For a long time, I resisted it. When I used AI for development at all, it was only as autocomplete through Copilot. The idea that I would eventually have to clean up a pile of cobbled-together code put me off.
But after a while, that familiar feeling of “adapt or get left behind” took over. Using AI and its surrounding toolset is becoming a basic skill. It is starting to matter almost as much as knowing Git or any other traditional development tool inside out.
First attempts
A few months ago, I tried to build a long-running hobby project with various Claude models. Naturally, I went straight in: here is a half-page prompt, off you go. Just like in the demos.
The result was, at best, pathetic, and the disappointment was inevitable. “What did I expect?” I thought. I then spent a few more days trying to bash it into something usable together with Claude.
It did not convince me that I wanted to use this every day. To be honest, I had not really looked into how to use it well yet, either.
I know, it has a strong vibe-coding smell to it – and I would rather warn people away from that. But at the time, that was all anyone seemed to talk about.
A second try, after doing some homework
Weeks went by, and more and more of the content I had previously brushed aside found its way to me. This time, I started reading and watching it.
Eventually, I took a proper dive into my old pet project with ChatGPT and the Codex harness. I had been planning it for years, so it was already thoroughly documented.
Once I understood a little more about what I was doing, the first results immediately improved. Not long after that, the 5.6 Sol model arrived, and that was what finally convinced me.
I am not personally fond of CLI harnesses; I prefer a proper GUI. That is why I used Copilot for so long, and later Codex. For a while, I did not realise how much this matters. It is not the experience itself, but how well I can work with the tools, skills, workflows and all the other things I had not even heard of before.
The rabbit hole
I tend to take a proper head-first dive down the rabbit hole, whatever the subject.
I really enjoy that gradual discovery: when you come across more and more information on your own, and then something suddenly clicks. There is a special kind of magic in the newness of it. I learned a number of programming languages and technologies this way, and it has always been a rewarding way for me to learn.
Little by little, I went deeper:
- Delegating subagents that use different models to particular types of work, which helps reduce burning through limits
- Creating my own skills – something I did not think I would need at first
- Building more complex workflows around my own needs: planning, implementation, review, UI/UX design and brainstorming
- Keeping MCPs to a minimum and reducing token hunger
- Learning the tools that support the work, such as Beads, RTK and LSPs
- Using worktrees properly
- Sandbox permissions
I am probably still at the beginning. I have only been putting it through its paces for a few weeks, eight to ten hours a day – thanks to being unemployed – but I already find myself looking a little strangely at some creators who are only slightly behind me.
I went from being just a developer to PO, BA, QA and developer at once
While Codex is working in one worktree – or even two in parallel – I am planning in another. When it is done, I can start the code review.
It is efficient, but it is also a thousand times more exhausting than I expected. I am doing two or three things at once, keeping far more in my head, and constantly making or revisiting decisions. Codex asks a follow-up question and I realise: “this is not going to work.”
It is true that what came together in a few weeks would probably have taken me months to build alone, and it likely would not have been half as stable. Still, there are quite a few parts that feel a bit overengineered, and I would have solved them much more simply. I just did not have the energy to make GPT refactor them again after it had already spent forty or fifty minutes working on them.
So there is code that works well, but that I do not necessarily like.
For now, I am enjoying it immensely, but it is exhausting. I almost look forward to hitting the weekly limit.
Any specifics?
Once the details have settled a little more, I promise I will write more about concrete workflows, skills and similar things.
At the moment, I refine the processes every two or three days, because I keep noticing small things that could make them better.
I would also like to talk about the project itself. I have been thinking for a long time about how it could work in practice and as a business – perhaps it could even make money one day.