My AI Side Hustle, Half a Year In

Today is November 12.

From today through the 14th, Fanqie Novel (Tomato Novel) is adjusting the levels of every author on the platform. Once that's done, for most people, the road of writing novels with AI will have reached its end.

It took me three months to figure out how to write a good AI novel. From figuring it out to today is only a little over two months.

This piece is about those six months.

Getting In

At the start of this year I began using AI constantly, mostly agent products. The newest models then were GPT-5.4 and Claude Opus 4.6. I built a blog with AI, cleaned junk off my computer, wrote articles, and then finally started wondering how to make money with it.

To this day I still think there's one rule at the core of making money with AI: bring AI's capabilities into a field you could already work in yourself.

In early May I decided to go deep on AI-written novels.

The reason was simple. Writing is what I'm good at, and it's what AI is good at. Text is still AI's main interface: you send a model a paragraph, the model and its agents do a series of things behind the scenes, and a paragraph comes back. So text work is the easiest place to start with AI. As for novels, I wrote them for four or five years back in school. My results were nothing to brag about, but I know the industry and the process of writing web fiction inside out.

So I started.

One month earlier, in April, Qidian had published an "Announcement on Further Strengthening the Management of Non-Human Automated Creation."

Paying Tuition

Once I got into it, I found that plenty of people were already doing this. Of course they were. If I could think of it, people who found AI before me could think of it too.

On Bilibili I found a solo creator. His videos were simple: once a month he showed his royalties, which started at 40,000 yuan a month back then. Along with the royalty screenshots he sold a course, really a community, with his tutorials inside, for 299 yuan.

I hesitated for a while, then paid.

As I was about to open the tutorial, I felt a side hustle worth 10,000 yuan a month waving at me, and I was thrilled.

My experience with AI was still thin, and the tutorial's steps were too much for me. I followed the SOP to the letter and started two books. Both fell apart, and both went dark before chapter 100.

The Detour

I stopped and took the two books apart.

The first one failed on genre. I ignored the group owner's instruction to write women's web fiction and started a men's book instead. During its debut every number was zero, so I cut it immediately.

For the second I switched to women's web fiction. But I'd never read any, so I wrote fan fiction for an IP I knew well. That book actually got decent traffic, but around chapter 100 the storyline began to collapse.

There were two problems I simply couldn't control.

First, the main plot moved too fast. AI told a volume's worth of story in half a volume's worth of chapters, and then had nothing left, so it dragged the second volume forward. One volume's length ended up covering two volumes' worth. It couldn't pace itself, and the writing wasn't good either.

Second, it broke its own continuity. AI's context window is finite, and a web novel runs to hundreds of thousands of characters, full of big and small tasks, with butterfly effects between events that change how characters feel and act. AI not only forgets a character's state, it invents things: the protagonist suddenly has a sister, a character who died comes back to life. Smaller things too: a place name wrong, a surname wrong.

To fix this, I decided to take the tutorial I'd bought and rebuild it into something more complete.

My idea was to have AI write me a piece of software that managed a book's progress, with several spreadsheets tracking characters' relationships, states, and possessions. Writing one chapter meant checking a great many things.

I named the project the Book of Truth.

On May 4 I first mentioned it to Liang Mouyin.

From that day until August I didn't start a single book.

I went down a rabbit hole with no end, spending every day figuring out how to make that software better. The only thing that came out of it was this: I sent the beta to Liang Mouyin, he used it to start a book, and he earned 300 yuan in royalties.

Later I fell into a bad spiral. I was addicted to reinventing the wheel and never satisfied with the wheel I built. By late August, the Book of Truth, by then renamed Seedraft, had more or less been dropped.

I spent a fortune on AI subscriptions, burned through over 20 billion tokens, and put in three months of time and energy. In the end I had nothing to show for it.

That's how it goes. Effort alone doesn't produce income. Sometimes you have to stop and think, and look around.

I call those three months my detour. I was set on building my own agent, and I badly underestimated how hard it is to build one well.

That August, Qidian also began taking works down.

Making Money

I stopped building software and went back to basics. I took the prompts I'd accumulated over those three months and put them into agents that already existed.

The AI world was a different place from May.

Claude released Fable 5.1 and Opus 4.8. On September 3, GPT released 6Astra. Behind every version bump was a leap in what models could do. Things that took real effort three months earlier could now be done in one sentence. Codex, Claude Code, Antigravity, all had huge upgrades. Domestically, DeepSeek Harness was about to launch, and Proma, built by Chinese developers, was strong too.

I put my refined prompt system into those big-company agents and let it run.

Suddenly everything was clear.

The problem I'd spent three months trying to solve with software, the new models had solved on their own. Three months of work turned into nothing in an instant.

Because I leaned heavily on Gemini, I ended up choosing Antigravity. It wasn't as mature or as easy to use as the other two, but it could call Gemini directly.

The first book I started with it hit 10,000 reads two days into its trial run.

I began to think that maybe I didn't need to build my own software to make this work.

After that I kept posting my numbers and my technical notes in related communities, and slowly gathered people who wanted to make money from AI novels. They were willing to pay for my system.

Less than half a month after I gave up on the software, I did what three months hadn't: I had a paid AI novel community of my own.

The Industry Shifts

Let's put the numbers on the table.

So far I've received three royalty payments: 98 yuan, 762 yuan, 643 yuan. That's 1,503 yuan in total.

Community membership is 298 yuan. Eighteen people joined, 5,364 yuan.

Together, 6,867 yuan, most of it tuition from students.

In theory, with my feel for fiction, this should have been a business I could run for a long time.

But just as I finally found the technical optimum after three months, the industry started to change.

The AI novel business, put plainly, depends on one platform: Fanqie Novel. Fanqie is famous in the field for low quality. Even before AI, the platform was full of low-quality content, and plenty of hand-written novels couldn't even get a contract. AI-written stuff gets signed without a second look, never mind how much it ends up earning.

Qidian moved first.

On August 18, Qidian publicly took down a large number of works, including Shanhe Jian (Mirror of Mountains and Rivers), which averaged over 30,000 paid subscriptions, and several titles near the top of the monthly ticket rankings. April's announcement was policy. August's removals were action. The biggest name in paid Chinese web fiction had started to crack down on AI's impact on the industry.

On September 7, Fanqie added a "completed-book contract" model and published its "Explanation of Work Contract Standards," which for the first time listed "AI slop with messy formatting, broken structure, and empty padding" as a separate category ineligible for a contract.

At that point I thought it was fine. Even if Fanqie later moved almost entirely to human review, most of the novels my system produced could still land contracts.

October 12.

A devastating day for the AI novel industry. Fanqie announced two things.

First, Author Level 3.0 went live. Eight levels, advanced by accumulating points through publishing, signing contracts, and the like. Levels will be adjusted across the board November 12 to 14.

Second, new works by Lv. 0 and Lv. 1 authors can only use the completed-book contract: write the full 150,000 or 300,000 characters the category requires, finish the whole book, then apply. The whole book is reviewed at once, and each book can be submitted at most three times. Readers can't see the book before it's signed. The old serial contract let you apply at 20,000 characters, with chapters reviewed one by one. The trick of "write the opening yourself for a contract recommendation, then let AI mass-produce the rest" is closed.

Lv. 2 and above can choose between the serial and completed-book models; books already started are unaffected.

Why do these two changes hit AI novels so hard?

Start with the contract model. The biggest advantage of AI novels used to be how fast they paid out. Sign at 20,000 characters, debut at 80,000, and see the numbers the next day. Bad numbers, drop it and start another. That's how I cut my first book.

It was a gacha system. One pull cost only tens of thousands of characters, and AI writers lived by pulling over and over.

Now you have to finish all 150,000 or 300,000 characters before you're even eligible to apply. Even for AI, writing a book from start to finish without it collapsing takes a long time. Readers can't see the book before it's signed, so there's no data to go on. It's writing the whole book blind and only then learning whether it works. And you get three chances.

Review also went from single chapters to the whole book. The place AI novels fail most is the middle and late stretches: the main plot drifts, the continuity breaks. Before, by the time those problems showed up, the book was long signed. Now they're all laid out in front of the reviewer.

Then the author levels. To get around the completed-book contract you have to reach Lv. 2. Right now Lv. 2 takes 3,000 yuan in royalties from one book. Most AI writers, and plenty of ordinary human writers too, can't reach that line.

The cost of one pull went from tens of thousands of characters to a hundred thousand, several hundred thousand. Gacha has been banned outright.

Looking Back

These six months I got some things right and some things wrong. But looking back, the things that really decided the outcome were never in my hands: when models upgrade, when platforms change their rules.

The only thing I can decide is when I step in.

Getting in early matters more than anything.

When I came across that creator, he was already making at least 40,000 a month. Was he much smarter than me? I don't think so. He was just earlier.

I decided to do AI novels in early May, and from May to August I didn't start a single book. By the time I actually got it working in September, Fanqie had changed its contract standards on September 7 and sealed off serial contracts for newcomers on October 12.

If I hadn't gone down the software rabbit hole for those three months, if I'd just started one book after another, even with lousy tools, I'd at least have a few more signed books, and three more months of royalties and experience.

There's a line in the new rules I read over and over: Lv. 2 and above can choose for themselves, and books already started are unaffected. Whenever rules tighten, it's old rules for the old guard, new rules for the newcomers. The people who got in early made a few more months of money and picked up an exemption ticket along the way.

Getting in means actually stepping onto the field: starting a book, publishing, watching the numbers. For three months I was busy with AI novels every day, and the whole time I was standing outside the field. Liang Mouyin took the beta I myself couldn't stand, started a book, and earned 300 yuan. That was the project's only income in three months, and I didn't even earn it.

Combine AI with the hobby you've put the most into.

For the second book I switched to women's web fiction, but I'd never read any, so I went back to a familiar IP and wrote fan fiction, and that's when the traffic came. The farther you get from what you know, the harder it is to keep writing.

After the book fell apart I could see where it broke: the plot ran too fast, the continuity broke. That's a judgment only someone who's written for four or five years has. Someone who hasn't written will just think AI writes badly, can't say where, and certainly can't fix it. Later, my prompt system, the thing people in the community paid for, came out of that judgment.

AI can do the work for you. You still have to judge whether it's any good.

I wrote novels for four or five years in school and never did well. Before AI, that experience was basically wasted. After AI, it became the most valuable thing I own.

Model capability decides product quality.

The two problems I spent three months building software to solve were, plainly, the model's shortcomings at the time: the context wasn't long enough, and it couldn't remember things. What I was doing was patching the model with spreadsheets and checklists.

The problem I was fighting so hard to solve in May was answered by the August model.

Model companies are more anxious about the model's shortcomings than you are, and they have far more people and money. If you go up against it alone, odds are the next generation arrives before you're done.

Those three months weren't all for nothing. What got scrapped was the software, the spreadsheets and workflows. What's left is the prompts and my understanding of why a novel falls apart. Things tied to a model's shortcomings expire. Things tied to the industry don't.

Making money depends on AI. How long you can make it depends on the industry.

The technology didn't fail. The industry changed.

The whole business stood on Fanqie's leniency. The foundation belongs to someone else. They change one rule and the building shakes.

And the platform tightened up precisely because AI was too easy to use. The barrier was so low anyone could walk in, content flooded, the flood hurt the platform, and the platform acted. Qidian announced in April and removed works in August; Fanqie changed contract standards in September and author levels in October, each step tighter than the last. The speed at which AI makes something easy and the speed at which the industry starts guarding against AI are the same speed.

If I size up an AI side hustle now, I ask two questions. Can AI do it? Will the industry let AI do it, and for how long? The person who rushed in after answering only the first question was me in September.

One more thing I have to be honest about: most of my income is tuition; royalties are the small part. Teaching pays better than doing, and that's common in AI side hustles. But tuition is rooted in the industry too. When Fanqie's door closes, fewer people will want to learn.

Price with confidence.

The 299 yuan I paid bought a tutorial that, followed by the SOP, produced two books. Both collapsed. Looking back, most of it was information asymmetry, and the packaging wasn't exactly polished.

I don't regret it. It showed me that someone had made it work, and roughly how. The information gap itself was worth money.

My community was built by constantly posting numbers and notes in the group. I priced it at 298 yuan, one yuan less than the fee I paid.

I didn't set that price. My first customer did. I think that's good, letting the users price it for me. It's worth whatever the people willing to pay say it's worth.

What they buy is the road I already walked. What they save is their own three months.

If your product came out of work you actually did yourself, it's already more solid than plenty of paid products on the market, and you don't have to feel sheepish about the price.

November 12

Today the level adjustment began.

My judgment is that the old model of AI novels no longer works.

My best-performing book makes 500 to 600 yuan a month so far. At that rate it takes nearly half a year to save 3,000 yuan and reach Lv. 2.

Actually, five or six hundred a month from one book is already decent. It covers a month of AI subscriptions, or some odds and ends.

Going forward, I think the future belongs to polished work. And with AI's current ability, it's hard to write a high-level novel with real character.

AI's strength in fiction is that it can tell a story very clearly. With a solid outline it writes smoothly, clear from beginning to end, with no dragging.

That's its strength and its weakness. What Fanqie calls a "stiff style" in its contract standards is exactly this.

I often say AI won't make dumplings for a dish of vinegar. It won't wrap a whole batch just to have something to dip. When it tells a story, it goes by the book.

People are different. People will often go ahead and make the dumplings. Shao Song spends an entire book wrapping dumplings, all for that last poem, Linjiang Xian.