王建硕：Markdown，是新时代的编程语言

Jian Shuo Wang: Markdown, the Programming Language of the New Era

Jian Shuo Wang started by telling me: "Xu Wenhao talked about how to manage 'shit code' in the last episode—I don't quite agree with that."

His rebuttal is: shit code is unmaintainable code, and there's no need to maintain it. "Code doesn't have errors; it just doesn't align with your intention—what we call a bug is either a bug itself or a feature written incorrectly." All methods, skills, and best practices for managing shit code—Jian Shuo Wang agrees with them all—but he believes this approach is misguided: you shouldn't be modifying the source code layer; you should be modifying the natural language layer above it.

After making this point, he talked to me for an hour about "what a true programmer looks like in the AI era."

The most insane part was when he showed me his app on his phone—

The most important feature is called "User Feedback". Users can write something like "Make the font bigger" or "Add an XX feature," and the app updates according to the feedback within 5 minutes. The AI receives GitHub Issues, automatically pulls code, writes code, builds, and pushes to TestFlight.

I thought this was already absurd enough. A few days later, he told me:

"I don't need user feedback anymore. It decides what to update and updates itself."

He connected a PM Agent to that app. The AI determines what the next version should do and adds the features itself. Jian Shuo Wang has transformed from a developer into a curious user who checks what his app has become each day— "I'm curious to see what it looks like today."

This episode is for all product managers, programmers, and entrepreneurs. If you're still "chatting" with AI today—please listen carefully to this episode.

• * *

[Guest]

Jian Shuo Wang | Internet veteran, founder of BaiShiNet. In the past one or two months, he has once again "taken up programming"—but this time he's not writing C or Python; he's writing a bunch of .md files to let large models run on schedule. He runs four agent teams (PM / Engineering / Marketing / Customer Service) by himself, where the PM Agent decides what the next version should do without his direct command. His app updates every five minutes, reviewed and launched daily.

[Highlights of This Episode]

"Xu Wenhao talks about managing 'shit code,' but I don't quite agree"—starting with a rhetorical challenge

Jian Shuo Wang immediately challenged the previous guest, Xu Wenhao. Xu Wenhao previously advocated using skills, best practices, and various management tools to make AI-written code maintainable. Jian Shuo Wang's rebuttal is—shit code is unmanageable code, and there's no need to maintain it. His point is roughly that as long as shit code isn't seen by people, it's not shit code. He acknowledges that Xu Wenhao's management methods are correct, but he believes this approach is misguided: you shouldn't manage at the source code level; you should manage at the natural language level above it.

"Code doesn't have errors; it just doesn't align with your intention"

Jian Shuo Wang redefines what a bug is: so-called errors are either bugs themselves or your features being written incorrectly—essentially, your intentions (intention) aren't clear enough. He gives an example: in C language, if you say print color without specifying the color, the compiler prints white—is that a bug? No, it's because you didn't specify blue. Most of the time when people say AI-written code is incorrect, it's actually due to unclear intentions, not AI. That's why he disagrees with "managing source code"—he believes the wrong object is being managed.

"Users write feedback, and the app updates according to their requests within 5 minutes"

This is the most insane part of the episode. Jian Shuo Wang's app has a feature called "User Feedback." Users can write anything—"make the font bigger," "add search functionality"—and within 5 minutes, the app generates a new version based on this feedback. The technical process is quite straightforward: user feedback directly goes to GitHub Issues, triggering a hook, Claude Code automatically pulls code, writes code, runs tests, merges into the main branch, builds, and pushes to TestFlight. "From submitting an issue to testing it on your phone, the app updates every five minutes."

"I don't need user feedback anymore—it decides what to update and updates itself"

Even more shocking is the next layer. Jian Shuo Wang connected a PM Agent to the app—when no feedback is provided, the AI determines what the next version should do and adds the features itself. "Every one or two hours, TestFlight will notify me of a new version release. I'm always curious to see what it looks like today." He has transformed from a developer into a curious user who checks what his app has become each day.

"I wrote tens of thousands of lines of natural language to get this system running"

Is this hands-off management? No. Jian Shuo Wang repeatedly emphasizes: "The natural language I mentioned is truly handwritten, not generated by a large language model. Writing natural language line by line, you might write at least hundreds of lines." His working directory is layered as Model / View / Control, with subfolders containing tens of thousands of .md files. The granularity of his handwritten natural language is the granularity of his control over the project. "Your dissatisfaction is with its default values—but those default values aren't your content."

"Chatting isn't programming; chatting is debugging"

Jian Shuo Wang made a statement that kept me thinking: programming is what you write, is certain, can be executed periodically or triggered, and gives you results. The opposite of chatting is writing programs. "Chatting is a very fast debugging method, but it's not how you write programs." He estimates that now 90% of people using AI are still at the "opening Python command line to learn the language" stage—no accumulation, no沉淀, and they start over the next day.

"Natural language is the new 'assembly,' and Python is becoming the new 'shit code'"

Jian Shuo Wang drew an industry evolution diagram: machine code, assembly, C, Python, natural language—each layer up makes the layer below "shit code"—unreadable and unmaintainable, but it doesn't affect human life. LLM is the new compiler, and natural language is the new 'assembly'. This diagram is also the theoretical foundation for why he disagrees with Xu Wenhao's theory of managing "shit code"—you're not fixing the foundation; you're fixing the layer above it; let the foundation sink.

"Programmers are the porters of the previous generation"

The most resonant metaphor throughout the episode. "We thought carrying the sedan chair was real work; suddenly it became a car, and we thought we had control over the car too—but we've already left." Programmers are now trapped in the "godfather" position—their coding abilities can be completely replaced by AI, but due to emotional bias, their willingness to learn driving is much lower than an ordinary person. "Workers in industries closest to new technology suffer the most."

"How many Tokens you burn daily is the GDP of the AI era"

Jian Shuo Wang's only hard metric is Token usage multiplied by unit price. Regardless of how sensible your PPT is or how advanced your AI cognition claims to be—Organization A and Organization B look at Token usage to determine the winner. He burns around $2,000 per day. "Clearing out Token waste is like treating a pus-filled sore—you won't see the real problem until you clear it." Recently, his poster reads: Embrace Waste by Using Tokens.

[Timestamps]

01:40 Jian Shuo Wang challenges Xu Wenhao: shit code doesn't need management

02:24 True definition of shit code: code that can't be maintained by people

03:18 Shit code isn't shit code if no one sees it

06:00 Redefining "shit code": code that can't be maintained by people

06:51 Code doesn't have errors; it just doesn't align with intention

07:06 Jian Shuo Wang acknowledges Xu Wenhao's methods are right, but the layer is wrong

12:12 Chatting isn't programming; it's debugging

14:43 What a true "programmer" looks like: working directories, folder engineering

21:47 Porters, godfathers, and drivers

33:08 Jian Shuo Wang's new identity—one person running four agent teams

37:18 Users write feedback, and the app automatically updates (core segment)

38:17 "I don't need user feedback anymore—it decides what to update and updates itself"

38:30 Breaking down the entire process (feedback, GitHub Issue, Claude Code, TestFlight)

39:38 PM Agent: letting AI make decisions instead of user feedback

45:30 "I don't download apps from the App Store anymore" + throwaway software

50:25 Everyone in the 21st century needs to know how to use computers—everyone will use AI in 10 years

1:02:01 The only hard metric: how many Tokens you burn daily

1:04:20 Embrace Waste: Use Tokens to Embrace Waste

Subscribe to the podcast "AI Alchemy," as well as the same-named WeChat Official Account and Video Account

"AI Alchemy" is a podcast co-hosted by Xu Wenhao and Ren Xin—two old friends and seasoned professionals in the AI field. Here, we discuss AI and entrepreneurship, inviting leading entrepreneurs, product managers, and researchers to delve into how AI reshapes industries and changes lives, and how to create AI-native products from scratch.

Our discussions cover multiple topics: from how AI will change the future of the world to finding the Product-Market Fit (PMF) for AI startups; from using AI to reduce costs and increase efficiency to integrating AI technology into daily life... If you're interested in AI, products, and entrepreneurship, there's plenty of valuable content and first-hand experience here. Feel free to subscribe and recommend it to your friends to explore endless possibilities together!

Business Cooperation: Contact information can be obtained via the [Business] menu in the "AI Alchemy" WeChat Official Account.

Podcast Hosts:

Xu Wenhao: Co-founder of an AI startup company, developing AI applications for the global market. Serial entrepreneur, involved in multiple startups, early employee of Pinduoduo. Algorithm and data director of advertising technology company MediaV, later acquired by Qihoo 360. After leaving, he joined Pinduoduo, which was less than a month old at the time. Later, he founded AI-based overseas customer service chatbot company BotHub.AI and overseas social e-commerce platform Bukito, both of which failed. Returned to entrepreneurship in 2023.

Ren Xin: Partner at a US dollar fund, primarily investing in and incubating AI applications for the global market. Previously a serial entrepreneur, he created "Tonight Hotel Discounts" to sell hotel surplus rooms at discounted prices via mobile internet applications; after being acquired by JD, he internally started "JD To Home" for local instant retail business; in 2015, he started Get to explore conversational AI assistants, which was unsuccessful; sold the company in 2021 and returned to the AI world in 2023.

Podcast Editing:

Sixteen Sugars

BGM:

Opening: Shortwire - Reconfig

Ending: Sam Wills, Honey Mooncie - Traingazing