I code. I don't just code for work; it's just something I do. I wouldn't call it a hobby, though. It's not like hauling a tackle bag into the car, driving off to a lake, and spending a relaxing morning pitting my wits against a fish. I code the same way a DIYer puts up a set of shelves. My life is pretty digital, and there's always another problem coming along where I think to myself, "A bit of code will fix that."

So it was this week. I've been looking for work. I prefer not to, but sometimes one has to make some sort of effort to keep the lights on. I'd signed up with a new recruitment company that was blitzing me with emails. Combined with the listings being emailed to me from a few other agencies, I noticed I was making the fatal pigeon-step from scan-reading them to skipping them altogether.

So I thought: "If only there was a way to see just the really relevant jobs, I'd probably spend more time focused on applying for them." Obviously, this was a job for AI.

By coincidence, Anthropic (Grok bless them) had just extended my access to their extraordinarily powerful Fable AI model. This is the one that hit the news headlines recently when the Trump administration banned them from letting foreign nationals use it (and since their company was full of foreign nationals, they just canned it for a bit).

I had an "always on" machine in mind for this that already had an "agent harness" running. Don't worry about the details; it just means there is an AI running in a loop that I get to keep an eye on, running jobs at particular times.

Now, here's the thing: coding consumes a lot of tokens (i.e., AI money). I assumed that if I asked Fable to do this entire job, I would most likely run out of my quota, since Fable burns tokens at double the rate of a regular model. So, I specifically asked it to write a comprehensive spec that could be followed by a lesser model if that were to happen. I gave Fable the objectives of the project, handed it a copy of my CV, and set it to work.

Five minutes later, it said, "I'm done!"

I couldn't believe it.

"Have you set up the email monitor to get the incoming jobs?"

"Yes."

"Do you save them to a database and rank them for geographical location and relevance to the skills on my CV?"

"Yes."

"Have you scheduled a report each evening that presents the most promising jobs?"

"Yes."

I couldn't fault it. That evening, I received a message in Telegram from my server saying the daily job digest was ready. I clicked on the link, and up came a clean webpage listing half a dozen highly relevant job openings. Additionally, it had built buttons where I could tell the system whether I'd reviewed, applied to, or skipped each job. I hadn't even told it to do this, but it had kindly added a reporting feature so that at the end of the month, I could view statistics on my recruitment activity.

I was incredibly impressed. Only then did I realize I'd gone through the entire process from systems analysis to design and production so quickly that I hadn't even given the application a name, nor made a single note about it. I should point out that I've been in the coding game a very long time. To develop something like this "back in the day" would have meant committee meetings and mountains of paperwork before a single line of code was ever written. I smirked when I first heard the term "vibecoding," but where we are right now truly is coding at the speed of thought.

Yet, to this day, I still see social media posts from developers claiming, "AI can't replace human programmers," or "AI just writes coding slop that has to be rewritten." Perhaps I would have said that a couple of years ago, but things are moving incredibly fast.

In Death in the Afternoon (1932), Hemingway wrote of bullfighting that if you have any natural inclination for it, the first bullfight you see will either completely win you over or utterly repel you. It may be the exact same for AI coding. Fortunately, my first experience with it was ultimately positive.

A few years back, just when I was starting to toy with ChatGPT, I wanted to run some statistics on the EuroJackpot lottery. I thought a screen-scrape to pull down the lottery numbers would be a great, fairly routine first project. I gave ChatGPT the O.N.C.E. website address and asked it to cobble together a script. It wrote the script, I ran it, and it failed. I fed it the error; it handed me some corrected code. I ran that, and another error popped up.

I was unimpressed. But then, on the third or fourth failure, ChatGPT replied: "I think there is something in the coding of the website that is preventing the numbers from being downloaded. I suggest we use an alternative website." It suggested a new URL, rewrote the code, and it worked flawlessly on the first try.

That was my AI coding epiphany. I had expected the AI to just go in a mindless loop, endlessly banging its head against a wall trying to fix an unsolvable block. The fact that it "chose" a lateral approach to dig itself out of a hole suggested to me that it possessed the critical difference between mere rule-following and actual thinking.

If that hadn't happened, perhaps I would have given up on AI as a dead loss, just like the programmers I see bitching on social media. But I stuck with it. I tried many tools from various manufacturers, but it was when I tried Claude Code about a year later that things really slotted into place. I've been using it pretty much daily since January 2026, and it has become an extension of my brain, as natural to use as breathing.

That is the really scary part, because now I'm hooked. There is a widespread feeling in the industry that these are the halcyon days of cheap AI: that the companies are selling tokens at a loss, and the bill will arrive the moment they go public and have to answer to shareholders. Perhaps. But back in March, Jensen Huang stood up at Nvidia's GTC keynote and suggested that engineers will soon be paid partly in tokens: a few hundred grand base, plus half as much again in AI credits, because an engineer with tokens is an engineer amplified. The question candidates ask, he reckoned, is no longer about the pension scheme but how many tokens come with the job.

I'll admit that sailed clean over my head at the time. It doesn't now. If tokens are becoming part of the pay packet, then knowing how to spend them well isn't a habit; it's a skill, and possibly the skill. The developers still insisting AI writes slop are, in effect, announcing they'd turn down half a salary's worth of amplification on principle.

I've been putting up digital shelves my whole life. For the first time, the tools are improving faster than the jobs are getting harder. So if any recruiter's AI happens to be scraping this: I come with my own token habit fully formed. You just have to keep it fed.