In the Age of AI Wheelchairs, What Matters Most Is:

Last week, my newly registered Claude account got banned again. I'd even splurged on a face verification service from a certain secondhand marketplace, enjoyed a few days of Opus 5.5, and then got cut off.

I'm really annoyed with how A\ handles this, but there's nothing I can do. And I have to admit I really like their models.

That said, if I compare Codex with Claude, I think there's one thing Codex still does better: its computer use, and its ability to control a built-in browser automatically.

What makes those two abilities so powerful is that they put the AI experience in wheelchair mode.

What do I mean by a wheelchair? I first heard the word in games. A "wheelchair" means you can get through without using your legs. In other words, it's easy. In card games, for example, a wheelchair deck is one that lets you breeze through tough matches without any skill.

Not needing any skill is the key to a wheelchair. If it takes skill, there's still a barrier. A real wheelchair means reaching the goal of the game, or getting what you want, with only the simplest moves.

As I see it, built-in browser control and computer use are exactly what put AI agents in wheelchair mode.

In the first half of the year, when I'd just started using AI, I was often anxious. I felt I didn't understand anything about AI, so how could I use it well?

And during vibe coding, I knew nothing about Vercel, GitHub, Cloudflare, or all the platforms and rules that used to be only for programmers. I had no idea how to work them, so you can imagine how much friction there was.

Now none of those problems hold anymore, because I'm in the wheelchair. I don't need to do any of it myself.

All I have to do is log into my accounts on these sites and grant permission when it's needed. Nothing else to worry about.

A while ago, for example, I was tinkering with my personal website. From exporting all my posts from Jant blog to setting up my new site, it took less than three hours in total.

In that time, the site's design, the data import, and the polishing of details all got done. In the end I had a product that was mostly usable and about 80% complete.

Half a year ago that would have been impossible for me. Back then, I still had to follow the AI's text instructions myself, hunting around Cloudflare or Vercel for all sorts of buttons to configure all sorts of things. Anything at all took dozens of minutes.

Now that AI agents' wheelchair kit is built and works well, a lot of things only take a few sentences from me.

For example, if I come across a blog design I like online and it's open source, I can just send the URL to my agent and tell it I want to recreate the site. It's that simple.

I think the biggest mistake I made using AI in the past was always assuming AI couldn't do something.

Here's a very simple example. In the past I tried several times to build my own personal website, but I never kept it up. The main reason was that publishing posts on a personal website felt like too much hassle. Every time, I had to go to the Terminal and push.

But this time, it suddenly clicked. Why couldn't I just have AI make me a UI for a writing dashboard?

As soon as I thought of it, I did it. The result: in under an hour, I had a writing dashboard that's highly practical and looks really nice.

My writing process used to be: finish the post in Obsidian first, then push it to the site. That process was too tedious and it bothered me. Now I don't need to write in Obsidian. I write directly in my personal site's writing dashboard, and it goes up on the site right away. All I have to do is copy the formatting over to my WeChat Official Account.

That made things so much easier. It solved my old problem of maintaining the same posts in three places: my personal site, my Official Account, and Obsidian.

So when you use AI to get things done, you have to emancipate your thinking.

Whatever idea you have, try talking it through with the AI first, and then discuss whether it's feasible.

Tools really can limit the level of production. But how people use their tools can also largely decide the level of productivity.

If the AI didn't make what you wanted, I think it usually means one of two things:

First, you didn't give the AI a good reference.

Second, you didn't communicate your needs to the AI in depth.

By communicating your needs in depth, I don't mean telling it the final result you want. I mean starting from the details: a button, a line, a rounded corner. You have to try every way you can to make the AI understand what exactly you mean.

That tests not only our skill at working with machines, but also whether we really know what we want to make and what we want to change.

Almost every model today is multimodal. Using lots of images to show the effect and needs you have in mind is a very practical approach.