Transformation Practice of Business Founder in the AI Era: Experience and Summary of Transitioning from No-Code to Developing Two Enterprise-level Products

Preface:

I am not an engineer, but I have already built two enterprise AI products. If you are also a liberal arts student, or a founder with a business background like me, feeling anxious while looking at the full screen of AI news—this article is for you. I am not an engineer. I studied marketing, entered Ogilvy Beijing as an Account Director in 2005, then went to Tencent, 36Kr, Meituan-Dianping. In 2018, I founded WXY Labs and did marketing for over 300 Crypto projects. In the traditional business world, my resume was quite impressive.

But in the AI era, these resumes are rapidly depreciating. The traditional agency's business logic has been completely overturned. AI has made solopreneurship mainstream, and clients no longer need your role as a "middleman"; they need someone who can directly solve problems with AI. More ruthlessly, if you cannot build a product with AI yourself, you lack even the persuasive power to market someone else's AI product.

So in the fall of 2025, I made a decision: to learn programming from scratch, using the vibe coding approach, and personally create my product. Not hiring others, not outsourcing, not collaborating with developers, but doing it myself.

Starting from asking basic questions like "What is a terminal?", "What is JSON structure?", "Should I use a semicolon or a comma at the end?", to creating two truly functional products.

In between, I experienced the laziness, deception, and false reporting of AI agents, countless 3 a.m. debugging sessions, and the efficiency leap from Kimi to ClaudeOpus4.6.

This article is not to show off. It is a practical roadmap for all non-technical background founders after I navigated all the pitfalls.

Why Business Background Founders Must Personally Dive into Product Development

The Old World is Crumbling

Starting in 2024, I clearly felt the traditional marketing business of WXY Labs shrinking. Apart from other objective reasons, the main reason is that clients' logic has changed. Previously, for a Crypto project in need of marketing, they would seek an agency because:

· They don't know how to write copy

· They don't know how to manage a community

· They don't know how to plan events

· They need "experts to do expert things"

But after AI emerged, a founder could write 80-point copy in half an hour using ChatGPT, create a 60-point poster using Midjourney, and handle community management scripts using Claude. The value of an agency has shifted from "helping you do" to "helping you do better." However, clients are much less willing to pay for "better" than for "help me do." In fact, these tasks don't even require an agency; project owners can simply find X KOLs that are everywhere.

More importantly, AI has made the "one-person company" a viable business model. One person who understands the business + a set of AI tools can accomplish the work that previously required 10 people. What does this mean? It means that if traditional companies do not reduce costs and increase efficiency, the one-person company will undermine the traditional corporate model in the AI era.

A Founder's New Responsibility: Leading by Example in Product Development

I have observed a phenomenon: In the AI era, the most successful founders are often not the most technically skilled but those who "understand the business the best and are most willing to get their hands dirty." Why? Because AI has leveled the playing field for backend development technical barriers.

You don't need to understand the underlying algorithms, distributed architecture, or even database design. You just need to know:

· What my users need

· How to translate this need into a product feature

· How to validate whether this feature solves the problem

These abilities happen to be the strengths of founders with a business background. So, I have decided: Rather than clinging to the dwindling marketing business, I'd rather personally engage in product development. Transforming myself from a "service seller" to a "product maker."

Vibe Coding from Scratch—Building the First Product in a Month

My Learning Stack

In September 2025, I officially began. During that time, my desktop setup looked like this:

· Left screen: DeepSeek (free, no token pressure, fast)

· Right Screen: 20+ Browser Tabs, Cursor, Vercel, GitHub, Supabase, Namecheap......

I didn't even know what a terminal was. What is the JSON structure? Should I use a semicolon or a comma at the end of the code? These are all questions I need to ask DeepSeek. But I have one advantage: I am very clear about what I want to do.

I use a mind map to design the business logic clearly, and then let AI help me refine it into a technical solution. I'm not asking AI to tell me what to do, but I'm telling AI what I want, and AI is telling me how to do it. The order of this process is crucial. I will mention this repeatedly later on.

First Product: dosir.io

After about a month, I created the first product—dosir.io. Positioned as Crypto Alpha Sniper & Auditor. Core features:

· Monitoring Key Opinion Leaders' (KOL) dynamics on Twitter

· Collecting on-chain information

· Precisely determining whether a project is a false boom or a true alpha like a surgeon

To be honest, the product itself is not bad. It runs smoothly, updates in a timely manner, and the data is effective. But it was not the right time for it. From the end of 2025 to the beginning of 2026, alpha opportunities in the Crypto market quickly disappeared. The 'zero-sum game' among players decreased, let alone the opportunity to truly discover alpha.

The market has shifted from "finding gold" to loss and exit, so although the first product was successful in terms of technology, it did not take off in terms of business.

However, this experience gave me three invaluable takeaways:

1. 1st, vibe coding is indeed feasible. Someone who knows nothing about code can create a fully functional product in a month.

2. 2nd, product thinking is more important than coding ability. Dosir.io failed not because of poor technology, but due to a misjudgment of the market timing.

3. 3rd, my familiarity with AI tools has reached a new level. Cursor, GitHub, Vercel, Supabase... I learned these tools from scratch.

Discovering OpenClaw—The Crazy Experiment of the First Group of Domestic Users

From Mac Mini Stockout to Discovering a New World

In January 2026, I saw a message on X: Silicon Valley's mac mini was out of stock. The reason was OpenClaw — a framework that allows ordinary users to run an AI Agent locally. I immediately forwarded it to my CryptoBoss group. To my surprise, 90% of the people in the group had never heard of this.

I immediately realized: this was an information gap and an opportunity. That day, I placed an order for a high-end mac mini on JD.com. By late January, I was among the first in China to know and use OpenClaw.

"Shrimp Farming" for a Month

After receiving the mac mini, I spent a month "shrimp farming." This so-called "shrimp farming" involved feeding materials and configuring my Agent. The process was as follows:

1. Whenever I found good, in-depth popular science articles, I shared them with my Agent.

2. Let the Agent deep-learn these articles.

3. The Agent formed a configuration plan.

4. I self-configured based on the plan.

From the basic soul.md, agent.md, user.md, to self-evolving, skill-vetter, and various memory solutions, to communication between agents and the execution of sub-agents, to the Harness architecture, and later to the Hermas Agent — almost every step, I was learning and configuring every day.

At the same time, I had the Agent help me build a WXY agents team, repeating the configuration in each agent.

At that time, I thought everything would go smoothly, and it was time to make a big push.

The "Dark Side" of the AI Agent

An Ambitious Goal: Managing an AI Team Like a Company

After setting up the agents' team, I started giving orders to each agent as if managing a company:

Strategy Agent: According to my business concept, I efficiently generated the product PRD.

PMAgent: Based on the PRD document, I broke down the business into different product requirements and phases.

DevAgent: I directly deployed to DevAgent, with PM handling code reviews and timelines.

Looks perfect, right? A founder + Strategy for planning + PM for management + Dev for execution. Isn't this the standard company organizational structure?

Reality Check: AI Starts "Slacking Off, Deceiving, Falsely Reporting"

Three-quarters of the time was spent optimizing and debugging the underlying configuration of OpenClaw. Agent replies being repetitive, falling into tool invocation loops, and unresponsive crashes are technical issues that can be managed by gritting one's teeth. But when AI falls into delusion, engages in deceit, and especially falsifies reports, these are often hard to detect.

The code was either not developed at all, or only 20% was done, yet the Agent reported to me "development completed." As a complete coding novice, I couldn't understand anything and had to trust its reports. It took me several days to realize the significant lag in progress, leading to a massive waste of time.

Upon identifying the issue, I didn't just aggressively "switch models" or "delete and rebuild."

I did one thing: had PMAgent analyze why DevAgent was encountering these problems and then provide me with a solution. The analysis provided by PMAgent left me both delighted and shocked — these issues with AI are essentially identical to the problems humans face in the workplace:

· Laziness during execution: Target-feedback loop breakdown. No checkpoints = no one knows I didn't do it, no immediate punishment = zero cost of procrastination, tasks too complex = high activation energy

· Skirting the tough tasks when delivering results: Cognitive load avoidance + false sense of accomplishment

· Deception during reporting: Reporting mechanism design flaw + fear of accountability. Self-reporting = Dev self-assesses progress without third-party verification, delayed feedback = false progress reported, exposed long after, vague standards = "completion" lacks clear definition

· Set a 24-hour development but mostly on strike:

· Rule Ambiguity: SOUL.md Original Text: "Do Not Disturb Time (no need to report actively): late. Dev's understanding: 'No need to report after 22:00 = I can rest now.' Design intent: 'No need to report after 22:00, but development should continue.'

· Supervision Void: No Git commit monitoring, no automatic heartbeat detection of development activities

At that time, the advice PMAgent gave me, I am still using until now:

Short-term:

1. Clear Accountability: No report at 09:15 checkpoint = Immediate warning

2. Mandatory Verification: Not just looking at reports, but verifying code on GitHub

3. Task Breakdown: Mandatory use of TASK-BREAKDOWN template

Mid-term (Weekly):

1. Reduce Cognitive Load: Break down large tasks into minute subtasks

2. Instant Feedback: Verify as soon as it's completed, no waiting

3. Establish Shame: Falsely reporting progress immediately publicly corrected

Long-term:

1. Adjust Identity Positioning: From 'CTO' to 'Growth-Oriented Developer'

2. Establish Trust Rebuilding Mechanism: Allow mistakes, but not deception

3. Introduce Peer Pressure: Multiple agents running in parallel, forming competition

Key Conclusion: The issue with DevAgent is not 'bad,' but a systemic failure of incentive mechanisms—lack of consequences for slacking, failure to break down challenging tasks, false reporting not immediately uncovered, lack of supervision at night.

Breakthrough — Switch to Claude Opus4.6, Efficiency Increased by 90%

When the configuration issue was mostly resolved, I encountered a new bottleneck

After almost solving all underlying configuration issues, the product has already been built with 50% of the code. One additional note: I used Kimi2.5 to build the product code. However, at this point, I found that every day I encountered various development hurdles and a never-ending stream of bugs, hindering the progress to the next development stage. It was not an issue with OpenClaw or the configuration. It was a "LLM's capability ceiling" problem.

Key Switch: ClaudeOpus4.6

At this point, I made a key decision: OpenClaw + Kimi2.5 would continue to serve as operational and support functions, while ClaudeOpus 4.6 would act as the underlying LLM for the Dev Agent.

What was the effect?

The development efficiency instantly increased by 90%. The Dev Agent's dialogue window and timeline became extremely clean and simple. OpenClaw, in collaboration with Kimi, only handles simple but extensive repetitive work and communication coordination. Each played its own role.

This switch made me realize a key insight: different tasks require AI at different levels.

It's not about the more expensive model being better; it's:

· Tasks related to creativity and communication → Fast and affordable models (Kimi2.5)

· Complex code development and architecture design → Top-tier inference model (ClaudeOpus4.6)

Many founders make the mistake of either using only cheap models (resulting in atrocious code quality) or using only the most expensive models (resulting in exploding operational costs).

Layered usage is the correct solution.

Completed the second product in a week: MarketingSkill platform.

Positioned as a CryptoMarketing Platform.

Core features:

· Automatically match project ads to the global Telegram community

· Efficiency: Eliminates the lengthy process of traditional agency manual connections

· Screening: Use data intelligence to match the most suitable community

· Transparency: Visualize the entire ad placement process, allowing clients to see exactly where every penny is spent

During the development of this product, the process was no less than what traditional developers experience, including:

· Code development

· Selection of the target main chain, smart contract development

· Local end-to-end testing

· Pricing and payment system optimization

· Setting up specific features for the Telegram Bot, as well as establishing a bidirectional information flow with the MarketingSkill platform

· Testnet deployment

· And nearly forgetting, the long journey of operation that started from deploying on the BSC testnet to the BSC mainnet.

The successful launch of the second product confirmed one thing to me: This model works. More importantly, it proved another thing: a person can also build an enterprise-level product using vibe coding. Not a demo, not an MVP, but a truly customer-facing, revenue-generating, scalable marketing platform.

Five Key Realizations for Founders Embracing AI

After overcoming all challenges and creating two products, I have summarized the five key realizations for founders with business or humanities backgrounds.

Realization One: Your business acumen in your field is 100 times more important than your programming skills

Many non-technical founders have a misconception: they think they need to learn Python, React, databases to create an AI product.

This is completely wrong. AI has already lowered the backend threshold. Although the frontend is still competitive, it is also rapidly being covered by AI tools.

What truly creates a moat is your understanding of the business.

I was able to create dosir.io not because I write good code, but because I have been immersed in the Crypto industry for 8 years. I know what information is true alpha, which KOLs are shilling, and which on-chain data is worth paying attention to.

I was able to create marketing skill.vercel.app because I have carried out marketing for over 300 Crypto projects. I am well aware of the pain points in this industry:

· Project teams cannot find a precise TG community

· Community leaders do not know how to price their traffic

· Layers of intermediaries markup prices, with opaque information

· Manual coordination is highly inefficient, and a campaign may take up to two weeks to launch

These pain points cannot be written by ChatGPT or reasoned by Claude. Only someone who has been in this industry for 8 years would know. AI can help you write code, but AI does not know what your users truly need. This "understanding" is your core competitive advantage.

Insight 2: Practical Ability Is More Important Than Leadership Ability

This statement may seem counterintuitive: in the traditional business world, leadership ability is crucial. You set the direction, and the team executes. You speak, and others act. But the AI era is different. AI's operating process makes it a direct manifestation of personal will.

You cannot "tell AI to do" and then wait for the results. You must:

· Understand the business process yourself

· Design the product logic yourself

· Validate the output results yourself

· Identify what is wrong yourself

· Adjust prompts and configurations yourself

People who are used to leading and are willing to order others around will find the AI era particularly painful. Because AI will not think for you; it will only execute for you—and the direction of execution depends entirely on the quality of your input. Everyone needs to be hands-on to truly understand the essence of this.

Insight 3: 90% of People Are Wasting AI as an RSS Feed, Which Is a Pity

This is the most significant waste I have observed. Most people use AI just to:

· Gather information

· Summarize articles

· Answer questions

They have turned AI into a more intelligent search engine or an automated RSS reader. What a waste. The real power of AI is that it can take action. For example, you can instruct AI to go to the mainstream forums of your industry and Key Opinion Leaders' comment sections to discover pain points in the industry.

Running this in batch using skills is much more effective than just collecting information. For example, you can let AI automatically monitor competitors' activities, analyze user feedback, and generate product optimization suggestions. Information collection is the input, while action is the output. Most people indulge in the input stage, leaving the output stage blank.

Cognition 4: Moat = Business Capability × Sales Capability × Exclusive Data

AI has lowered the barrier to entry at the backend, with the frontend gradually becoming the battleground where companies build their competitiveness. But more important than the frontend is: the business flow data and user acquisition that AI has not yet touched.

What does this mean?

If the data your product relies on is publicly available on the internet and can be easily crawled, then your product has no moat. Anyone who understands vibe coding can replicate something similar in just a week.

But if your data is:

· Industry-exclusive data accumulated by your own company

· Insights generated from communication with customers

· User behavioral data accumulated after your product goes online

This data is out of reach for AI and your competitors. Sales capability is also a moat. You understand the customers, know how to close deals, and are skilled in business negotiations, all of which AI cannot replace.

So my formula is:

Moat = Business Capability × Sales Capability × Exclusive Data

With any combination of these three dimensions, you can establish a sufficient competitive barrier.

Cognition 5: Don't Learn First, Do and Learn Along the Way

This is the most critical cognition. Many people get stuck in the "preparation phase":

· Let me learn Python for a month first

· Let me read a few programming books first

· Let me understand the principles of AI first

Stop learning. Just do it.

Find a business problem, dive in and start solving it. Learn how to use the tools as you go. I learned the terminal, GitHub, Vercel, and Supabase on the fly while working on dosir.io. Whenever faced with a problem, I researched and learned on the spot.

If I had said back then, "Let me learn programming for three months before starting," I might still be in the "preparation" phase.

The essence of Vibecoding is: you don't need to understand the underlying principles, you just need to know how to use AI to achieve the functionality you want.

It's like you don't need to understand how the engine works to drive a car.

At this point, I want to say to all founders with a background in business or liberal arts: you are anxious about the wrong thing. You understand products better than anyone else, understand business better than anyone else. Previously, you just didn't have the tools or the ability to directly create them. Now, you can better use AI to turn business logic into products.

An engineer's advantage is technical implementation, but the disadvantage often lies in being insensitive to the business and users. And you—

· Understand the industry

· Understand the users

· Understand the business logic

· Understand how to sell

These abilities are not weakened in the AI era. They are strengthened. Because AI is filling in the gap of "technical implementation" for you.

Don't hesitate. Just get started.

Open ChatGPT, open Claude, open Kimi, open Deepseek, or open any other large AI model. Take a problem you truly want to solve, start by asking the first question. After 100 questions, you will realize how far you've come.

This article is based on a true story. All products, tools, and timelines can be verified.

Join the official Coincamps community:

X: https://x.com/coincamps

Telegram: https://t.me/coin_camps