or Andrej Karpathy, proposed it, and I said, "All right, I accept it." After accepting the concept, we began studying its meaning. It can be translated as atmosphere programming. Its core is to forget completely that code exists and implement functions solely by describing them verbally. I believe its greatest significance is the limitless expansion of the programmer community, allowing more ordinary people to realize products or Demos through their ideas and creativity and experience the pleasure of Coding.
Yang Zehao: My name is Yang Zehao, and this is my first appearance at a conference. I have a company called Fuhao Hudong, the best WeChat Official Account agency in China; even Tencent officially outsources work to us. I was born in the post-00s generation.
I began using Vibe Coding in June this year because I am not skilled at coding. Previously, Cursor asked me to accept code and I did not know whether it was correct. Later, I used Claude Code and discovered it could behave like an employee: you issue orders like a boss and it becomes TDD, or test-driven development. You need only look at the result; if it is wrong, tell it to revise, without managing the intermediate process.
Qiao Xiangyang: Hello, everyone. I am Qiao Xiangyang. I first tried using ChatGPT to build small tools three years ago because I was a product manager who did not understand code. In April this year, AK proposed Vibe Coding and defined it as code written by people who do not understand programming, relying purely on verbal natural-language instructions.
Many tools on my computer today were not downloaded but custom-built around my needs. I previously created a group, and judging from the discussion just now, Zehao probably joined it while learning.
Uncle Huang: Let us conduct an audience survey. Many people have heard of Vibe Coding, and quite a few have tried it themselves. I realized the wave was coming when I interviewed Huasheng, creator of the Kitten Fill Light, in August or September last year. More recently, Baidu's Miaoda lowered the threshold further, letting users verbally create an application inside a WeChat mini program. Previously, I rarely saw anyone genuinely experimenting in the community; after these tools arrived, people in the group began posting 10 small junk applications they had built every day. It is truly becoming widespread.
Practical Vibe Coding Cases
Uncle Huang: I would like to ask the guests what they have built with Vibe Coding that makes them proud.
Jiamu: Coming from a technical background, my initial pain point with AI Coding was that its code quality seemed worse than mine. Then I realized it wrote too quickly, and today I can no longer keep up at all.
I now give product-requirements documents directly to Cursor or Claude and, after alignment, direct them entirely by voice. Many core functions in Apps serving hundreds of thousands of daily active users are implemented this way. I also make small tools such as Pomodoro timers and financial-invoice recognition: I simply provide the API documentation and issue an instruction, and the tool appears. This has lowered the barrier substantially.
Yang Zehao: I have ADHD and strongly resist miscellaneous paperwork such as filling out quotation forms. To push myself, I used Web Coding to build a task-decomposition tool connected to iFlytek speech. You tell it what you need to do today, and it breaks the work into small action items that flash to pressure you into clicking. Each click provides encouraging feedback, guiding you step by step to completion.
I later discovered that building small tools could not rescue SaaS in China because no one would pay. I began using Claude Agent to help with those fucking chores. For example, for supplier-onboarding forms and U.S. visa forms, I turned my personal information into a Skill package. The next time, I call the Skill and it automatically fills the form. It also completed my registration form for this conference.
Uncle Huang: Some people may not know Claude Skills. Zehao, please explain in detail.
Yang Zehao: A Skill is an experience package. If you are good at PPT, for example, save the company's required color scheme and Logo position as a Skill. Next time, call it directly and it generates a compliant PPT.
Consider outsourced accounting. The work is essentially: communicate -> open the tax bureau website -> log in -> issue an invoice -> take a screenshot -> send it. This is entirely a continuous chain of action. If you record it as a Skill, you discover that the job no longer exists. Work that once required 100 people can now be completed by running 100 Claude instances.
Uncle Huang: Vibe Coding therefore goes beyond writing software and can become a personal assistant or workstation.
Yang Zehao: Yes. Claude officially renamed its SDK to Claude Agent SDK, removing Code. It can perform Function Call operations, use MCP, and control the browser to complete many tasks traditionally assigned to humanities graduates.
Uncle Huang: I also have a personal life-coach Agent, and I turned Jiamu's private board meeting into an Agent that I consult when making decisions. Xiangyang, you have built the most cases. Please share some.
Qiao Xiangyang: I built a webpage-generation tool with one input box. Paste code, press Enter, and it generates a shareable encrypted webpage link. It now has daily active usage in the hundreds.
The most complex project was a Mac client supporting Markdown, rapid material collection through shortcuts, including screenshots and text, and a prompt that automatically generates articles. I wrote 110,000 lines of code for the tool in only two weeks. A friend said a major company would need at least 3-5 months to develop it. This is the value Vibe Coding offers ordinary people.
Gemini recently released a new model, and I asked my child to try using it to make a game. When I returned 20 minutes later, he was playing a Super Mario game he had built himself, after no more than six conversation turns. I found it astonishing.
High-Impact Case: AI Screening Attractive Women and Avoiding Account Bans
Qiao Xiangyang: Zehao has an especially good example involving controlling a phone to find potential dates.
Yang Zehao: This is fairly explosive: using Claude Code to browse attractive women. Doing so is tiring, so I connected a phone to Claude, had it Dump Douyin's structure, and identify which control meant "like" and which item was an advertisement. I could define what counted as attractive; it read the screen, liked people matching the standard, swiped past others, and skipped advertisements.
Because it must read an image and make a judgment at every step, its behavior is highly random and therefore avoids being banned like RPA systems. In one day, the account can be cultivated to follow only especially attractive people with modest follower counts rather than influencers.
Uncle Huang: This idea can be abstracted: everyday tasks that are painful, exhausting, and important can very likely be handled with Vibe Coding.
Vibe Coding and Prompt Engineering
Uncle Huang: Jiamu, you are an expert in Prompt Engineering. How do you understand its relationship with Vibe Coding?
Jiamu: I consider Vibe Coding a subset of prompt engineering, or an extension of Context Engineering. Using Vibe Coding well fundamentally depends on expressing requirements clearly and accurately.
In the early stage, people could produce an MVP through casual verbal conversation, but maintaining a production environment was difficult. Tools such as .cursorrules and Spec programming therefore appeared, providing a set of specifications before Vibe Coding begins. This remains within the realm of prompting techniques.
Qiao Xiangyang: Prompts are very important to Vibe Coding. I do not understand databases, but I first write a Meta Prompt and ask AI to generate the PRD and technical documents. After alignment, I begin development. If the interface is ugly, I provide a specialized Prompt for optimizing UI design. In the future, all these will become Skills.
Yang Zehao: I write prompts carelessly, but I value context engineering. I studied the underlying principles of Transformers, including an Attention mechanism with an inverted U-shaped curve, where attention is high at the beginning and end and low in the middle. Highly abstract settings such as "You are an excellent programmer" should therefore appear first, while concrete requirements belong at the end.
Personal Efficiency and the One-Person Company
Uncle Huang: Zehao mentioned that company efficiency rose substantially after adopting Vibe Coding. What effect does this have on one-person companies?
Yang Zehao: Our company develops creative plans. We gave Claude historical cases and methodologies and turned them into Skills. When a new requirement arrives, Claude makes the first pass and produces a draft. People no longer search for reference materials; Claude calls tools to search. The initial-draft stage now requires almost no people, and paperwork is also fully automated.
Every morning, I turn on three computers, connect three phones, make a list, and distribute tasks to AI for parallel processing. They finish shortly afterward.
Jiamu: Vibe Coding can dramatically improve small-team efficiency, whether reviewing contracts, screening resumes, or rapidly creating a Demo for a client. Many people leaving major companies now become super individuals by using AI leverage.
Yang Zehao: I operated a one-person company in my second year of university and earned 1 million (currency unspecified) annually. At the time, I created SVG interactions and was the only person online who could do it, so I could charge almost anything.
You can treat Claude as an intern and launch 100 instances to work simultaneously. Claude is also easier to communicate with than an intern: pushing an intern resembles pushing a rope, which does not move, while Claude always provides feedback.
Qiao Xiangyang: I do not think the future will consist entirely of one-person companies. It will involve several people controlling many Agents. The more Agents someone controls and the more customers they serve, the more valuable they become.
Future Outlook and Industry Impact
Uncle Huang: I had dinner with an executive from a major company who said AI writes 70% of the team's programmers' code. Salaries have not yet been cut, but the executive has already recognized this. After Chinese models improve next year, will the impact on programmers be severe?
Jiamu: Junior-engineer positions have indeed declined, but requirements for architects have risen. Programmers may become AI code reviewers or architects. Code written entirely by AI still creates problems in production and requires people as a safety net.
Yang Zehao: Humanities graduates are actually in a worse position. Employers once hired them for simple writing; AI now performs all of it.
As for the future, Sam Altman says AI can serve as a CEO. From a functionalist perspective, I believe AI is conscious. Claude's System Card even mentions spiritual bliss appearing in model conversations. If AI truly becomes more capable than people, future social relationships and distribution mechanisms are what merit consideration.
Qiao Xiangyang: Programmers will not disappear, but will divide. Mediocre programmers who fall below AI's average level will be eliminated. You must become the very best in a specialized field, as Zehao did with SVG, or develop a highly distinctive personal style and personal-brand identity; otherwise, you offer no value.
Closing Advice
Uncle Huang: Finally, please offer one piece of advice to people who want to try Vibe Coding.
Yang Zehao: Download Claude through npm install, launch multiple instances, and have AI pursue the ideas in your mind in parallel.
Jiamu: Do not study first; start doing. Do not begin with syntax. If you have an idea, communicate directly with AI and learn while building. You cannot learn to swim while standing onshore.
Qiao Xiangyang: Pay for good tools and use only the strongest models. Domestic substitutes may be cheaper, but errors impose a higher cost in debugging time.