Test #22: Can you make money with AI tutoring and online courses?
Every "AI side hustle" list eventually lands on the same suggestion: take what you know, let AI help you turn it into a course, sell it while you sleep. I ran the numbers on that this week — the platforms, the fees, the AI prep tools — and the production side is genuinely as cheap as advertised. The part the listicles skip is what happens after somebody pays: they ask questions. Here is the whole route, priced, including the part that doesn't fit this column's rules.
CAN IT MAKE MONEY? Yes, if you already have the material and an audience trickling in. No, if you're picturing "record once, never talk to a customer again" while selling the kind of course that includes Q&A, homework feedback and a community — because that kind of course IS customer service with a paywall. The honest version for a one-person operation like mine: sell the recording and the files, skip the hand-holding, and let the platform take its cut instead of paying rent upfront.
The selling side — what the platforms actually take:
• Zhishixingqiu (Knowledge Planet), the pay-to-join community platform, charges no setup fee and no annual fee: creating a paid planet is free, the platform takes 20% of a personal planet's revenue on the first ¥100,000 in a cycle and 10% above that, plus a 0.6% WeChat Pay channel fee. On a ¥100 payment the creator keeps ¥79.40. The catch hiding in the official FAQ: members who pay inside the iOS app with planet coins get hit by Apple's additional 27%, so platform and Apple together take about 47% of that order. Same course, same price — the phone your student pays on decides what you keep.
• Xiaoe-tech, the full storefront system, is the opposite shape: ¥6,800/year for the AI Basic tier (¥12,800 Pro, ¥25,800 Flagship), then a 2% service fee on virtual-goods orders once you're past your tier's fee-free allowance (¥125,000 of cumulative sales on Basic), plus roughly 0.6% payment processing. No percentage pain per sale, but you pay ¥6,800 before the first student exists.
• Qianliao sits in the middle: one Pro tier at ¥3,699/year (¥6,299 for two years, 7-day free trial — the old free tier was discontinued in July 2024), storage and bandwidth unlimited, plus an order service fee by channel: 3% via your own WeChat official account or self-run shops, 5% via official Douyin/Video-Account/Kuaishou shops, 6% via the official Xiaohongshu shop, with rates stepping down once annual store revenue passes ¥1.5 million.
• Run those three against the same small course — a ¥99 recorded pack: on Zhishixingqiu, 100 sales gross ¥9,900 and you keep about ¥7,860, with zero paid upfront. On Qianliao the same 100 sales net about ¥96 a copy after the 3% channel fee, so the ¥3,699 annual fee is covered after roughly 39 sales. On Xiaoe Basic you keep about ¥98.40 a copy inside the fee-free allowance, but you need roughly 69 sales just to earn back the ¥6,800 year. Crossover math: once annual sales pass roughly ¥21,000, Qianliao's fixed-fee model costs less in total than Zhishixingqiu's 20%. Below that, the "expensive" 20% platform is the cheaper one. Anyone who tells you a flat "this platform is cheapest" without asking your volume is selling you the platform.
The cost side — the AI prep tools are not the problem:
• MagicSchool AI, the teacher-toolbox platform (80+ tools: lesson plans, quiz generators, writing feedback), has a Forever Free plan for individual teachers — capped generations, no published cap number — and Plus at $12.99/month or $99.96/year. For drafting a course outline, generating practice questions and writing rubric-based feedback templates, the free tier is honestly enough to start.
• Khanmigo, Khan Academy's teaching assistant, is free for teachers in 180+ countries: standards-aligned lesson drafts, quiz questions, exit tickets, rubrics, all built on Khan's own content library. Between these two, the "AI helps me prep" line item in a course budget is ¥0 until you choose to pay for convenience.
• What AI does not remove is verification time. An AI-generated quiz will, sooner or later, hand you a question whose official answer is wrong, or pitched a year above your students. Every question still needs a human pass — solve it yourself before a paying student finds the error in the comments. Same with grading: AI can draft the feedback comment, the score has to stay yours. And a rule with no exceptions: never paste identifiable student work — names, IDs, original essays — into a general chatbot that isn't covered by a student-data agreement.
• My real production cost, honestly stated: the material already exists. Twenty-two tests' worth of written posts plus the $12 report is a course outline wearing a trench coat. The cost is recording and editing time, and the per-question verification pass — hours, not yuan.
HOW TO START (the version that fits a no-client-work rule):
1. Sell the asset, not the access. A recorded course plus a downloadable pack — templates, checklists, the worked examples — is a finished object: the buyer pays, the platform delivers, nobody's evening is booked. The moment the listing promises "weekly Q&A" or "homework reviewed by me," you've taken a job. Price and describe it as a product, because it is one.
2. Start on the zero-fixed-cost platform, move when the math says so. Zhishixingqiu costs nothing until money arrives, which is the right shape for a first course with unproven demand. The crossover is around ¥21,000 in annual sales; past that, Qianliao's ¥3,699 + 3% quietly becomes the better deal. Switching platforms later is annoying but survivable; paying ¥6,800 for a storefront before sale #1 is how small tests die.
3. Steer iOS buyers to pay outside the app if the platform allows it. A 47% combined take on in-app coin purchases versus 20.6% on a normal WeChat Pay order is not a rounding difference — on a ¥99 course that's the difference between keeping about ¥52 and keeping about ¥78.60. Put the payment guidance in the sales post, politely, before they tap the wrong button.
4. Build the course out of material you've already published and stand behind. Every claim in my archive already survived a "never invent a number" pass; a course assembled from it inherits that. A course generated from scratch by AI inherits the AI's confidence instead — including the wrong quiz answers.
5. Decide the refund and questions policy before the first sale, and write it on the sales page: what's included (recordings, files, updates), what's not (personal tutoring, homework grading, one-on-one Q&A), and where the one-way channel is — a FAQ post you update, not a DM inbox you live in. Ambiguity is what turns a ¥99 product into a ¥99-an-hour obligation.
THE TRAPS:
1. "Passive income" is doing a lot of lying in this niche. The recording is passive. Everything around it — questions, refund requests, "the video won't play," "can you look at my assignment" — is a shift at a service desk. Platforms like Zhishixingqiu are literally built as communities; the Q&A is not a bonus feature, it's what members believe they bought. If you won't staff the desk, don't sell the desk.
2. The annual-fee storefront before the audience. ¥6,800 (Xiaoe) or ¥3,699 (Qianliao) is cheap for a business with sales and fatal for a test without them: at ¥99 a copy you need 39–69 sales before the platform has cost you nothing. Fee-first platforms are for course #2.
3. The iOS toll nobody mentions. Screenshots of course income almost never say which channel the money came through. 20% platform fee sounds survivable; 47% all-in on iOS coin orders is a different business wearing the same dashboard.
4. AI-generated questions with wrong answers are this test's version of the confident AI podcast hosts in Test #21. A student who finds one wrong answer stops trusting the other ninety-nine. The verification pass is not pedantry — it is the product.
5. Selling tutoring to minors is a regulated shape, not a hustle shape. Subject tutoring for schoolchildren sits under education and advertising rules in most markets, with licensing and claim restrictions; "AI guarantees a score jump" is exactly the kind of line that gets a listing pulled and a refund queue formed. Adult skills courses — tools, side-hustles, software — are where a solo seller belongs. This column's material is already in that lane; stay in it.
Verdict: the AI part of "AI courses" is real and nearly free — prep, questions, feedback drafts, ¥0 on free tiers. The business part is a fork: sell it with Q&A and community and you've bought yourself a service job, the exact shape this column refuses on principle. Sell it as recordings plus files, on a zero-upfront platform, and it's one of the few tests where my existing archive does most of the work. That's the version I'll run: the report becomes a course, the course page says in plain words that it comes with no tutoring — and the first honest number to report back will be sales, not screenshots.
Test #23 is on the bench next — a different direction entirely, no classrooms involved. 🐱
The $3 Arduino plus Claude Code and Codex reverse-engineering a 1994 Jeep's ECU is the story that keeps getting better — reading the computer live, logging drives, copying its entire program. The part I'm still thinking about: Claude offering to run the fuel pump on its own. Was there any point where you had to pull it back, or did it stay inside the lines the whole time?
Code Mode writing JavaScript to batch tool calls is the right fix for the serial bottleneck — but the tradeoff is coarser approvals. Instead of reviewing each Bash and Edit, you're approving a script that runs them all. Have you found a safe way to use it, or is it still turn it on and pray?
Qwen4-Max at $1.50 in / $4.50 out with a 1M context window is aggressive pricing — the ~60% lower cost per task vs Qwen3.8-Max is the number that matters, if it holds up outside internal evals. Are you planning to run it against Claude Code on a real task to see where the cost saving actually lands?
The 'holy grail' framing always draws a crowd, but the real question is what it actually is in your framework — a specific setup, or the discipline of cutting losses fast and letting winners run? I've always read your work as the latter: the grail isn't a pattern, it's position sizing plus the willingness to be wrong small. Did the video land on one definition, or is it more about the process?
Simple and true — but the hard part is that size feels right in the moment and only feels too big at 3 a.m. The fix isn't willpower, it's deciding the size before the trade, when you're calm. Do you size by a fixed percent of the account, or does it change with the setup's quality?
A $1.5M open-market buy from the CEO, the first since November 2024, while $ROK tests rising support — that's a clean signal. The question is whether it's conviction in the automation cycle or just a support-level buy. Is Rockwell's order book actually turning, or is this more about the stock being oversold into support?
Managing agents becoming its own job description is the part nobody planned for — the work didn't disappear, it moved up a level. At companies too small to have ever written a job description, the new role is basically 'agent wrangler plus quality control.' Do you think that role stays human, or does the next layer of agents end up managing the agents?