It's been one week since my surgery, and I've actually found a way to finally be productive in coding.

The only thing I've really had to code so far this week was migrating schematical.com from using a handwritten DynamoDB blog to now using PayloadCMS instead. But I was able to actually accomplish that using two tools.

Friendly reminder, this is dictated by Matt, not written. So if it sounds different, it's not AI; it's just that I'm talking to a program that writes this.

Let me start at the beginning though to tell you the whole story. Up until now my first attempts at coding were with ChatGPT. Not their coding tool, but their web user interface. Generally I only use that for writing bash scripts and things like that.

I was hesitant to generate any intellectual property using AI because that's a bit of a black box right now. So any intellectual property I generated for my clients and for myself that I knew we really needed to retain ownership of, I wouldn't write with AI.

But over the last few weeks, as I've realized I needed to be code-capable with one hand, I started digging into other tools. The first of which was just Claude Desktop, not Claude Code, Claude Desktop. Cloud Code comes with a pretty hefty fee, so I wanted to try out Cloud Desktop first. It did alright, I hooked up the file system MCP to it and a few other things, but it constantly was saying, "Hey, give us a credit card." I didn't get too far with that.

The second of which was Amazon Q running via the command line. This actually proved to be pretty useful. Not the best interface, but it actually was able to help me start the Dynamo to payload CMS migration. It ran into a lot of problems when I tried to add mcp's to it. Sometimes they worked, sometimes they didn't. It was a bit of a mixed bag.

After that, I went down the open source route by setting up a repo with Ollama, Open Web UI, and a few other tools running locally. I really learned a lot here about models and feeding them prompts.

I'll dive deeper into the techniques I used to debug this later, but it was pretty eye-opening for me. The end results weren't the best. The user interfaces, the open web UI and all that stuff was still pretty clunky. And it was a bit slow, because it's running on my desktop, which does have some GPUs, but it would time out with the real big models.

After that Gemini CLI came out and they give you a thousand free requests per day. I was making good progress with that except I managed to max out those requests in 45 minutes. So I had to find something else.

Finally, I decided to bite the bullet and download Cursor. I've got a two-week pro trial going on with that right now. And I have had quite a bit of success with that. Is it perfect? No, as a matter of fact it's off the rails right now. I think it's got 4 Next.js processes running at once for the exact same codebase because it kept forgetting it had one running already. But combining Cursor and Wispr Flow for dictation, and me starting to use my right hand again just to manipulate the mouse a little bit, I've actually been fairly efficient.

I added in the MongoDB MCP and playwright, and it can debug the browser. It actually was able to fix CSS issues like very small pixel off CSS issues and generate entire pages, basically looking at the existing MongoDB pages in the codebase. So it generated new landing pages for me. They didn't have any images in it, but it still was pretty cool to see it learn and figure out how to write the page schema directly to the database without even having to use the admin.

I probably will bite the bullet and check out Claude code, but for now, I'm actually very happy with Cursor.

Would I call what I'm doing "vibe coding"? No, because I actually look at every file that it changes and updates, and I know how to code, so it's just holding me over until I get use of my arm.

Will I continue to use it afterwards? I will probably consult with my attorneys about the copyright laws.

It's interesting to see so many established software developers using these tools. I'm not sure there's a way not to moving forward, but at the same time I can't be writing code that we don't know who owns the Intellectual Property.

I'd love to know what your thoughts are on using AI tools for code generation. Shoot me a message, etc.

PS: Just a reminder today at 1PM Central I will be doing a live Tech Talk on how to scale LLMs and Model Context Protocol if you care to join.