Recently, I helped quite a few friends install OpenClaw (nicknamed "little crawfish" in Chinese tech circles), with Claude Code automating the entire process so I barely had to lift a finger.
It reminded me of that Tencent video where a bunch of engineers sat there installing it manually for people one by one, getting through hundreds in an afternoon. With Claude Code paired with a strong model, it takes about ten-odd minutes per install and costs roughly ten to twenty yuan—maybe just a few yuan in the future. Remote installation services on the market charge 300+ yuan per unit. Given that every computer environment is different and networks are unpredictable, half a day of manual labor really can't get much cheaper. But AI has driven that cost down to one-tenth.
No Method, Just Volume
After installation, many friends ask: How do you use this thing? How do you actually learn AI?
My answer has always been simple: There's no secret method—just use it a lot.
I bought the Claude Code Max × 20 membership and basically hit the weekly limit every time. One person's usage is comparable to a small team. Use it enough and you'll naturally know what it can and can't do, then you'll start getting used to letting it handle more and more things. I used it to build this website from scratch, and I used it to write 300,000 lines of code in 10 days and then delete them all—that process itself is the learning.
The shift happens at the mindset level. When you run into a problem, your first reaction becomes: Can I get AI to help me with this?
People who haven't reached this stage can read all the articles they want, but when it's time to actually work, they'll still fall back to old habits. They can't resist micromanaging and constantly want to supervise. AI can't perform under that kind of usage.
Worker and Manager
Many people think this wave of hype is like the earlier DeepSeek craze—it'll fade away. I don't think so.
But I do believe most people shouldn't start with OpenClaw right now; they should start with Claude Code.
OpenClaw is positioned more like a Manager, responsible for delegation and coordination. Coding agents like Claude Code are the Worker, responsible for execution.
Hook up a good model to Claude Code and you'll feel what execution power really means—give it a vague, complex task, like remotely installing OpenClaw for someone, and it'll keep at it for two or three hours, troubleshooting on its own until it's done. Seeing this capability makes you realize you no longer need to write complex prompts or break down tasks; just give it a general direction and it'll deliver.
Only when the Worker reaches this level does the Manager become valuable. It can direct multiple Workers simultaneously, each task needing only a simple instruction, then you just wait two or three hours for the results.
The Chain Reaction of a Dumb Worker
Think about it the other way around. What if the Worker is dumb?
Give it a vague task, it heads in the wrong direction, and breaks things along the way. It gets stuck spinning its wheels down some weird branch, constantly asking for help.
What can the Manager do at this point? It doesn't know the details either. It assigned the task to a Worker who can't get the job done, and the Worker reports back blankly. Both sides are talking past each other with pitifully little information about a complex task—the result is guaranteed to be a mess.
This is exactly how many people feel using OpenClaw right now.
The Economics
OpenClaw was originally called ClawdBot, connected to what was then the strongest model, Opus 4.5. Opus 4.5 had end-to-end execution capability, which is why the results were so astonishing.
Now there are stronger models. Opus 4.6 came out this January, and just a couple of days ago OpenAI released ChatGPT 5.4—after using it, honestly, I was shocked. Not just a little stronger, but a lot stronger.
But the problem is money.
Claude Code runs on a coding plan, similar to an all-you-can-eat buffet. I pay 2,000 if billed separately via API—a tenfold difference. These companies are subsidizing it, using this approach to get technical users on board first.
OpenClaw can't use the coding plan. It has to go through API pay-as-you-go billing. For the same usage, that's $2,000 a month.
So almost no one on the market connects the most expensive models to OpenClaw. For convenience, people connect domestic models. Domestic models are indeed improving fast, but there's still a gap compared to the very top tier.
When the Worker isn't strong enough, the Manager becomes a mere decoration. Many people feel like OpenClaw is just a toy.
But This Won't Be a Flash in the Pan
Model progress is happening much faster than imagined. Opus 4.6 came out in January, and two months later ChatGPT 5.4 surpassed it. Domestic players will catch up soon too.
The infrastructure is already laid out, and a huge number of people have installed OpenClaw. When new models come out, you only need to swap the connection. Last month you might have thought it was a bit dumb; two months later, switch the model and you'll realize it's practically a different species. In half a year, at the current pace, ordinary people will be able to experience that sense of awe from the strongest models.
And the barrier to entry is genuinely low. Install OpenClaw locally, spend a few dozen yuan per month on API fees—domestic cloud providers offer subsidies, ranging from twenty to forty yuan. When DeepSeek 4 comes out, just plug it in and it's ready to go. No longer just a chatbot.
OpenClaw lets non-coders use AI agents too. Everyone knows how to use a chat interface; the barrier has been lowered to the minimum. If you haven't installed it yet, check out this one-line command tutorial to set up OpenClaw.
It may not be the best experience right now. But people have already peeked through the crack in the door and seen what AI can do. As models upgrade, that door will open wider and wider. Install it first, and start using it.