I have some of my old contacts from my prior life flying airplanes for a living. I started there because I know the field extremely well. These are my first customers so far.
The first thing was just some really simple stuff a bush airline I used to work for needed too, like, their software is through a DB run by this other company, they wanted a status board customers could view. That shouldn't be a huge lift, but the company that runs the enterprise software doesn't have the time to build it.
I sent a series of emails, got permission to hit the API, and was able to connect things so now this little bush airline has a customer facing schedule app and people don't call the office 30 times an hour to see if the flight is late or on time or early. Even in the middle of nowhere, if they have Wifi the can check the flight schedule on their phone. That has spread to "hey, do you think you could use this data to auto-populate flight and duty logs?" Yup, not a huge deal. Then onto the next one. Every month it seems I take on a new project for them and the scope of their tooling keeps growing and the recurring costs I charge to maintain things is low enough where I'm worth it. There's a dashboard of data science stuff, then a compliance auditing tool, and the list of bespoke features that are critical to them continues to grow, and they continue to pay me. It's pretty cool.
This has lead to another customer pinging me that wants me to work on an app for their factory floor to help their technicians. Nothing crazy, just a kind of wrapper over USB tool they have and a CRUD app. 99% of the real work is going to be testing out like 30 different layouts and making sure that it works properly in practice, but a big company would never bother to do this. I will go down to their factory this week, set up computer, and talk with their technicians while I vibe code it out with Codex and draw process diagrams and think. 90% of it is really just thinking about what's a prudent choice.
The SaaS the first company is paying for is incredibly necessary to run their business, those guys will probably have their hooks into that operation for many more years because of the inertia to change, but there is tons of room to fix some of the little small annoyances that not having bespoke custom software creates. Also, the software they are kind of locked into is 10s of thousands of dollars a month. I reckon in the long run I'll end up trying to build a replacement for it entirely then charging way less to give them exactly what they need.
Then there's the existential angst of vibe coding this stuff. The truth is, I could write all this code myself. It's mostly Python, and JS, but it would take me a month to do what I can do in a week and I'd be working myself to the bone. Instead, this is more like an extremely fun part-time job that's growing in scope and pay but not growing in time required of me. Seriously, these tools are cool! They're like I have a team of idiot savants/interns working for me but the entire company so far is literally just me and my wife (and she isn't really involved in the technical stuff at all). Codex is dumb and does not understand the use case at all, but good lord does it churn out boilerplate code that solves real engineering problems for customers. My job is largely playing "software plumber foreman" and making sure all the lego pieces fit together nicely and that they're good architectural choices.
For example, I was skimming the code base last week and noticed a ton of just unused code from an early iteration. I spent a bunch of time pruning that as a human, then also having codex refactor code smells I didn't like. "This file is ridiculous, it's like a monolith of 30 different concepts hammered into one place - refactor all this stuff and spread it out, move function X to a separate file, use a functional style" etc. Stuff like that is kind of mandatory, otherwise your codebase will give you a stroke and you can grow it to an extraordinary size that will hurt your ability to iterate because you'll be running into context length issues. But the robot doesn't do too horrible of a job.
I could write all of the code, but the customers don't care if it's written by a human or not? They just want it to work. So I spend a lot of the time coming up with test-cases, then interactively evaluating what the robot is building? Kind of like a really slow REPL? But I'm definitely less of an engineer and more of an architect now. That pains me a bit? But all things must come to an end.
One thing I'd say is important if you're going to do this... use the dumbest possible solution you can. You'll need to specify that to these tools otherwise they'll build you a cathedral? You probably do not need some monster system with 80 layers of abstraction. KISS is important.