The Guitar That Wasn’t Worth Playing
I used to know a guitar player who could make any instrument sing. When he picked up a student’s guitar during lessons, people would stop what they were doing just to listen. He’d play a quick lick between exercises, just enough to remind everyone who the teacher was, but he never showed off. He was holding back most of the time.
His own guitar? A piece of cheap wood with strings on it. Nothing special. Not a brand name, not quality hardware. Just a basic instrument that was actually hard to play because of how poorly it was made.
But when he picked up a good guitar—one of those expensive models with the right wood, the right setup—he would fly.
I’ve also met guitar players with $4,000 signature models who couldn’t make them sound like anything special. They had the best tools money could buy, but they lacked the skills and talent to match.
The Same Pattern with AI Tools
This week, I was in a conversation about frontier language models—the expensive ones, the tools people pay $200 a month to use. Some developers swear by them. Others are building impressive things with the cheaper versions.
I don’t understand all the variations. When I open Cursor or Windsurf and see the model selector, there are so many options, and I honestly don’t yet know how to decide. So I just pick one and go with it.
When Cursor introduced auto mode—where it selects the model based on the request, favoring their free model—I started using that exclusively. And I’ve been more productive than I’ve ever been. I’m getting meaningful work done within a sprint that I couldn’t have imagined before.
I think it’s because of the techniques I’m using, the approach I’m taking, and the guardrails I’ve built around how I work with AI (grounded in hard-earned skills). At that level, the cheap model works just fine.
When the Expensive Model Matters
For months, I only used auto mode. I wasn’t even touching my $20 monthly budget for premium models.
Then, about two months ago, I hit a wall with dark mode styling. Cursor kept trying and failing. My teammate asked, “Have you tried selecting a different model?”
I switched to a premium model. Within two minutes, it solved the problem. Cost: about a dollar.
That’s when I reapplied a previously learned lesson. Once the problem was solved, I had the AI create a markdown file documenting how to handle similar issues in the future. I don’t need an expensive model to do that.
Now I know when to reach for the thinking models. When I have a complex problem that needs deeper reasoning, I’ll tell the tool to use a premium model. Then I switch back to auto.
The $9 Feature
A few sprints ago, I had a complex feature to implement. I knew I’d be giving Cursor a lot of context—not just volume, but complexity. Multiple places to look, specifications to parse, meeting notes to analyze.
I used a premium model. Two hours later, the feature was done. Cost: $9.
If I’d done it myself (without the AI tool)? Easily a full week. Reading specs, parsing notes, finding the right code, designing the solution. Instead, I spent $9 and two hours, and while the AI was working, I was doing something else. I’d check in occasionally, redirect it when needed, but mostly I was productive on other work.
Building Calluses on Cheap Guitars
I haven’t used a $200-a-month tool yet. I don’t know what I’d be able to do with it. But with my $20-a-month setup, I’m getting a lot (of meaningful work) done.
I recently learned about Devin—a cloud-based tool where you can run three agents in parallel, like having three junior developers working simultaneously. It’s expensive and looks promising.
But when I saw what it does, I thought, “I’m already doing that on my computer.” I have multiple monitors set up, each running a different AI assistant that works on different tasks (stories, features) in parallel. I should print out faces to put on each monitor because that’s how I think about them—different assistants, each with their own workspace.
The main difference is that I’m using my local machine, so they’re not both running at full speed, and my computer gets busy. Devin would offload that processing to the cloud, keeping my machine free for other intensive tasks.
But here’s what I’m noticing: even if I could run three assistants at full speed, I might be the bottleneck. I still need to review their work, make sure they have proper context, and ensure they’re working on the right things. If they’re producing code fast, but it’s the wrong code for the wrong feature, that’s not better.
I need to organize how I work with these assistants. I’ve added another screen to my setup, and I’m documenting this part of my system separately. But it’s still cheaper than paying for cloud tools.
When I Pick Up the Expensive Guitar
By the time I have this process figured out—how to properly orchestrate multiple AI assistants, how to be the bottleneck-free conductor—I hope I’ll be like my guitar-playing friend.
He plays his cheap guitar every day, building calluses, developing technique, and learning every limitation of that instrument. But when he picks up the expensive one, he melts faces.
Maybe that will be my case. I’m doing well with my cheap setup. I’m flexing these muscles, creating the calluses on my fingertips. When I have a better tool, it will flow. I’ll glide through things.
The expensive guitar doesn’t make you a better player. But if you’re already good, it lets you fly.
Leave a Reply