Last week I had dinner with a friend who runs Xiaohongshu accounts. She said she wanted to build a tool that auto-writes viral titles and charge for it, then asked: "Do I need to learn Python first?" I almost spit out my Coke.
Two years ago I thought the same thing. I wanted to build a small tool helping freelancers manage their quote sheets — spent three months just "finding a tech cofounder." They ghosted. The product died.
Now it's different: you're not building the engine, you're building the steering wheel
Recently saw a concept called "harness" — the idea: AI models are like car engines. Smart, they run, but a bare engine can't hit the road. What you build is the steering wheel, seats, dashboard — the "stuff users can touch."
Abroad there's a company called Harvey. They make "AI assistants for lawyers." Lawyers don't need to understand AI — they just open the interface, drop in a contract, and ask "is this clause fair to my client?" Harvey's core isn't training their own model. It's stitching together a general model + legal-domain data + lawyer workflows. Last year they raised $150M on this model.
In China you can see the same play: all those "Xiaohongshu viral title generators," "WeChat article formatting AI," "e-commerce customer service scripts" — same playbook. Anyone can plug in a model, but "which scenario, for whom, in what tone" — that's your job.
To replicate this today, what do you need?
Money: Take a cheap model like GPT-4o mini. 1,000 normal Q&As run about $1-2 (around 7 RMB). A few thousand calls a month costs less than a bubble tea.
Time: With drag-and-drop platforms (ByteDance Coze, Alibaba Dify), and not building complex features, you can ship a minimum version in one or two evenings.
Tech barrier: No code. I built an "auto-categorize customer inquiries" tool in Coze myself — all mouse clicks.
First step: Go to coze.cn, register with your phone, click "Create Bot," pick a scenario you know cold (help students revise resumes, help clients write weekly reports — anything), follow the three-step prompts, and you're live.
Three types, three plays
Just starting: You probably don't even know "what to sell" yet. My advice: don't build a tool first. Post 10 pieces of content on Xiaohongshu or WeChat Moments with the format "I used AI to do X, saved Y time." See what gets the most traction — then decide if it's worth a product.
Have 1-2 clients: You already know their specific pain. Build a minimum version in Coze for them to use. Charge even 99 RMB/month. The point isn't profit — it's validating "will they pay for this workflow."
Scaling up: Starting to worry "what if the model changes," "how do I manage user data." Time to build your moat — accumulate client history, maintain your own prompt library. The model vendor can't take these from you.
Final word: This playbook isn't for everyone. If your current business doesn't really need "automating some workflow," forcing it just wears you out. First ask: among the annoying repetitive things you do daily, is there one worth "outsourcing to an AI shell"?